QS8 IGEMM microkernels and infrastructure
- QS8 IGEMM microkernels for SSE2/SSSE3/SSE4.1
- Updated GEMM tester to support QS8 IGEMM
- Updated weights packing functions to support QS8 IGEMM
PiperOrigin-RevId: 324289358
diff --git a/BUILD.bazel b/BUILD.bazel
index 86def78..6fa565a 100644
--- a/BUILD.bazel
+++ b/BUILD.bazel
@@ -1642,6 +1642,8 @@
"src/qs8-requantization/q31-sse2.c",
"src/qs8-gemm/gen/1x4c2-minmax-sse2.c",
"src/qs8-gemm/gen/4x4c2-minmax-sse2.c",
+ "src/qs8-igemm/gen/1x4c2-minmax-sse2.c",
+ "src/qs8-igemm/gen/4x4c2-minmax-sse2.c",
"src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
"src/qu8-avgpool/9x-minmax-sse2-c8.c",
"src/qu8-igemm/4x4c2-minmax-sse2.c",
@@ -1678,6 +1680,8 @@
SSSE3_UKERNELS = [
"src/qs8-gemm/gen/1x4c2-minmax-ssse3.c",
"src/qs8-gemm/gen/4x4c2-minmax-ssse3.c",
+ "src/qs8-igemm/gen/1x4c2-minmax-ssse3.c",
+ "src/qs8-igemm/gen/4x4c2-minmax-ssse3.c",
"src/qs8-requantization/precise-ssse3.c",
"src/qs8-requantization/q31-ssse3.c",
"src/qu8-requantization/precise-ssse3.c",
@@ -1705,6 +1709,8 @@
"src/f32-vrnd/gen/vrndd-sse41-x8.c",
"src/qs8-gemm/gen/1x4c2-minmax-sse41.c",
"src/qs8-gemm/gen/4x4c2-minmax-sse41.c",
+ "src/qs8-igemm/gen/1x4c2-minmax-sse41.c",
+ "src/qs8-igemm/gen/4x4c2-minmax-sse41.c",
"src/qs8-requantization/fp32-sse4.c",
"src/qs8-requantization/precise-sse4.c",
"src/qs8-requantization/q31-sse4.c",
@@ -5007,16 +5013,6 @@
)
xnnpack_unit_test(
- name = "qs8_requantization_test",
- srcs = [
- "src/xnnpack/requantization-stubs.h",
- "test/qs8-requantization.cc",
- "test/requantization-tester.h",
- ] + MICROKERNEL_TEST_HDRS,
- deps = MICROKERNEL_TEST_DEPS,
-)
-
-xnnpack_unit_test(
name = "qs8_gemm_minmax_test",
srcs = [
"test/qs8-gemm-minmax.cc",
@@ -5027,6 +5023,26 @@
)
xnnpack_unit_test(
+ name = "qs8_igemm_minmax_test",
+ srcs = [
+ "test/qs8-igemm-minmax.cc",
+ "test/gemm-microkernel-tester.h",
+ "src/xnnpack/AlignedAllocator.h",
+ ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
+ deps = MICROKERNEL_TEST_DEPS + [":packing"],
+)
+
+xnnpack_unit_test(
+ name = "qs8_requantization_test",
+ srcs = [
+ "src/xnnpack/requantization-stubs.h",
+ "test/qs8-requantization.cc",
+ "test/requantization-tester.h",
+ ] + MICROKERNEL_TEST_HDRS,
+ deps = MICROKERNEL_TEST_DEPS,
+)
+
+xnnpack_unit_test(
name = "qu8_avgpool_minmax_test",
srcs = [
"test/qu8-avgpool-minmax.cc",