implement qs8 x8c8 pack using avxvnniint8

google · Sep 26, 2024 · 2c0a33a · 2c0a33a
1 parent 467338f
commit 2c0a33a
Show file tree

Hide file tree

Showing 8 changed files with 848 additions and 1 deletion.
diff --git a/cmake/gen/avxvnniint8_microkernels.cmake b/cmake/gen/avxvnniint8_microkernels.cmake
@@ -15,6 +15,7 @@ SET(PROD_AVXVNNIINT8_MICROKERNEL_SRCS
   src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x8c8-minmax-fp32-avxvnniint8-prfm.c
   src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-5x8c8-minmax-fp32-avxvnniint8-prfm.c)
 
-SET(NON_PROD_AVXVNNIINT8_MICROKERNEL_SRCS)
+SET(NON_PROD_AVXVNNIINT8_MICROKERNEL_SRCS
+  src/qs8-packw/gen/qs8-packw-x8c8-gemm-goi-avxvnniint8.c)
 
 SET(ALL_AVXVNNIINT8_MICROKERNEL_SRCS ${PROD_AVXVNNIINT8_MICROKERNEL_SRCS} + ${NON_PROD_AVXVNNIINT8_MICROKERNEL_SRCS})
diff --git a/gen/avxvnniint8_microkernels.bzl b/gen/avxvnniint8_microkernels.bzl
@@ -13,6 +13,7 @@ PROD_AVXVNNIINT8_MICROKERNEL_SRCS = [
 ]
 
 NON_PROD_AVXVNNIINT8_MICROKERNEL_SRCS = [
+    "src/qs8-packw/gen/qs8-packw-x8c8-gemm-goi-avxvnniint8.c",
 ]
 
 ALL_AVXVNNIINT8_MICROKERNEL_SRCS = PROD_AVXVNNIINT8_MICROKERNEL_SRCS + NON_PROD_AVXVNNIINT8_MICROKERNEL_SRCS
diff --git a/scripts/generate-x8-packw.sh b/scripts/generate-x8-packw.sh
@@ -23,4 +23,6 @@ tools/xngen src/x8-packw/kr-scalar.c.in -D NR=16 -D KR=4 -D TYPE=int8_t -o src/q
 tools/xngen src/x8-packw/kr-scalar.c.in -D NR=32 -D KR=4 -D TYPE=int8_t -o src/qs8-packw/gen/qs8-packw-x32c4-gemm-goi-scalar.c &
 tools/xngen src/x8-packw/kr-scalar.c.in -D NR=64 -D KR=4 -D TYPE=int8_t -o src/qs8-packw/gen/qs8-packw-x64c4-gemm-goi-scalar.c &
 
+tools/xngen src/x8-packw/kr-avxvnniint8.c.in -D NR=8 -D KR=8 -D TYPE=int8_t -o src/qs8-packw/gen/qs8-packw-x8c8-gemm-goi-avxvnniint8.c &
+
 wait
diff --git a/src/configs/hardware-config.c b/src/configs/hardware-config.c
@@ -324,6 +324,7 @@ static void init_hardware_config(void) {
     if (hardware_config.use_x86_avx256skx) hardware_config.arch_flags |= xnn_arch_x86_avx256skx;
     if (hardware_config.use_x86_avx256vnni) hardware_config.arch_flags |= xnn_arch_x86_avx256vnni;
     if (hardware_config.use_x86_avx256vnnigfni) hardware_config.arch_flags |= xnn_arch_x86_avx256vnnigfni;
+    if (hardware_config.use_x86_avxvnniint8) hardware_config.arch_flags |= xnn_arch_x86_avxvnniint8;
   #endif
   #if XNN_ARCH_RISCV
     if (hardware_config.use_riscv_vector) hardware_config.arch_flags |= xnn_arch_riscv_vector;