FP32 IMAGIC variants of scalar QC8/QS8/QU8 GEMM/IGEMM/DWCONV microkernels
PiperOrigin-RevId: 419687973
diff --git a/test/qc8-dwconv-minmax-fp32.cc b/test/qc8-dwconv-minmax-fp32.cc
index 1103df3..d30f371 100644
--- a/test/qc8-dwconv-minmax-fp32.cc
+++ b/test/qc8-dwconv-minmax-fp32.cc
@@ -11834,468 +11834,6 @@
#endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_eq_1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(1)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_gt_1) {
- for (uint32_t channels = 2; channels < 10; channels++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_gt_1_with_qmin) {
- for (uint32_t channels = 2; channels < 10; channels++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_gt_1_with_qmax) {
- for (uint32_t channels = 2; channels < 10; channels++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .width(3)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_step) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- for (size_t step = 2; step <= 9; step++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .width(3)
- .step(step)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_output_stride) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(1)
- .width(5)
- .output_stride(7)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_qmin) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .width(3)
- .qmin(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_qmax) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .width(3)
- .qmax(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, input_offset) {
- for (uint32_t channels = 2; channels < 16; channels += 3) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .input_offset(48)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 9; mz++) {
- for (uint32_t channels = 2; channels < 16; channels += 3) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .input_offset(48)
- .zero_index(mz)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_eq_2) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(2)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_div_2) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_div_2_with_qmin) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_div_2_with_qmax) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_lt_2) {
- for (uint32_t channels = 1; channels < 2; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_gt_2) {
- for (uint32_t channels = 3; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_gt_2_with_qmin) {
- for (uint32_t channels = 3; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_gt_2_with_qmax) {
- for (uint32_t channels = 3; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .width(3)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_step) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- for (size_t step = 2; step <= 9; step++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .width(3)
- .step(step)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_output_stride) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(2)
- .width(5)
- .output_stride(13)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_qmin) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .width(3)
- .qmin(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_qmax) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .width(3)
- .qmax(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, input_offset) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .input_offset(80)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 9; mz++) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .input_offset(80)
- .zero_index(mz)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_eq_4) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(4)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_div_4) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_div_4_with_qmin) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_div_4_with_qmax) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_lt_4) {
- for (uint32_t channels = 1; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_gt_4) {
- for (uint32_t channels = 5; channels < 8; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_gt_4_with_qmin) {
- for (uint32_t channels = 5; channels < 8; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_gt_4_with_qmax) {
- for (uint32_t channels = 5; channels < 8; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .width(3)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_step) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- for (size_t step = 2; step <= 9; step++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .width(3)
- .step(step)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_output_stride) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(4)
- .width(5)
- .output_stride(23)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_qmin) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .width(3)
- .qmin(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_qmax) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .width(3)
- .qmax(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, input_offset) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .input_offset(112)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 9; mz++) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .input_offset(112)
- .zero_index(mz)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_FMAGIC, c_eq_1) {
DWConvMicrokernelTester()
.cr(1)
@@ -12758,6 +12296,930 @@
}
}
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, c_eq_1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(1)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, c_gt_1) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, c_gt_1_with_qmin) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, c_gt_1_with_qmax) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, multipixel) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ for (size_t step = 2; step <= 9; step++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(1)
+ .width(5)
+ .output_stride(7)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, input_offset) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .input_offset(48)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 9; mz++) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .input_offset(48)
+ .zero_index(mz)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_eq_2) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(2)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_div_2) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_div_2_with_qmin) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_div_2_with_qmax) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_lt_2) {
+ for (uint32_t channels = 1; channels < 2; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_gt_2) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_gt_2_with_qmin) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_gt_2_with_qmax) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, multipixel) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ for (size_t step = 2; step <= 9; step++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(2)
+ .width(5)
+ .output_stride(13)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, input_offset) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .input_offset(80)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 9; mz++) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .input_offset(80)
+ .zero_index(mz)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_eq_4) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(4)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_div_4) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_div_4_with_qmin) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_div_4_with_qmax) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_lt_4) {
+ for (uint32_t channels = 1; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_gt_4) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_gt_4_with_qmin) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_gt_4_with_qmax) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, multipixel) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ for (size_t step = 2; step <= 9; step++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(4)
+ .width(5)
+ .output_stride(23)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, input_offset) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .input_offset(112)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 9; mz++) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .input_offset(112)
+ .zero_index(mz)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_eq_1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(1)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_gt_1) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_gt_1_with_qmin) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_gt_1_with_qmax) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ for (size_t step = 2; step <= 9; step++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(1)
+ .width(5)
+ .output_stride(7)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, input_offset) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .input_offset(48)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 9; mz++) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .input_offset(48)
+ .zero_index(mz)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_eq_2) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(2)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_div_2) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_div_2_with_qmin) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_div_2_with_qmax) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_lt_2) {
+ for (uint32_t channels = 1; channels < 2; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_gt_2) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_gt_2_with_qmin) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_gt_2_with_qmax) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ for (size_t step = 2; step <= 9; step++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(2)
+ .width(5)
+ .output_stride(13)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, input_offset) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .input_offset(80)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 9; mz++) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .input_offset(80)
+ .zero_index(mz)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_eq_4) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(4)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_div_4) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_div_4_with_qmin) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_div_4_with_qmax) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_lt_4) {
+ for (uint32_t channels = 1; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_gt_4) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_gt_4_with_qmin) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_gt_4_with_qmax) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ for (size_t step = 2; step <= 9; step++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(4)
+ .width(5)
+ .output_stride(23)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, input_offset) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .input_offset(112)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 9; mz++) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .input_offset(112)
+ .zero_index(mz)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
#if XNN_ARCH_ARM || XNN_ARCH_ARM64
TEST(QC8_DWCONV_MINMAX_FP32_UP8X25__NEON_MUL8_LD64, c_eq_8) {
TEST_REQUIRES_ARM_NEON;
@@ -24572,468 +25034,6 @@
#endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_eq_1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(1)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_gt_1) {
- for (uint32_t channels = 2; channels < 10; channels++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_gt_1_with_qmin) {
- for (uint32_t channels = 2; channels < 10; channels++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_gt_1_with_qmax) {
- for (uint32_t channels = 2; channels < 10; channels++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .width(3)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_step) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- for (size_t step = 2; step <= 25; step++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .width(3)
- .step(step)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_output_stride) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(1)
- .width(5)
- .output_stride(7)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_qmin) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .width(3)
- .qmin(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_qmax) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .width(3)
- .qmax(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, input_offset) {
- for (uint32_t channels = 2; channels < 16; channels += 3) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .input_offset(48)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 25; mz++) {
- for (uint32_t channels = 2; channels < 16; channels += 3) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .input_offset(48)
- .zero_index(mz)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_eq_2) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(2)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_div_2) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_div_2_with_qmin) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_div_2_with_qmax) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_lt_2) {
- for (uint32_t channels = 1; channels < 2; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_gt_2) {
- for (uint32_t channels = 3; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_gt_2_with_qmin) {
- for (uint32_t channels = 3; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_gt_2_with_qmax) {
- for (uint32_t channels = 3; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .width(3)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_step) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- for (size_t step = 2; step <= 25; step++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .width(3)
- .step(step)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_output_stride) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(2)
- .width(5)
- .output_stride(13)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_qmin) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .width(3)
- .qmin(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_qmax) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .width(3)
- .qmax(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, input_offset) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .input_offset(80)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 25; mz++) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .input_offset(80)
- .zero_index(mz)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_eq_4) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(4)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_div_4) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_div_4_with_qmin) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_div_4_with_qmax) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_lt_4) {
- for (uint32_t channels = 1; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_gt_4) {
- for (uint32_t channels = 5; channels < 8; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_gt_4_with_qmin) {
- for (uint32_t channels = 5; channels < 8; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_gt_4_with_qmax) {
