QS8/QU8 VMUL[C] microkernels in SSE2/SSE4.1/AVX implementation
PiperOrigin-RevId: 388242650
diff --git a/BUILD.bazel b/BUILD.bazel
index acb6048..aacf557 100644
--- a/BUILD.bazel
+++ b/BUILD.bazel
@@ -3659,6 +3659,10 @@
"src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
"src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
"src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
+ "src/qs8-vmul/gen/minmax-fp32-sse2-mul16-ld64-x8.c",
+ "src/qs8-vmul/gen/minmax-fp32-sse2-mul16-ld64-x16.c",
+ "src/qs8-vmulc/gen/minmax-fp32-sse2-mul16-ld64-x8.c",
+ "src/qs8-vmulc/gen/minmax-fp32-sse2-mul16-ld64-x16.c",
"src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
"src/qu8-avgpool/9x-minmax-sse2-c8.c",
"src/qu8-dwconv/gen/up8x9-minmax-fp32-sse2-mul16.c",
@@ -3706,6 +3710,10 @@
"src/qu8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
"src/qu8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
"src/qu8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
+ "src/qu8-vmul/gen/minmax-fp32-sse2-mul16-ld64-x8.c",
+ "src/qu8-vmul/gen/minmax-fp32-sse2-mul16-ld64-x16.c",
+ "src/qu8-vmulc/gen/minmax-fp32-sse2-mul16-ld64-x8.c",
+ "src/qu8-vmulc/gen/minmax-fp32-sse2-mul16-ld64-x16.c",
"src/u8-maxpool/9p8x-minmax-sse2-c16.c",
"src/u8-rmax/sse2.c",
"src/u8-vclamp/sse2-x64.c",
@@ -3995,6 +4003,10 @@
"src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
"src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
"src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
+ "src/qs8-vmul/gen/minmax-fp32-sse41-mul16-ld64-x8.c",
+ "src/qs8-vmul/gen/minmax-fp32-sse41-mul16-ld64-x16.c",
+ "src/qs8-vmulc/gen/minmax-fp32-sse41-mul16-ld64-x8.c",
+ "src/qs8-vmulc/gen/minmax-fp32-sse41-mul16-ld64-x16.c",
"src/qu8-dwconv/gen/up8x9-minmax-fp32-sse41-mul16.c",
"src/qu8-dwconv/gen/up8x9-minmax-fp32-sse41-mul32.c",
"src/qu8-dwconv/gen/up8x25-minmax-fp32-sse41-mul16.c",
@@ -4045,6 +4057,10 @@
"src/qu8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
"src/qu8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
"src/qu8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
+ "src/qu8-vmul/gen/minmax-fp32-sse41-mul16-ld64-x8.c",
+ "src/qu8-vmul/gen/minmax-fp32-sse41-mul16-ld64-x16.c",
+ "src/qu8-vmulc/gen/minmax-fp32-sse41-mul16-ld64-x8.c",
+ "src/qu8-vmulc/gen/minmax-fp32-sse41-mul16-ld64-x16.c",
]
PROD_AVX_MICROKERNEL_SRCS = [
@@ -4378,6 +4394,10 @@
"src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x16.c",
"src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x24.c",
"src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x32.c",
+ "src/qs8-vmul/gen/minmax-fp32-avx-mul16-ld64-x8.c",
+ "src/qs8-vmul/gen/minmax-fp32-avx-mul16-ld64-x16.c",
+ "src/qs8-vmulc/gen/minmax-fp32-avx-mul16-ld64-x8.c",
+ "src/qs8-vmulc/gen/minmax-fp32-avx-mul16-ld64-x16.c",
"src/qu8-dwconv/gen/up8x9-minmax-fp32-avx-mul16.c",
"src/qu8-dwconv/gen/up8x9-minmax-fp32-avx-mul32.c",
"src/qu8-dwconv/gen/up8x25-minmax-fp32-avx-mul16.c",
@@ -4422,6 +4442,10 @@
"src/qu8-vaddc/gen/minmax-avx-mul16-ld64-x16.c",
"src/qu8-vaddc/gen/minmax-avx-mul32-ld32-x8.c",
"src/qu8-vaddc/gen/minmax-avx-mul32-ld32-x16.c",
+ "src/qu8-vmul/gen/minmax-fp32-avx-mul16-ld64-x8.c",
+ "src/qu8-vmul/gen/minmax-fp32-avx-mul16-ld64-x16.c",
+ "src/qu8-vmulc/gen/minmax-fp32-avx-mul16-ld64-x8.c",
+ "src/qu8-vmulc/gen/minmax-fp32-avx-mul16-ld64-x16.c",
]
PROD_XOP_MICROKERNEL_SRCS = [
@@ -5613,6 +5637,7 @@
"src/xnnpack/unpool.h",
"src/xnnpack/vadd.h",
"src/xnnpack/vbinary.h",
+ "src/xnnpack/vmul.h",
"src/xnnpack/vmulcaddc.h",
"src/xnnpack/vscale.h",
"src/xnnpack/vscaleexpminusmax.h",
@@ -9282,6 +9307,24 @@
)
xnnpack_unit_test(
+ name = "qs8_vmul_minmax_fp32_test",
+ srcs = [
+ "test/qs8-vmul-minmax-fp32.cc",
+ "test/vmul-microkernel-tester.h",
+ ] + MICROKERNEL_TEST_HDRS,
+ deps = MICROKERNEL_TEST_DEPS,
+)
+
+xnnpack_unit_test(
+ name = "qs8_vmulc_minmax_fp32_test",
+ srcs = [
+ "test/qs8-vmulc-minmax-fp32.cc",
+ "test/vmulc-microkernel-tester.h",
+ ] + MICROKERNEL_TEST_HDRS,
+ deps = MICROKERNEL_TEST_DEPS,
+)
+
+xnnpack_unit_test(
name = "qu8_avgpool_minmax_test",
srcs = [
"test/qu8-avgpool-minmax.cc",
@@ -9410,6 +9453,24 @@
)
xnnpack_unit_test(
+ name = "qu8_vmul_minmax_fp32_test",
+ srcs = [
+ "test/qu8-vmul-minmax-fp32.cc",
+ "test/vmul-microkernel-tester.h",
+ ] + MICROKERNEL_TEST_HDRS,
+ deps = MICROKERNEL_TEST_DEPS,
+)
+
+xnnpack_unit_test(
+ name = "qu8_vmulc_minmax_fp32_test",
+ srcs = [
+ "test/qu8-vmulc-minmax-fp32.cc",
+ "test/vmulc-microkernel-tester.h",
+ ] + MICROKERNEL_TEST_HDRS,
+ deps = MICROKERNEL_TEST_DEPS,
+)
+
+xnnpack_unit_test(
name = "u8_lut32norm_test",
srcs = [
"test/u8-lut32norm.cc",