Support qs8 5x8c8 wasmdot microkernels

yolanda15 · Oct 9, 2024 · 3b45ce1 · 3b45ce1
1 parent 14ce5c4
commit 3b45ce1
Show file tree

Hide file tree

Showing 21 changed files with 2,889 additions and 53 deletions.
diff --git a/bench/qs8-qc8w-gemm-fp32.cc b/bench/qs8-qc8w-gemm-fp32.cc
@@ -4093,6 +4093,28 @@
 
   BENCHMARK_GEMM(qs8_qc8w_gemm_minmax_fp32_ukernel_4x8c8__wasmusdot_u2)
 
+  static void qs8_qc8w_gemm_minmax_fp32_ukernel_5x8c8__wasmusdot(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qs8_qc8w_gemm_minmax_fp32_ukernel_5x8c8__wasmusdot,
+      xnn_init_qs8_qc8w_conv_minmax_fp32_scalar_params,
+      xnn_pack_qs8_to_qu8_gemm_goi_w,
+      /*mr=*/5, /*nr=*/8, /*kr=*/8, /*sr=*/1,
+      benchmark::utils::CheckWAsmUSDOT);
+  }
+
+  BENCHMARK_GEMM(qs8_qc8w_gemm_minmax_fp32_ukernel_5x8c8__wasmusdot)
+
+  static void qs8_qc8w_gemm_minmax_fp32_ukernel_5x8c8__wasmusdot_u2(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qs8_qc8w_gemm_minmax_fp32_ukernel_5x8c8__wasmusdot_u2,
+      xnn_init_qs8_qc8w_conv_minmax_fp32_scalar_params,
+      xnn_pack_qs8_to_qu8_gemm_goi_w,
+      /*mr=*/5, /*nr=*/8, /*kr=*/8, /*sr=*/1,
+      benchmark::utils::CheckWAsmUSDOT);
+  }
+
+  BENCHMARK_GEMM(qs8_qc8w_gemm_minmax_fp32_ukernel_5x8c8__wasmusdot_u2)
+
   static void qs8_qc8w_gemm_minmax_fp32_ukernel_1x8c8__wasmsdot(benchmark::State& state, const char* net) {
     GEMMBenchmark(state,
       xnn_qs8_qc8w_gemm_minmax_fp32_ukernel_1x8c8__wasmsdot,
@@ -4180,6 +4202,28 @@
   }
 
   BENCHMARK_GEMM(qs8_qc8w_gemm_minmax_fp32_ukernel_4x8c8__wasmsdot_u2)
+
+  static void qs8_qc8w_gemm_minmax_fp32_ukernel_5x8c8__wasmsdot(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qs8_qc8w_gemm_minmax_fp32_ukernel_5x8c8__wasmsdot,
+      xnn_init_qs8_qc8w_conv_minmax_fp32_scalar_params,
+      xnn_pack_qs8_gemm_goi_w,
+      /*mr=*/5, /*nr=*/8, /*kr=*/8, /*sr=*/1,
+      benchmark::utils::CheckWAsmSDOT);
+  }
+
+  BENCHMARK_GEMM(qs8_qc8w_gemm_minmax_fp32_ukernel_5x8c8__wasmsdot)
+
+  static void qs8_qc8w_gemm_minmax_fp32_ukernel_5x8c8__wasmsdot_u2(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qs8_qc8w_gemm_minmax_fp32_ukernel_5x8c8__wasmsdot_u2,
+      xnn_init_qs8_qc8w_conv_minmax_fp32_scalar_params,
+      xnn_pack_qs8_gemm_goi_w,
+      /*mr=*/5, /*nr=*/8, /*kr=*/8, /*sr=*/1,
+      benchmark::utils::CheckWAsmSDOT);
+  }
+
+  BENCHMARK_GEMM(qs8_qc8w_gemm_minmax_fp32_ukernel_5x8c8__wasmsdot_u2)
 #endif  // XNN_ARCH_WASMRELAXEDSIMD
 