- for (uint32_t channels = 5; channels < 8; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .width(3)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_step) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- for (size_t step = 2; step <= 25; step++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .width(3)
- .step(step)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_output_stride) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(4)
- .width(5)
- .output_stride(23)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_qmin) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .width(3)
- .qmin(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_qmax) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .width(3)
- .qmax(128)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, input_offset) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .input_offset(112)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 25; mz++) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .input_offset(112)
- .zero_index(mz)
- .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_FMAGIC, c_eq_1) {
DWConvMicrokernelTester()
.cr(1)
@@ -25494,4 +25494,928 @@
.Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_fmagic, xnn_init_qs8_minmax_scalar_fmagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
}
}
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, c_eq_1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(1)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, c_gt_1) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, c_gt_1_with_qmin) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, c_gt_1_with_qmax) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, multipixel) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ for (size_t step = 2; step <= 25; step++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(1)
+ .width(5)
+ .output_stride(7)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, input_offset) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .input_offset(48)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 25; mz++) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .input_offset(48)
+ .zero_index(mz)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_eq_2) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(2)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_div_2) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_div_2_with_qmin) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_div_2_with_qmax) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_lt_2) {
+ for (uint32_t channels = 1; channels < 2; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_gt_2) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_gt_2_with_qmin) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_gt_2_with_qmax) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, multipixel) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ for (size_t step = 2; step <= 25; step++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(2)
+ .width(5)
+ .output_stride(13)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, input_offset) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .input_offset(80)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 25; mz++) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .input_offset(80)
+ .zero_index(mz)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_eq_4) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(4)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_div_4) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_div_4_with_qmin) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_div_4_with_qmax) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_lt_4) {
+ for (uint32_t channels = 1; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_gt_4) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_gt_4_with_qmin) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_gt_4_with_qmax) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, multipixel) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ for (size_t step = 2; step <= 25; step++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(4)
+ .width(5)
+ .output_stride(23)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, input_offset) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .input_offset(112)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 25; mz++) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .input_offset(112)
+ .zero_index(mz)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_eq_1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(1)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_gt_1) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_gt_1_with_qmin) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_gt_1_with_qmax) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ for (size_t step = 2; step <= 25; step++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(1)
+ .width(5)
+ .output_stride(7)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, input_offset) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .input_offset(48)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 25; mz++) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .input_offset(48)
+ .zero_index(mz)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_eq_2) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(2)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_div_2) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_div_2_with_qmin) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_div_2_with_qmax) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_lt_2) {
+ for (uint32_t channels = 1; channels < 2; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_gt_2) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_gt_2_with_qmin) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_gt_2_with_qmax) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ for (size_t step = 2; step <= 25; step++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(2)
+ .width(5)
+ .output_stride(13)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, input_offset) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .input_offset(80)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 25; mz++) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .input_offset(80)
+ .zero_index(mz)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_eq_4) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(4)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_div_4) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_div_4_with_qmin) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_div_4_with_qmax) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_lt_4) {
+ for (uint32_t channels = 1; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_gt_4) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_gt_4_with_qmin) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_gt_4_with_qmax) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ for (size_t step = 2; step <= 25; step++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(4)
+ .width(5)
+ .output_stride(23)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, input_offset) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .input_offset(112)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 25; mz++) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .input_offset(112)
+ .zero_index(mz)
+ .Test(xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
}
\ No newline at end of file
diff --git a/test/qc8-dwconv-minmax-fp32.yaml b/test/qc8-dwconv-minmax-fp32.yaml
index b0f0bb2..15dc85a 100644
--- a/test/qc8-dwconv-minmax-fp32.yaml
+++ b/test/qc8-dwconv-minmax-fp32.yaml
@@ -131,18 +131,24 @@
init: xnn_init_qs8_minmax_wasmsimd_params
- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__wasmsimd_mul16_add16
init: xnn_init_qs8_minmax_wasmsimd_params
-- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
-- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
-- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_fmagic
init: xnn_init_qs8_minmax_scalar_fmagic_params
- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_fmagic
init: xnn_init_qs8_minmax_scalar_fmagic_params
- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_fmagic
init: xnn_init_qs8_minmax_scalar_fmagic_params
+- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul8_ld64
init: xnn_init_qs8_minmax_neon_params
- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul8_ld64
@@ -271,15 +277,21 @@
init: xnn_init_qs8_minmax_wasmsimd_params
- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up24x25__wasmsimd_mul16_add16
init: xnn_init_qs8_minmax_wasmsimd_params
-- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
-- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
-- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_fmagic
init: xnn_init_qs8_minmax_scalar_fmagic_params
- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_fmagic
init: xnn_init_qs8_minmax_scalar_fmagic_params
- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_fmagic
init: xnn_init_qs8_minmax_scalar_fmagic_params
+- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+- name: xnn_qc8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
diff --git a/test/qc8-gemm-minmax-fp32.cc b/test/qc8-gemm-minmax-fp32.cc
index 80b2a7d..880e442 100644
--- a/test/qc8-gemm-minmax-fp32.cc
+++ b/test/qc8-gemm-minmax-fp32.cc
@@ -77180,2654 +77180,6 @@
#endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .a_stride(3)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .a_stride(11)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_strided_a) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_strided_a) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .a_stride(3)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .a_stride(11)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_strided_a) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_strided_a) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .a_stride(3)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .a_stride(11)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_strided_a) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_strided_a) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .a_stride(3)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .a_stride(11)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_strided_a) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_strided_a) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .a_stride(3)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .a_stride(11)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_strided_a) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_strided_a) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .a_stride(3)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .a_stride(11)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_strided_a) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_strided_a) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .a_stride(3)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .a_stride(11)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_strided_a) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_strided_a) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .a_stride(3)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .a_stride(11)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_strided_a) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_strided_a) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_FMAGIC, k_eq_1) {
GemmMicrokernelTester()
.mr(1)
@@ -82474,3 +79826,5299 @@
.cm_stride(7)
.Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_fmagic, xnn_init_qs8_minmax_scalar_fmagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
}
+
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
diff --git a/test/qc8-gemm-minmax-fp32.yaml b/test/qc8-gemm-minmax-fp32.yaml
index 23e095c..bfba07b 100644
--- a/test/qc8-gemm-minmax-fp32.yaml
+++ b/test/qc8-gemm-minmax-fp32.yaml
@@ -539,30 +539,6 @@
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16_ld128
init: xnn_init_qs8_minmax_wasmsimd_params
k-block: 8
-- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
- k-block: 1
-- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
- k-block: 1
-- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
- k-block: 1
-- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
- k-block: 1
-- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
- k-block: 1
-- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
- k-block: 1
-- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
- k-block: 1
-- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
- k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_fmagic
init: xnn_init_qs8_minmax_scalar_fmagic_params
k-block: 1
@@ -587,3 +563,51 @@
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_fmagic
init: xnn_init_qs8_minmax_scalar_fmagic_params
k-block: 1
+- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+ k-block: 1
+- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+ k-block: 1
+- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+ k-block: 1
+- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+ k-block: 1
+- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+ k-block: 1
+- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+ k-block: 1
+- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+ k-block: 1
+- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+ k-block: 1
+- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+ k-block: 1
diff --git a/test/qc8-igemm-minmax-fp32.cc b/test/qc8-igemm-minmax-fp32.cc
index de7517e..f26b2b8 100644
--- a/test/qc8-igemm-minmax-fp32.cc
+++ b/test/qc8-igemm-minmax-fp32.cc
@@ -75029,2990 +75029,6 @@
#endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_small_kernel) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(7)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 1; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(7)
- .zero_index(mz)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_small_kernel) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(13)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 2; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(13)
- .zero_index(mz)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_small_kernel) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(17)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 3; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(17)
- .zero_index(mz)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_small_kernel) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(23)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 4; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(23)
- .zero_index(mz)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_small_kernel) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(7)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 1; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(7)
- .zero_index(mz)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_small_kernel) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(13)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 2; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(13)
- .zero_index(mz)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_small_kernel) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(17)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 3; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(17)
- .zero_index(mz)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_small_kernel) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(23)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 4; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(23)
