commit bb00b1dda26b55c80f7a737ad374dfc312bd92b0
author Marat Dukhan <maratek@google.com> Mon Aug 10 11:37:23 2020 -0700
committer XNNPACK Team <xnnpack-github-robot@google.com> Mon Aug 10 11:38:00 2020 -0700
tree 260af38d3e6c650055de9d4b11d2e95779c26653
parent facecc526bf35d09ac3319f44072d640101b2b70

AVX512 variants of QS8 GEMM and IGEMM microkernels

PiperOrigin-RevId: 325850791

CMakeLists.txt

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 293b79f..9497263 100755
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1977,6 +1977,16 @@
   src/math/sqrt-avx512f-nr2fma.c
   src/math/sqrt-avx512f-nr1fma1adj.c)
 
+SET(XNNPACK_AVX512SKX_MICROKERNEL_SRCS
+  src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c
+  src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c
+  src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c
+  src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c
+  src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c
+  src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c
+  src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c
+  src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c)
+
 SET(XNNPACK_AARCH32_ASM_MICROKERNEL_SRCS
   src/f32-gemm/4x4-aarch32-vfp-ld64.S
   src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S
@@ -2098,6 +2108,7 @@
   LIST(APPEND XNNPACK_MICROKERNEL_SRCS ${XNNPACK_FMA3_MICROKERNEL_SRCS})
   LIST(APPEND XNNPACK_MICROKERNEL_SRCS ${XNNPACK_AVX2_MICROKERNEL_SRCS})
   LIST(APPEND XNNPACK_MICROKERNEL_SRCS ${XNNPACK_AVX512F_MICROKERNEL_SRCS})
+  LIST(APPEND XNNPACK_MICROKERNEL_SRCS ${XNNPACK_AVX512SKX_MICROKERNEL_SRCS})
 ENDIF()
 
 IF(XNNPACK_LIBRARY_TYPE STREQUAL "default")
@@ -2142,6 +2153,7 @@
     SET_PROPERTY(SOURCE ${XNNPACK_FMA3_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " /arch:AVX ")
     SET_PROPERTY(SOURCE ${XNNPACK_AVX2_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " /arch:AVX2 ")
     SET_PROPERTY(SOURCE ${XNNPACK_AVX512F_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " /arch:AVX512 ")
+    SET_PROPERTY(SOURCE ${XNNPACK_AVX512SKX_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " /arch:AVX512 ")
   ELSE()
     SET_PROPERTY(SOURCE ${XNNPACK_PSIMD_ACCMATH_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -msse2 ")
     SET_PROPERTY(SOURCE ${XNNPACK_PSIMD_FASTMATH_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -msse2 ")
@@ -2154,9 +2166,11 @@
     SET_PROPERTY(SOURCE ${XNNPACK_FMA3_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -mfma ")
     SET_PROPERTY(SOURCE ${XNNPACK_AVX2_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -mfma -mavx2 ")
     SET_PROPERTY(SOURCE ${XNNPACK_AVX512F_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -mavx512f ")
+    SET_PROPERTY(SOURCE ${XNNPACK_AVX512SKX_MICROKERNEL_SRCS} APPEND_STRIDE PROPERTY COMPILE_FLAGS " -mavx512f -mavx512cd -mavx512bw -mavx512dq -mavx512vl ")
     IF(MINGW OR CMAKE_SYSTEM_NAME MATCHES "^(CYGWIN|MSYS)$")
       # Work-around for https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65782
       SET_PROPERTY(SOURCE ${XNNPACK_AVX512F_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -fno-asynchronous-unwind-tables ")
+      SET_PROPERTY(SOURCE ${XNNPACK_AVX512SKX_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -fno-asynchronous-unwind-tables ")
     ENDIF()
   ENDIF()
 ENDIF()