 

diff --git a/cmake/gen/wasmrelaxedsimd_microkernels.cmake b/cmake/gen/wasmrelaxedsimd_microkernels.cmake
@@ -530,6 +530,10 @@ SET(NON_PROD_WASMRELAXEDSIMD_MICROKERNEL_SRCS
   src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-4x8c8-minmax-fp32-wasmusdot.c
   src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-4x8c16-minmax-fp32-wasmsdot.c
   src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-4x8c16-minmax-fp32-wasmusdot.c
+  src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-5x8c8-minmax-fp32-wasmsdot-u2.c
+  src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-5x8c8-minmax-fp32-wasmsdot.c
+  src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-5x8c8-minmax-fp32-wasmusdot-u2.c
+  src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-5x8c8-minmax-fp32-wasmusdot.c
   src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x4c16-minmax-fp32-wasmsdot.c
   src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x4c16-minmax-fp32-wasmusdot.c
   src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x8c8-minmax-fp32-wasmusdot.c
@@ -556,6 +560,10 @@ SET(NON_PROD_WASMRELAXEDSIMD_MICROKERNEL_SRCS
   src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x8c8-minmax-fp32-wasmusdot.c
   src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x8c16-minmax-fp32-wasmsdot.c
   src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x8c16-minmax-fp32-wasmusdot.c
+  src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-5x8c8-minmax-fp32-wasmsdot-u2.c
+  src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-5x8c8-minmax-fp32-wasmsdot.c
+  src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-5x8c8-minmax-fp32-wasmusdot-u2.c
+  src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-5x8c8-minmax-fp32-wasmusdot.c
   src/qs8-vcvt/gen/qs8-vcvt-wasmrelaxedsimd-u8.c
   src/qs8-vcvt/gen/qs8-vcvt-wasmrelaxedsimd-u16.c
   src/qs8-vlrelu/gen/qs8-vlrelu-wasmrelaxedsimd-arm-u16.c

diff --git a/gen/wasmrelaxedsimd_microkernels.bzl b/gen/wasmrelaxedsimd_microkernels.bzl
@@ -527,6 +527,10 @@ NON_PROD_WASMRELAXEDSIMD_MICROKERNEL_SRCS = [
     "src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-4x8c8-minmax-fp32-wasmusdot.c",
     "src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-4x8c16-minmax-fp32-wasmsdot.c",
     "src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-4x8c16-minmax-fp32-wasmusdot.c",
+    "src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-5x8c8-minmax-fp32-wasmsdot-u2.c",
+    "src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-5x8c8-minmax-fp32-wasmsdot.c",
+    "src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-5x8c8-minmax-fp32-wasmusdot-u2.c",
+    "src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-5x8c8-minmax-fp32-wasmusdot.c",
     "src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x4c16-minmax-fp32-wasmsdot.c",
     "src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x4c16-minmax-fp32-wasmusdot.c",
     "src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x8c8-minmax-fp32-wasmusdot.c",
@@ -553,6 +557,10 @@ NON_PROD_WASMRELAXEDSIMD_MICROKERNEL_SRCS = [
     "src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x8c8-minmax-fp32-wasmusdot.c",
     "src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x8c16-minmax-fp32-wasmsdot.c",
     "src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x8c16-minmax-fp32-wasmusdot.c",
+    "src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-5x8c8-minmax-fp32-wasmsdot-u2.c",
+    "src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-5x8c8-minmax-fp32-wasmsdot.c",
+    "src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-5x8c8-minmax-fp32-wasmusdot-u2.c",
+    "src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-5x8c8-minmax-fp32-wasmusdot.c",
     "src/qs8-vcvt/gen/qs8-vcvt-wasmrelaxedsimd-u8.c",
     "src/qs8-vcvt/gen/qs8-vcvt-wasmrelaxedsimd-u16.c",
     "src/qs8-vlrelu/gen/qs8-vlrelu-wasmrelaxedsimd-arm-u16.c",