- .zero_index(mz)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_FMAGIC, k_eq_1) {
GemmMicrokernelTester()
.mr(1)
@@ -80995,3 +78011,5971 @@
.cm_stride(7)
.Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_fmagic, xnn_init_qs8_minmax_scalar_fmagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
}
+
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 1; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .zero_index(mz)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 2; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .zero_index(mz)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 3; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .zero_index(mz)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 4; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .zero_index(mz)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 1; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .zero_index(mz)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 2; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .zero_index(mz)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 3; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .zero_index(mz)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 4; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .zero_index(mz)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_minmax_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 1; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .zero_index(mz)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 2; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .zero_index(mz)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 3; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .zero_index(mz)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 4; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .zero_index(mz)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 1; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .zero_index(mz)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 2; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .zero_index(mz)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 3; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .zero_index(mz)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 4; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .zero_index(mz)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QC8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_minmax_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
diff --git a/test/qc8-igemm-minmax-fp32.yaml b/test/qc8-igemm-minmax-fp32.yaml
index 6c1861e..1c1611d 100644
--- a/test/qc8-igemm-minmax-fp32.yaml
+++ b/test/qc8-igemm-minmax-fp32.yaml
@@ -506,30 +506,6 @@
- name: xnn_qc8_igemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16_ld128
init: xnn_init_qs8_minmax_wasmsimd_params
k-block: 8
-- name: xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
- k-block: 1
-- name: xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
- k-block: 1
-- name: xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
- k-block: 1
-- name: xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
- k-block: 1
-- name: xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
- k-block: 1
-- name: xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
- k-block: 1
-- name: xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
- k-block: 1
-- name: xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf
- init: xnn_init_qs8_minmax_scalar_lrintf_params
- k-block: 1
- name: xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_fmagic
init: xnn_init_qs8_minmax_scalar_fmagic_params
k-block: 1
@@ -554,3 +530,51 @@
- name: xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_fmagic
init: xnn_init_qs8_minmax_scalar_fmagic_params
k-block: 1
+- name: xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+ k-block: 1
+- name: xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+ k-block: 1
+- name: xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+ k-block: 1
+- name: xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+ k-block: 1
+- name: xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+ k-block: 1
+- name: xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+ k-block: 1
+- name: xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+ k-block: 1
+- name: xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic
+ init: xnn_init_qs8_minmax_scalar_imagic_params
+ k-block: 1
+- name: xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qc8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qc8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qc8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf
+ init: xnn_init_qs8_minmax_scalar_lrintf_params
+ k-block: 1
diff --git a/test/qs8-dwconv-minmax-fp32.cc b/test/qs8-dwconv-minmax-fp32.cc
index cc067d6..accc2a9 100644
--- a/test/qs8-dwconv-minmax-fp32.cc
+++ b/test/qs8-dwconv-minmax-fp32.cc
@@ -9602,468 +9602,6 @@
#endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_eq_1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(1)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_gt_1) {
- for (uint32_t channels = 2; channels < 10; channels++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_gt_1_with_qmin) {
- for (uint32_t channels = 2; channels < 10; channels++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_gt_1_with_qmax) {
- for (uint32_t channels = 2; channels < 10; channels++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .width(3)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_step) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- for (size_t step = 2; step <= 9; step++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .width(3)
- .step(step)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_output_stride) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(1)
- .width(5)
- .output_stride(7)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_qmin) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .width(3)
- .qmin(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_qmax) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .width(3)
- .qmax(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, input_offset) {
- for (uint32_t channels = 2; channels < 16; channels += 3) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .input_offset(48)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 9; mz++) {
- for (uint32_t channels = 2; channels < 16; channels += 3) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .input_offset(48)
- .zero_index(mz)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_eq_2) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(2)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_div_2) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_div_2_with_qmin) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_div_2_with_qmax) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_lt_2) {
- for (uint32_t channels = 1; channels < 2; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_gt_2) {
- for (uint32_t channels = 3; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_gt_2_with_qmin) {
- for (uint32_t channels = 3; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_gt_2_with_qmax) {
- for (uint32_t channels = 3; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .width(3)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_step) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- for (size_t step = 2; step <= 9; step++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .width(3)
- .step(step)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_output_stride) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(2)
- .width(5)
- .output_stride(13)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_qmin) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .width(3)
- .qmin(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_qmax) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .width(3)
- .qmax(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, input_offset) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .input_offset(80)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 9; mz++) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .input_offset(80)
- .zero_index(mz)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_eq_4) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(4)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_div_4) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_div_4_with_qmin) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_div_4_with_qmax) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_lt_4) {
- for (uint32_t channels = 1; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_gt_4) {
- for (uint32_t channels = 5; channels < 8; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_gt_4_with_qmin) {
- for (uint32_t channels = 5; channels < 8; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_gt_4_with_qmax) {
- for (uint32_t channels = 5; channels < 8; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .width(3)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_step) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- for (size_t step = 2; step <= 9; step++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .width(3)
- .step(step)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_output_stride) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(4)
- .width(5)
- .output_stride(23)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_qmin) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .width(3)
- .qmin(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_qmax) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .width(3)
- .qmax(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, input_offset) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .input_offset(112)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 9; mz++) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .input_offset(112)
- .zero_index(mz)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_FMAGIC, c_eq_1) {
DWConvMicrokernelTester()
.cr(1)
@@ -10526,6 +10064,930 @@
}
}
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, c_eq_1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(1)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, c_gt_1) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, c_gt_1_with_qmin) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, c_gt_1_with_qmax) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, multipixel) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ for (size_t step = 2; step <= 9; step++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(1)
+ .width(5)
+ .output_stride(7)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, input_offset) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .input_offset(48)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 9; mz++) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .input_offset(48)
+ .zero_index(mz)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_eq_2) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(2)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_div_2) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_div_2_with_qmin) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_div_2_with_qmax) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_lt_2) {
+ for (uint32_t channels = 1; channels < 2; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_gt_2) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_gt_2_with_qmin) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_gt_2_with_qmax) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, multipixel) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ for (size_t step = 2; step <= 9; step++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(2)
+ .width(5)
+ .output_stride(13)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, input_offset) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .input_offset(80)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 9; mz++) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .input_offset(80)
+ .zero_index(mz)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_eq_4) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(4)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_div_4) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_div_4_with_qmin) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_div_4_with_qmax) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_lt_4) {
+ for (uint32_t channels = 1; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_gt_4) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_gt_4_with_qmin) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_gt_4_with_qmax) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, multipixel) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ for (size_t step = 2; step <= 9; step++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(4)
+ .width(5)
+ .output_stride(23)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, input_offset) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .input_offset(112)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 9; mz++) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .input_offset(112)
+ .zero_index(mz)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_eq_1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(1)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_gt_1) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_gt_1_with_qmin) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_gt_1_with_qmax) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ for (size_t step = 2; step <= 9; step++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(1)
+ .width(5)
+ .output_stride(7)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, input_offset) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .input_offset(48)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 9; mz++) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .input_offset(48)
+ .zero_index(mz)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_eq_2) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(2)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_div_2) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_div_2_with_qmin) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_div_2_with_qmax) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_lt_2) {
+ for (uint32_t channels = 1; channels < 2; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_gt_2) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_gt_2_with_qmin) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_gt_2_with_qmax) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ for (size_t step = 2; step <= 9; step++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(2)
+ .width(5)
+ .output_stride(13)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, input_offset) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .input_offset(80)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 9; mz++) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .input_offset(80)
+ .zero_index(mz)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_eq_4) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(4)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_div_4) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_div_4_with_qmin) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_div_4_with_qmax) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_lt_4) {
+ for (uint32_t channels = 1; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_gt_4) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_gt_4_with_qmin) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_gt_4_with_qmax) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ for (size_t step = 2; step <= 9; step++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(4)
+ .width(5)
+ .output_stride(23)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, input_offset) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .input_offset(112)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 9; mz++) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .input_offset(112)
+ .zero_index(mz)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
#if XNN_ARCH_ARM || XNN_ARCH_ARM64
TEST(QS8_DWCONV_MINMAX_FP32_UP8X25__NEON_MUL16, c_eq_8) {
TEST_REQUIRES_ARM_NEON;
@@ -20108,468 +20570,6 @@
#endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_eq_1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(1)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_gt_1) {
- for (uint32_t channels = 2; channels < 10; channels++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_gt_1_with_qmin) {
- for (uint32_t channels = 2; channels < 10; channels++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_gt_1_with_qmax) {
- for (uint32_t channels = 2; channels < 10; channels++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .width(3)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_step) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- for (size_t step = 2; step <= 25; step++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .width(3)
- .step(step)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_output_stride) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(1)
- .width(5)
- .output_stride(7)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_qmin) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .width(3)
- .qmin(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_qmax) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .width(3)
- .qmax(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, input_offset) {
- for (uint32_t channels = 2; channels < 16; channels += 3) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .input_offset(48)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 25; mz++) {
- for (uint32_t channels = 2; channels < 16; channels += 3) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .input_offset(48)
- .zero_index(mz)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_eq_2) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(2)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_div_2) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_div_2_with_qmin) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_div_2_with_qmax) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_lt_2) {
- for (uint32_t channels = 1; channels < 2; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_gt_2) {
- for (uint32_t channels = 3; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_gt_2_with_qmin) {
- for (uint32_t channels = 3; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_gt_2_with_qmax) {
- for (uint32_t channels = 3; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .width(3)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_step) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- for (size_t step = 2; step <= 25; step++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .width(3)
- .step(step)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_output_stride) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(2)
- .width(5)
- .output_stride(13)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_qmin) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .width(3)
- .qmin(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_qmax) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .width(3)
- .qmax(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, input_offset) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .input_offset(80)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 25; mz++) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .input_offset(80)
- .zero_index(mz)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_eq_4) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(4)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_div_4) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_div_4_with_qmin) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_div_4_with_qmax) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_lt_4) {
- for (uint32_t channels = 1; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_gt_4) {
- for (uint32_t channels = 5; channels < 8; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_gt_4_with_qmin) {
- for (uint32_t channels = 5; channels < 8; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_gt_4_with_qmax) {
- for (uint32_t channels = 5; channels < 8; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .width(3)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_step) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- for (size_t step = 2; step <= 25; step++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .width(3)
- .step(step)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_output_stride) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(4)
- .width(5)
- .output_stride(23)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_qmin) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .width(3)
- .qmin(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_qmax) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .width(3)
- .qmax(128)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, input_offset) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .input_offset(112)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 25; mz++) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .input_offset(112)
- .zero_index(mz)
- .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_FMAGIC, c_eq_1) {
DWConvMicrokernelTester()
.cr(1)
@@ -21030,4 +21030,928 @@
.Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_fmagic, xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
}
}
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, c_eq_1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(1)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, c_gt_1) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, c_gt_1_with_qmin) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, c_gt_1_with_qmax) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, multipixel) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ for (size_t step = 2; step <= 25; step++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(1)
+ .width(5)
+ .output_stride(7)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, input_offset) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .input_offset(48)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 25; mz++) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .input_offset(48)
+ .zero_index(mz)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_eq_2) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(2)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_div_2) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_div_2_with_qmin) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_div_2_with_qmax) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_lt_2) {
+ for (uint32_t channels = 1; channels < 2; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_gt_2) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_gt_2_with_qmin) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_gt_2_with_qmax) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, multipixel) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ for (size_t step = 2; step <= 25; step++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(2)
+ .width(5)
+ .output_stride(13)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, input_offset) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .input_offset(80)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 25; mz++) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .input_offset(80)
+ .zero_index(mz)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_eq_4) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(4)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_div_4) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_div_4_with_qmin) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_div_4_with_qmax) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_lt_4) {
+ for (uint32_t channels = 1; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_gt_4) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_gt_4_with_qmin) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_gt_4_with_qmax) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, multipixel) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ for (size_t step = 2; step <= 25; step++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(4)
+ .width(5)
+ .output_stride(23)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, input_offset) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .input_offset(112)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 25; mz++) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .input_offset(112)
+ .zero_index(mz)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_eq_1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(1)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_gt_1) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_gt_1_with_qmin) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_gt_1_with_qmax) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ for (size_t step = 2; step <= 25; step++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(1)
+ .width(5)
+ .output_stride(7)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, input_offset) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .input_offset(48)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 25; mz++) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .input_offset(48)
+ .zero_index(mz)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_eq_2) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(2)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_div_2) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_div_2_with_qmin) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_div_2_with_qmax) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_lt_2) {
+ for (uint32_t channels = 1; channels < 2; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_gt_2) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_gt_2_with_qmin) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_gt_2_with_qmax) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ for (size_t step = 2; step <= 25; step++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(2)
+ .width(5)
+ .output_stride(13)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, input_offset) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .input_offset(80)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 25; mz++) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .input_offset(80)
+ .zero_index(mz)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_eq_4) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(4)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_div_4) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_div_4_with_qmin) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_div_4_with_qmax) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_lt_4) {
+ for (uint32_t channels = 1; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_gt_4) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_gt_4_with_qmin) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_gt_4_with_qmax) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ for (size_t step = 2; step <= 25; step++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(4)
+ .width(5)
+ .output_stride(23)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, input_offset) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .input_offset(112)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 25; mz++) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .input_offset(112)
+ .zero_index(mz)
+ .Test(xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
}
\ No newline at end of file
diff --git a/test/qs8-dwconv-minmax-fp32.yaml b/test/qs8-dwconv-minmax-fp32.yaml
index fea6679..1f02558 100644
--- a/test/qs8-dwconv-minmax-fp32.yaml
+++ b/test/qs8-dwconv-minmax-fp32.yaml
@@ -106,18 +106,24 @@
init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params
- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__wasmsimd_mul16_add16
init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params
-- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
-- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
-- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_fmagic
init: xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params
- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_fmagic
init: xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params
- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_fmagic
init: xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params
+- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16
init: xnn_init_qs8_conv_minmax_fp32_neon_params
- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16
@@ -222,15 +228,21 @@
init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params
- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__wasmsimd_mul16_add16
init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params
-- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
-- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
-- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_fmagic
init: xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params
- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_fmagic
init: xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params
- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_fmagic
init: xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params
+- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+- name: xnn_qs8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
diff --git a/test/qs8-gemm-minmax-fp32.cc b/test/qs8-gemm-minmax-fp32.cc
index 34b0430..e6f64f5 100644
--- a/test/qs8-gemm-minmax-fp32.cc
+++ b/test/qs8-gemm-minmax-fp32.cc
@@ -94107,2654 +94107,6 @@
#endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .a_stride(3)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .a_stride(11)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_strided_a) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_strided_a) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .a_stride(3)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .a_stride(11)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_strided_a) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_strided_a) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .a_stride(3)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .a_stride(11)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_strided_a) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_strided_a) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .a_stride(3)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .a_stride(11)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_strided_a) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_strided_a) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .a_stride(3)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .a_stride(11)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_strided_a) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_strided_a) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .a_stride(3)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .a_stride(11)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_strided_a) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_strided_a) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .a_stride(3)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .a_stride(11)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_strided_a) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_strided_a) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .a_stride(3)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .a_stride(11)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_strided_a) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_strided_a) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_FMAGIC, k_eq_1) {
GemmMicrokernelTester()
.mr(1)
@@ -99401,3 +96753,5299 @@
.cm_stride(7)
.Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_fmagic, xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
}
+
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
diff --git a/test/qs8-gemm-minmax-fp32.yaml b/test/qs8-gemm-minmax-fp32.yaml
index 96a0baa..20c3c34 100644
--- a/test/qs8-gemm-minmax-fp32.yaml
+++ b/test/qs8-gemm-minmax-fp32.yaml
@@ -650,30 +650,6 @@
- name: xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16
init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params
k-block: 8
-- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_fmagic
init: xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params
k-block: 1
@@ -698,3 +674,51 @@
- name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_fmagic
init: xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params
k-block: 1
+- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
diff --git a/test/qs8-igemm-minmax-fp32.cc b/test/qs8-igemm-minmax-fp32.cc
index 69350b7..3c21f5b 100644
--- a/test/qs8-igemm-minmax-fp32.cc
+++ b/test/qs8-igemm-minmax-fp32.cc
@@ -74093,2990 +74093,6 @@
#endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_small_kernel) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(7)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 1; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(7)
- .zero_index(mz)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_small_kernel) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(13)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 2; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(13)
- .zero_index(mz)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_small_kernel) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(17)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 3; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(17)
- .zero_index(mz)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_small_kernel) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(23)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 4; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(23)
- .zero_index(mz)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_small_kernel) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(7)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 1; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(7)
- .zero_index(mz)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_small_kernel) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(13)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 2; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(13)
- .zero_index(mz)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_small_kernel) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(17)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 3; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(17)
- .zero_index(mz)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_small_kernel) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(23)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 4; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(23)
- .zero_index(mz)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
- }
- }
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
-}
-
-
TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_FMAGIC, k_eq_1) {
GemmMicrokernelTester()
.mr(1)
@@ -80059,3 +77075,5971 @@
.cm_stride(7)
.Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_fmagic, xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
}
+
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 1; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 2; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 3; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 4; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 1; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 2; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 3; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 4; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 1; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 2; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 3; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 4; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 1; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 2; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 3; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 4; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
+
+TEST(QS8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+}
diff --git a/test/qs8-igemm-minmax-fp32.yaml b/test/qs8-igemm-minmax-fp32.yaml
index d81de8c..295c945 100644
--- a/test/qs8-igemm-minmax-fp32.yaml
+++ b/test/qs8-igemm-minmax-fp32.yaml
@@ -499,30 +499,6 @@
- name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16_ld128
init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params
k-block: 8
-- name: xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf
- init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
- name: xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_fmagic
init: xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params
k-block: 1
@@ -547,3 +523,51 @@
- name: xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_fmagic
init: xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params
k-block: 1
+- name: xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf
+ init: xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
diff --git a/test/qu8-dwconv-minmax-fp32.cc b/test/qu8-dwconv-minmax-fp32.cc
index 71b8ab4..ec3afde 100644
--- a/test/qu8-dwconv-minmax-fp32.cc
+++ b/test/qu8-dwconv-minmax-fp32.cc
@@ -5961,546 +5961,6 @@
#endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_eq_1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(1)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_gt_1) {
- for (uint32_t channels = 2; channels < 10; channels++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_gt_1_with_qmin) {
- for (uint32_t channels = 2; channels < 10; channels++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_gt_1_with_qmax) {
- for (uint32_t channels = 2; channels < 10; channels++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .width(3)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_step) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- for (size_t step = 2; step <= 9; step++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .width(3)
- .step(step)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_output_stride) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(1)
- .width(5)
- .output_stride(7)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_qmin) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .width(3)
- .qmin(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_qmax) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .width(3)
- .qmax(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, input_zero_point_only) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .width(3)
- .input_zero_point(255)
- .kernel_zero_point(0)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, kernel_zero_point_only) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .width(3)
- .input_zero_point(0)
- .kernel_zero_point(255)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, input_offset) {
- for (uint32_t channels = 2; channels < 16; channels += 3) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .input_offset(48)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 9; mz++) {
- for (uint32_t channels = 2; channels < 16; channels += 3) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(9)
- .channels(channels)
- .input_offset(48)
- .zero_index(mz)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_eq_2) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(2)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_div_2) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_div_2_with_qmin) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_div_2_with_qmax) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_lt_2) {
- for (uint32_t channels = 1; channels < 2; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_gt_2) {
- for (uint32_t channels = 3; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_gt_2_with_qmin) {
- for (uint32_t channels = 3; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_gt_2_with_qmax) {
- for (uint32_t channels = 3; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .width(3)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_step) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- for (size_t step = 2; step <= 9; step++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .width(3)
- .step(step)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_output_stride) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(2)
- .width(5)
- .output_stride(13)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_qmin) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .width(3)
- .qmin(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_qmax) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .width(3)
- .qmax(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, input_zero_point_only) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .width(3)
- .input_zero_point(255)
- .kernel_zero_point(0)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, kernel_zero_point_only) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .width(3)
- .input_zero_point(0)
- .kernel_zero_point(255)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, input_offset) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .input_offset(80)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 9; mz++) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(9)
- .channels(channels)
- .input_offset(80)
- .zero_index(mz)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_eq_4) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(4)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_div_4) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_div_4_with_qmin) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_div_4_with_qmax) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_lt_4) {
- for (uint32_t channels = 1; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_gt_4) {
- for (uint32_t channels = 5; channels < 8; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_gt_4_with_qmin) {
- for (uint32_t channels = 5; channels < 8; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_gt_4_with_qmax) {
- for (uint32_t channels = 5; channels < 8; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .width(3)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_step) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- for (size_t step = 2; step <= 9; step++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .width(3)
- .step(step)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_output_stride) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(4)
- .width(5)
- .output_stride(23)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_qmin) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .width(3)
- .qmin(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_qmax) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .width(3)
- .qmax(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, input_zero_point_only) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .width(3)
- .input_zero_point(255)
- .kernel_zero_point(0)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, kernel_zero_point_only) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .width(3)
- .input_zero_point(0)
- .kernel_zero_point(255)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, input_offset) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .input_offset(112)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 9; mz++) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(9)
- .channels(channels)
- .input_offset(112)
- .zero_index(mz)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_FMAGIC, c_eq_1) {
DWConvMicrokernelTester()
.cr(1)
@@ -7041,6 +6501,1086 @@
}
}
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, c_eq_1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(1)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, c_gt_1) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, c_gt_1_with_qmin) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, c_gt_1_with_qmax) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, multipixel) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ for (size_t step = 2; step <= 9; step++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(1)
+ .width(5)
+ .output_stride(7)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, input_zero_point_only) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(255)
+ .kernel_zero_point(0)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, kernel_zero_point_only) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(0)
+ .kernel_zero_point(255)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, input_offset) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .input_offset(48)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 9; mz++) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .input_offset(48)
+ .zero_index(mz)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_eq_2) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(2)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_div_2) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_div_2_with_qmin) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_div_2_with_qmax) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_lt_2) {
+ for (uint32_t channels = 1; channels < 2; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_gt_2) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_gt_2_with_qmin) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, c_gt_2_with_qmax) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, multipixel) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ for (size_t step = 2; step <= 9; step++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(2)
+ .width(5)
+ .output_stride(13)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, input_zero_point_only) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(255)
+ .kernel_zero_point(0)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, kernel_zero_point_only) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(0)
+ .kernel_zero_point(255)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, input_offset) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .input_offset(80)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 9; mz++) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .input_offset(80)
+ .zero_index(mz)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_eq_4) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(4)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_div_4) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_div_4_with_qmin) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_div_4_with_qmax) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_lt_4) {
+ for (uint32_t channels = 1; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_gt_4) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_gt_4_with_qmin) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, c_gt_4_with_qmax) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, multipixel) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ for (size_t step = 2; step <= 9; step++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(4)
+ .width(5)
+ .output_stride(23)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, input_zero_point_only) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(255)
+ .kernel_zero_point(0)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, kernel_zero_point_only) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(0)
+ .kernel_zero_point(255)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, input_offset) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .input_offset(112)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 9; mz++) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .input_offset(112)
+ .zero_index(mz)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_eq_1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(1)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_gt_1) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_gt_1_with_qmin) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, c_gt_1_with_qmax) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ for (size_t step = 2; step <= 9; step++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(1)
+ .width(5)
+ .output_stride(7)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, input_zero_point_only) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(255)
+ .kernel_zero_point(0)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, kernel_zero_point_only) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(0)
+ .kernel_zero_point(255)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, input_offset) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .input_offset(48)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X9__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 9; mz++) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(9)
+ .channels(channels)
+ .input_offset(48)
+ .zero_index(mz)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_eq_2) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(2)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_div_2) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_div_2_with_qmin) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_div_2_with_qmax) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_lt_2) {
+ for (uint32_t channels = 1; channels < 2; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_gt_2) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_gt_2_with_qmin) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, c_gt_2_with_qmax) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ for (size_t step = 2; step <= 9; step++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(2)
+ .width(5)
+ .output_stride(13)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, input_zero_point_only) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(255)
+ .kernel_zero_point(0)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, kernel_zero_point_only) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(0)
+ .kernel_zero_point(255)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, input_offset) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .input_offset(80)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X9__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 9; mz++) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(9)
+ .channels(channels)
+ .input_offset(80)
+ .zero_index(mz)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_eq_4) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(4)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_div_4) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_div_4_with_qmin) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_div_4_with_qmax) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_lt_4) {
+ for (uint32_t channels = 1; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_gt_4) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_gt_4_with_qmin) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, c_gt_4_with_qmax) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ for (size_t step = 2; step <= 9; step++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(4)
+ .width(5)
+ .output_stride(23)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, input_zero_point_only) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(255)
+ .kernel_zero_point(0)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, kernel_zero_point_only) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(0)
+ .kernel_zero_point(255)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, input_offset) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .input_offset(112)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X9__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 9; mz++) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(9)
+ .channels(channels)
+ .input_offset(112)
+ .zero_index(mz)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
#if XNN_ARCH_ARM || XNN_ARCH_ARM64
TEST(QU8_DWCONV_MINMAX_FP32_UP8X25__NEON_MUL16, c_eq_8) {
TEST_REQUIRES_ARM_NEON;
@@ -12982,546 +13522,6 @@
#endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_eq_1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(1)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_gt_1) {
- for (uint32_t channels = 2; channels < 10; channels++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_gt_1_with_qmin) {
- for (uint32_t channels = 2; channels < 10; channels++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_gt_1_with_qmax) {
- for (uint32_t channels = 2; channels < 10; channels++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .width(3)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_step) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- for (size_t step = 2; step <= 25; step++) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .width(3)
- .step(step)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_output_stride) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(1)
- .width(5)
- .output_stride(7)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_qmin) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .width(3)
- .qmin(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_qmax) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .width(3)
- .qmax(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, input_zero_point_only) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .width(3)
- .input_zero_point(255)
- .kernel_zero_point(0)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, kernel_zero_point_only) {
- for (size_t channels = 1; channels <= 5; channels += 1) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .width(3)
- .input_zero_point(0)
- .kernel_zero_point(255)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, input_offset) {
- for (uint32_t channels = 2; channels < 16; channels += 3) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .input_offset(48)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 25; mz++) {
- for (uint32_t channels = 2; channels < 16; channels += 3) {
- DWConvMicrokernelTester()
- .cr(1)
- .kr(25)
- .channels(channels)
- .input_offset(48)
- .zero_index(mz)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_eq_2) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(2)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_div_2) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_div_2_with_qmin) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_div_2_with_qmax) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_lt_2) {
- for (uint32_t channels = 1; channels < 2; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_gt_2) {
- for (uint32_t channels = 3; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_gt_2_with_qmin) {
- for (uint32_t channels = 3; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_gt_2_with_qmax) {
- for (uint32_t channels = 3; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .width(3)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_step) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- for (size_t step = 2; step <= 25; step++) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .width(3)
- .step(step)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_output_stride) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(2)
- .width(5)
- .output_stride(13)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_qmin) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .width(3)
- .qmin(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_qmax) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .width(3)
- .qmax(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, input_zero_point_only) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .width(3)
- .input_zero_point(255)
- .kernel_zero_point(0)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, kernel_zero_point_only) {
- for (size_t channels = 1; channels <= 10; channels += 1) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .width(3)
- .input_zero_point(0)
- .kernel_zero_point(255)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, input_offset) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .input_offset(80)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 25; mz++) {
- for (uint32_t channels = 4; channels < 32; channels += 6) {
- DWConvMicrokernelTester()
- .cr(2)
- .kr(25)
- .channels(channels)
- .input_offset(80)
- .zero_index(mz)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_eq_4) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(4)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_div_4) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_div_4_with_qmin) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_div_4_with_qmax) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_lt_4) {
- for (uint32_t channels = 1; channels < 4; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_gt_4) {
- for (uint32_t channels = 5; channels < 8; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_gt_4_with_qmin) {
- for (uint32_t channels = 5; channels < 8; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .qmin(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_gt_4_with_qmax) {
- for (uint32_t channels = 5; channels < 8; channels++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .qmax(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .width(3)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_step) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- for (size_t step = 2; step <= 25; step++) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .width(3)
- .step(step)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_output_stride) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(4)
- .width(5)
- .output_stride(23)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_qmin) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .width(3)
- .qmin(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_qmax) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .width(3)
- .qmax(128)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, input_zero_point_only) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .width(3)
- .input_zero_point(255)
- .kernel_zero_point(0)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, kernel_zero_point_only) {
- for (size_t channels = 1; channels <= 20; channels += 3) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .width(3)
- .input_zero_point(0)
- .kernel_zero_point(255)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, input_offset) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .input_offset(112)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 25; mz++) {
- for (uint32_t channels = 8; channels < 64; channels += 12) {
- DWConvMicrokernelTester()
- .cr(4)
- .kr(25)
- .channels(channels)
- .input_offset(112)
- .zero_index(mz)
- .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_FMAGIC, c_eq_1) {
DWConvMicrokernelTester()
.cr(1)
@@ -14060,4 +14060,1084 @@
.Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_fmagic, xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
}
}
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, c_eq_1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(1)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, c_gt_1) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, c_gt_1_with_qmin) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, c_gt_1_with_qmax) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, multipixel) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ for (size_t step = 2; step <= 25; step++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(1)
+ .width(5)
+ .output_stride(7)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, input_zero_point_only) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(255)
+ .kernel_zero_point(0)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, kernel_zero_point_only) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(0)
+ .kernel_zero_point(255)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, input_offset) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .input_offset(48)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 25; mz++) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .input_offset(48)
+ .zero_index(mz)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_eq_2) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(2)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_div_2) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_div_2_with_qmin) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_div_2_with_qmax) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_lt_2) {
+ for (uint32_t channels = 1; channels < 2; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_gt_2) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_gt_2_with_qmin) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, c_gt_2_with_qmax) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, multipixel) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ for (size_t step = 2; step <= 25; step++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(2)
+ .width(5)
+ .output_stride(13)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, input_zero_point_only) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(255)
+ .kernel_zero_point(0)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, kernel_zero_point_only) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(0)
+ .kernel_zero_point(255)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, input_offset) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .input_offset(80)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 25; mz++) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .input_offset(80)
+ .zero_index(mz)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_eq_4) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(4)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_div_4) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_div_4_with_qmin) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_div_4_with_qmax) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_lt_4) {
+ for (uint32_t channels = 1; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_gt_4) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_gt_4_with_qmin) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, c_gt_4_with_qmax) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, multipixel) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ for (size_t step = 2; step <= 25; step++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(4)
+ .width(5)
+ .output_stride(23)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, input_zero_point_only) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(255)
+ .kernel_zero_point(0)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, kernel_zero_point_only) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(0)
+ .kernel_zero_point(255)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, input_offset) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .input_offset(112)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 25; mz++) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .input_offset(112)
+ .zero_index(mz)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_eq_1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(1)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_gt_1) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_gt_1_with_qmin) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, c_gt_1_with_qmax) {
+ for (uint32_t channels = 2; channels < 10; channels++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ for (size_t step = 2; step <= 25; step++) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(1)
+ .width(5)
+ .output_stride(7)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, input_zero_point_only) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(255)
+ .kernel_zero_point(0)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, kernel_zero_point_only) {
+ for (size_t channels = 1; channels <= 5; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(0)
+ .kernel_zero_point(255)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, input_offset) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .input_offset(48)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP1X25__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 25; mz++) {
+ for (uint32_t channels = 2; channels < 16; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(1)
+ .kr(25)
+ .channels(channels)
+ .input_offset(48)
+ .zero_index(mz)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_eq_2) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(2)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_div_2) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_div_2_with_qmin) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_div_2_with_qmax) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_lt_2) {
+ for (uint32_t channels = 1; channels < 2; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_gt_2) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_gt_2_with_qmin) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, c_gt_2_with_qmax) {
+ for (uint32_t channels = 3; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ for (size_t step = 2; step <= 25; step++) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(2)
+ .width(5)
+ .output_stride(13)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, input_zero_point_only) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(255)
+ .kernel_zero_point(0)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, kernel_zero_point_only) {
+ for (size_t channels = 1; channels <= 10; channels += 1) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(0)
+ .kernel_zero_point(255)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, input_offset) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .input_offset(80)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP2X25__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 25; mz++) {
+ for (uint32_t channels = 4; channels < 32; channels += 6) {
+ DWConvMicrokernelTester()
+ .cr(2)
+ .kr(25)
+ .channels(channels)
+ .input_offset(80)
+ .zero_index(mz)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_eq_4) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(4)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_div_4) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_div_4_with_qmin) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_div_4_with_qmax) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_lt_4) {
+ for (uint32_t channels = 1; channels < 4; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_gt_4) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_gt_4_with_qmin) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, c_gt_4_with_qmax) {
+ for (uint32_t channels = 5; channels < 8; channels++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_step) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ for (size_t step = 2; step <= 25; step++) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .step(step)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_output_stride) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(4)
+ .width(5)
+ .output_stride(23)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_qmin) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmin(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, multipixel_with_qmax) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .qmax(128)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, input_zero_point_only) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(255)
+ .kernel_zero_point(0)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, kernel_zero_point_only) {
+ for (size_t channels = 1; channels <= 20; channels += 3) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .width(3)
+ .input_zero_point(0)
+ .kernel_zero_point(255)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, input_offset) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .input_offset(112)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_DWCONV_MINMAX_FP32_UP4X25__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 25; mz++) {
+ for (uint32_t channels = 8; channels < 64; channels += 12) {
+ DWConvMicrokernelTester()
+ .cr(4)
+ .kr(25)
+ .channels(channels)
+ .input_offset(112)
+ .zero_index(mz)
+ .Test(xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
}
\ No newline at end of file
diff --git a/test/qu8-dwconv-minmax-fp32.yaml b/test/qu8-dwconv-minmax-fp32.yaml
index 9ac55fd..fa1b330 100644
--- a/test/qu8-dwconv-minmax-fp32.yaml
+++ b/test/qu8-dwconv-minmax-fp32.yaml
@@ -58,18 +58,24 @@
init: xnn_init_qu8_conv_minmax_fp32_wasmsimd_params
- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__wasmsimd_mul16
init: xnn_init_qu8_conv_minmax_fp32_wasmsimd_params
-- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
-- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
-- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_fmagic
init: xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params
- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_fmagic
init: xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params
- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_fmagic
init: xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params
+- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16
init: xnn_init_qu8_conv_minmax_fp32_neon_params
- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul16
@@ -126,15 +132,21 @@
init: xnn_init_qu8_conv_minmax_fp32_wasmsimd_params
- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up24x25__wasmsimd_mul16
init: xnn_init_qu8_conv_minmax_fp32_wasmsimd_params
-- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
-- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
-- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_fmagic
init: xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params
- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_fmagic
init: xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params
- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_fmagic
init: xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params
+- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+- name: xnn_qu8_dwconv_minmax_fp32_ukernel_up4x25__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
diff --git a/test/qu8-gemm-minmax-fp32.cc b/test/qu8-gemm-minmax-fp32.cc
index 37c5a2f..b833f56 100644
--- a/test/qu8-gemm-minmax-fp32.cc
+++ b/test/qu8-gemm-minmax-fp32.cc
@@ -47843,3014 +47843,6 @@
#endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .a_stride(3)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .a_stride(11)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_strided_a) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_strided_a) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, no_a_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .a_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, no_b_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .b_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, no_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .a_zero_point(0)
- .b_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .a_stride(3)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .a_stride(11)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_strided_a) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_strided_a) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, no_a_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .a_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, no_b_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .b_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, no_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .a_zero_point(0)
- .b_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .a_stride(3)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .a_stride(11)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_strided_a) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_strided_a) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, no_a_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .a_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, no_b_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .b_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, no_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .a_zero_point(0)
- .b_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .a_stride(3)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .a_stride(11)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_strided_a) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_strided_a) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, no_a_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .a_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, no_b_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .b_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, no_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .a_zero_point(0)
- .b_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .a_stride(3)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .a_stride(11)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_strided_a) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_strided_a) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, no_a_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .a_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, no_b_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .b_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, no_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .a_zero_point(0)
- .b_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .a_stride(3)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .a_stride(11)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_strided_a) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_strided_a) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, no_a_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .a_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, no_b_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .b_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, no_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .a_zero_point(0)
- .b_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .a_stride(3)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .a_stride(11)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_strided_a) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_strided_a) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, no_a_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .a_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, no_b_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .b_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, no_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .a_zero_point(0)
- .b_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_strided_a) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .a_stride(3)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1_strided_a) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .a_stride(11)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_strided_a) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_strided_a) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .a_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, no_a_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .a_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, no_b_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .b_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, no_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .a_zero_point(0)
- .b_zero_point(0)
- .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_FMAGIC, k_eq_1) {
GemmMicrokernelTester()
.mr(1)
@@ -53857,4 +50849,6020 @@
.b_zero_point(0)
.Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_fmagic, xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
}
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_strided_a) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_strided_a) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_strided_a) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .a_stride(3)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1_strided_a) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .a_stride(11)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_strided_a) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_strided_a) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .a_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_GEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
}
\ No newline at end of file
diff --git a/test/qu8-gemm-minmax-fp32.yaml b/test/qu8-gemm-minmax-fp32.yaml
index 7e52b9a..62bfe78 100644
--- a/test/qu8-gemm-minmax-fp32.yaml
+++ b/test/qu8-gemm-minmax-fp32.yaml
@@ -292,30 +292,6 @@
- name: xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul32_ld128
init: xnn_init_qu8_conv_minmax_fp32_wasmsimd_params
k-block: 8
-- name: xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
- name: xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_fmagic
init: xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params
k-block: 1
@@ -340,3 +316,51 @@
- name: xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_fmagic
init: xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params
k-block: 1
+- name: xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
diff --git a/test/qu8-igemm-minmax-fp32.cc b/test/qu8-igemm-minmax-fp32.cc
index 3b5506c..de8824e 100644
--- a/test/qu8-igemm-minmax-fp32.cc
+++ b/test/qu8-igemm-minmax-fp32.cc
@@ -48995,3350 +48995,6 @@
#endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_small_kernel) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(7)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 1; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(7)
- .zero_index(mz)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, no_a_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .a_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, no_b_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .b_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, no_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(1)
- .n(2)
- .k(k)
- .a_zero_point(0)
- .b_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_small_kernel) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(13)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 2; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(13)
- .zero_index(mz)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, no_a_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .a_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, no_b_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .b_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, no_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(2)
- .n(2)
- .k(k)
- .a_zero_point(0)
- .b_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_small_kernel) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(17)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 3; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(17)
- .zero_index(mz)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, no_a_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .a_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, no_b_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .b_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, no_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(3)
- .n(2)
- .k(k)
- .a_zero_point(0)
- .b_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .cn_stride(5)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(2)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_subtile) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_strided_cn) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .cn_stride(5)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_subtile) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
- for (uint32_t n = 3; n < 4; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_small_kernel) {
- for (uint32_t n = 4; n <= 6; n += 2) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 2; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(5)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(23)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 4; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .ks(3)
- .a_offset(23)
- .zero_index(mz)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .qmin(128)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .qmax(128)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(1)
- .cm_stride(5)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, no_a_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .a_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, no_b_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .b_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, no_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(2)
- .kr(1)
- .sr(1)
- .m(4)
- .n(2)
- .k(k)
- .a_zero_point(0)
- .b_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_small_kernel) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(7)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 1; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(7)
- .zero_index(mz)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, no_a_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .a_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, no_b_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .b_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, no_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(1)
- .n(4)
- .k(k)
- .a_zero_point(0)
- .b_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_small_kernel) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(13)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 2; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(13)
- .zero_index(mz)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, no_a_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .a_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, no_b_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .b_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, no_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(2)
- .n(4)
- .k(k)
- .a_zero_point(0)
- .b_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_small_kernel) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 3; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(17)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 3; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(17)
- .zero_index(mz)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, no_a_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .a_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, no_b_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .b_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, no_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(3)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(3)
- .n(4)
- .k(k)
- .a_zero_point(0)
- .b_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cn) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .cn_stride(7)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(4)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(1)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1) {
- for (size_t k = 2; k < 10; k++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1_subtile) {
- for (size_t k = 2; k < 10; k++) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_subtile) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_strided_cn) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(n)
- .k(k)
- .cn_stride(7)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_subtile) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, small_kernel) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, small_kernel_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
- for (uint32_t n = 5; n < 8; n++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_small_kernel) {
- for (uint32_t n = 8; n <= 12; n += 4) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .ks(3)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm_subtile) {
- for (size_t k = 1; k <= 5; k += 2) {
- for (uint32_t m = 1; m <= 4; m++) {
- for (uint32_t n = 1; n <= 4; n++) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(7)
- .iterations(1)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, a_offset) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(23)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, zero) {
- for (uint32_t mz = 0; mz < 4; mz++) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .ks(3)
- .a_offset(23)
- .zero_index(mz)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmin) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .qmin(128)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmax) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .qmax(128)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(1)
- .cm_stride(7)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, no_a_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .a_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, no_b_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .b_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
-TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, no_zero_point) {
- for (size_t k = 1; k <= 5; k += 2) {
- GemmMicrokernelTester()
- .mr(4)
- .nr(4)
- .kr(1)
- .sr(1)
- .m(4)
- .n(4)
- .k(k)
- .a_zero_point(0)
- .b_zero_point(0)
- .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
- }
-}
-
TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_FMAGIC, k_eq_1) {
GemmMicrokernelTester()
.mr(1)
@@ -55681,4 +52337,6692 @@
.b_zero_point(0)
.Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_fmagic, xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
}
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 1; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .zero_index(mz)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_IMAGIC, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 2; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .zero_index(mz)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_IMAGIC, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 3; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .zero_index(mz)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_IMAGIC, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 4; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .zero_index(mz)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_IMAGIC, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 1; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .zero_index(mz)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_IMAGIC, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 2; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .zero_index(mz)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_IMAGIC, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 3; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .zero_index(mz)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_IMAGIC, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, zero) {
+ for (uint32_t mz = 0; mz < 4; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .zero_index(mz)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_IMAGIC, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic, xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 1; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .zero_index(mz)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X2__SCALAR_LRINTF, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 2; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .zero_index(mz)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X2__SCALAR_LRINTF, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 3; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .zero_index(mz)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X2__SCALAR_LRINTF, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(2)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_strided_cn) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_subtile) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_strided_cn) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_subtile) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_gt_2_small_kernel) {
+ for (uint32_t n = 3; n < 4; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, n_div_2_small_kernel) {
+ for (uint32_t n = 4; n <= 6; n += 2) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 2; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(5)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 4; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .zero_index(mz)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(1)
+ .cm_stride(5)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X2__SCALAR_LRINTF, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(2)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(2)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 1; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(7)
+ .zero_index(mz)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_1X4__SCALAR_LRINTF, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(1)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 2; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(13)
+ .zero_index(mz)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_2X4__SCALAR_LRINTF, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(2)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 3; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 3; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(17)
+ .zero_index(mz)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_3X4__SCALAR_LRINTF, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(3)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(3)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cn) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_m) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(4)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_eq_1_subtile_n) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(1)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1) {
+ for (size_t k = 2; k < 10; k++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, k_gt_1_subtile) {
+ for (size_t k = 2; k < 10; k++) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_strided_cn) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_subtile) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_strided_cn) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(n)
+ .k(k)
+ .cn_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_subtile) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, small_kernel) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, small_kernel_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_gt_4_small_kernel) {
+ for (uint32_t n = 5; n < 8; n++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, n_div_4_small_kernel) {
+ for (uint32_t n = 8; n <= 12; n += 4) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm_subtile) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ for (uint32_t m = 1; m <= 4; m++) {
+ for (uint32_t n = 1; n <= 4; n++) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(7)
+ .iterations(1)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, a_offset) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, zero) {
+ for (uint32_t mz = 0; mz < 4; mz++) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .ks(3)
+ .a_offset(23)
+ .zero_index(mz)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmin) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmin(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, qmax) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .qmax(128)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, strided_cm) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(1)
+ .cm_stride(7)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, no_a_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, no_b_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
+}
+
+TEST(QU8_IGEMM_MINMAX_FP32_4X4__SCALAR_LRINTF, no_zero_point) {
+ for (size_t k = 1; k <= 5; k += 2) {
+ GemmMicrokernelTester()
+ .mr(4)
+ .nr(4)
+ .kr(1)
+ .sr(1)
+ .m(4)
+ .n(4)
+ .k(k)
+ .a_zero_point(0)
+ .b_zero_point(0)
+ .Test(xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf, xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, xnn_init_qu8_requantization_fp32_params, xnn_qu8_requantize_fp32);
+ }
}
\ No newline at end of file
diff --git a/test/qu8-igemm-minmax-fp32.yaml b/test/qu8-igemm-minmax-fp32.yaml
index e38e345..9395b71 100644
--- a/test/qu8-igemm-minmax-fp32.yaml
+++ b/test/qu8-igemm-minmax-fp32.yaml
@@ -292,30 +292,6 @@
- name: xnn_qu8_igemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul32_ld128
init: xnn_init_qu8_conv_minmax_fp32_wasmsimd_params
k-block: 8
-- name: xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
-- name: xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf
- init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
- k-block: 1
- name: xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_fmagic
init: xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params
k-block: 1
@@ -340,3 +316,51 @@
- name: xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_fmagic
init: xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params
k-block: 1
+- name: xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_imagic
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params
+ k-block: 1
+- name: xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qu8_igemm_minmax_fp32_ukernel_3x2__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qu8_igemm_minmax_fp32_ukernel_4x2__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qu8_igemm_minmax_fp32_ukernel_2x4__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1
+- name: xnn_qu8_igemm_minmax_fp32_ukernel_4x4__scalar_lrintf
+ init: xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params
+ k-block: 1