diff --git a/scripts/generate-qs8-gemm.sh b/scripts/generate-qs8-gemm.sh
@@ -304,35 +304,40 @@ tools/xngen src/qs8-gemm/MRx4c16-wasmdot.c.in -D MR=2 NR=8 -D REQUANTIZATION=FP3
 tools/xngen src/qs8-gemm/MRx4c16-wasmdot.c.in -D MR=3 NR=8 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=0 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-3x8c16-minmax-fp32-wasmusdot.c &
 tools/xngen src/qs8-gemm/MRx4c16-wasmdot.c.in -D MR=4 NR=8 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=0 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-4x8c16-minmax-fp32-wasmusdot.c &
 
+tools/xngen src/qs8-gemm/MRx4c16-wasmdot.c.in -D MR=1 NR=4 -D REQUANTIZATION=     -D DATATYPE=QD8 -D SDOT=0 -o src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-1x4c16-minmax-wasmusdot.c &
+tools/xngen src/qs8-gemm/MRx4c16-wasmdot.c.in -D MR=2 NR=4 -D REQUANTIZATION=     -D DATATYPE=QD8 -D SDOT=0 -o src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-2x4c16-minmax-wasmusdot.c &
+tools/xngen src/qs8-gemm/MRx4c16-wasmdot.c.in -D MR=3 NR=4 -D REQUANTIZATION=     -D DATATYPE=QD8 -D SDOT=0 -o src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-3x4c16-minmax-wasmusdot.c &
+tools/xngen src/qs8-gemm/MRx4c16-wasmdot.c.in -D MR=4 NR=4 -D REQUANTIZATION=     -D DATATYPE=QD8 -D SDOT=0 -o src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-4x4c16-minmax-wasmusdot.c &
+
+tools/xngen src/qs8-gemm/MRx4c16-wasmdot.c.in -D MR=1 NR=8 -D REQUANTIZATION=     -D DATATYPE=QD8 -D SDOT=0 -o src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-1x8c16-minmax-wasmusdot.c &
+tools/xngen src/qs8-gemm/MRx4c16-wasmdot.c.in -D MR=2 NR=8 -D REQUANTIZATION=     -D DATATYPE=QD8 -D SDOT=0 -o src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-2x8c16-minmax-wasmusdot.c &
+tools/xngen src/qs8-gemm/MRx4c16-wasmdot.c.in -D MR=3 NR=8 -D REQUANTIZATION=     -D DATATYPE=QD8 -D SDOT=0 -o src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-3x8c16-minmax-wasmusdot.c &
+tools/xngen src/qs8-gemm/MRx4c16-wasmdot.c.in -D MR=4 NR=8 -D REQUANTIZATION=     -D DATATYPE=QD8 -D SDOT=0 -o src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-4x8c16-minmax-wasmusdot.c &
+
+### C8 micro-kernels
 tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=1 NR=8 -D UNROLL=0 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=0 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-1x8c8-minmax-fp32-wasmusdot.c &
 tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=2 NR=8 -D UNROLL=0 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=0 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-2x8c8-minmax-fp32-wasmusdot.c &
 tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=3 NR=8 -D UNROLL=0 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=0 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-3x8c8-minmax-fp32-wasmusdot.c &
 tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=4 NR=8 -D UNROLL=0 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=0 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-4x8c8-minmax-fp32-wasmusdot.c &
+tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=5 NR=8 -D UNROLL=0 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=0 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-5x8c8-minmax-fp32-wasmusdot.c &
 
 tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=1 NR=8 -D UNROLL=1 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=0 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-1x8c8-minmax-fp32-wasmusdot-u2.c &
 tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=2 NR=8 -D UNROLL=1 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=0 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-2x8c8-minmax-fp32-wasmusdot-u2.c &
 tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=3 NR=8 -D UNROLL=1 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=0 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-3x8c8-minmax-fp32-wasmusdot-u2.c &
 tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=4 NR=8 -D UNROLL=1 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=0 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-4x8c8-minmax-fp32-wasmusdot-u2.c &
+tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=5 NR=8 -D UNROLL=1 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=0 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-5x8c8-minmax-fp32-wasmusdot-u2.c &
 
 tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=1 NR=8 -D UNROLL=0 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=1 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-1x8c8-minmax-fp32-wasmsdot.c &
 tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=2 NR=8 -D UNROLL=0 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=1 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-2x8c8-minmax-fp32-wasmsdot.c &
 tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=3 NR=8 -D UNROLL=0 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=1 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-3x8c8-minmax-fp32-wasmsdot.c &
 tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=4 NR=8 -D UNROLL=0 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=1 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-4x8c8-minmax-fp32-wasmsdot.c &
+tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=5 NR=8 -D UNROLL=0 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=1 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-5x8c8-minmax-fp32-wasmsdot.c &
 
 tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=1 NR=8 -D UNROLL=1 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=1 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-1x8c8-minmax-fp32-wasmsdot-u2.c &
 tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=2 NR=8 -D UNROLL=1 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=1 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-2x8c8-minmax-fp32-wasmsdot-u2.c &
 tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=3 NR=8 -D UNROLL=1 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=1 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-3x8c8-minmax-fp32-wasmsdot-u2.c &
 tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=4 NR=8 -D UNROLL=1 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=1 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-4x8c8-minmax-fp32-wasmsdot-u2.c &
-
-tools/xngen src/qs8-gemm/MRx4c16-wasmdot.c.in -D MR=1 NR=4 -D REQUANTIZATION=     -D DATATYPE=QD8 -D SDOT=0 -o src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-1x4c16-minmax-wasmusdot.c &
-tools/xngen src/qs8-gemm/MRx4c16-wasmdot.c.in -D MR=2 NR=4 -D REQUANTIZATION=     -D DATATYPE=QD8 -D SDOT=0 -o src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-2x4c16-minmax-wasmusdot.c &
-tools/xngen src/qs8-gemm/MRx4c16-wasmdot.c.in -D MR=3 NR=4 -D REQUANTIZATION=     -D DATATYPE=QD8 -D SDOT=0 -o src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-3x4c16-minmax-wasmusdot.c &
-tools/xngen src/qs8-gemm/MRx4c16-wasmdot.c.in -D MR=4 NR=4 -D REQUANTIZATION=     -D DATATYPE=QD8 -D SDOT=0 -o src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-4x4c16-minmax-wasmusdot.c &
-
-tools/xngen src/qs8-gemm/MRx4c16-wasmdot.c.in -D MR=1 NR=8 -D REQUANTIZATION=     -D DATATYPE=QD8 -D SDOT=0 -o src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-1x8c16-minmax-wasmusdot.c &
-tools/xngen src/qs8-gemm/MRx4c16-wasmdot.c.in -D MR=2 NR=8 -D REQUANTIZATION=     -D DATATYPE=QD8 -D SDOT=0 -o src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-2x8c16-minmax-wasmusdot.c &
-tools/xngen src/qs8-gemm/MRx4c16-wasmdot.c.in -D MR=3 NR=8 -D REQUANTIZATION=     -D DATATYPE=QD8 -D SDOT=0 -o src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-3x8c16-minmax-wasmusdot.c &
-tools/xngen src/qs8-gemm/MRx4c16-wasmdot.c.in -D MR=4 NR=8 -D REQUANTIZATION=     -D DATATYPE=QD8 -D SDOT=0 -o src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-4x8c16-minmax-wasmusdot.c &
+tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=5 NR=8 -D UNROLL=1 -D REQUANTIZATION=FP32 -D DATATYPE=QC8 -D SDOT=1 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-5x8c8-minmax-fp32-wasmsdot-u2.c &
 
 tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=1 NR=8 -D UNROLL=0 -D REQUANTIZATION=     -D DATATYPE=QD8 -D SDOT=0 -o src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-1x8c8-minmax-wasmusdot.c &
 tools/xngen src/qs8-gemm/c8-wasmdot.c.in -D MR=2 NR=8 -D UNROLL=0 -D REQUANTIZATION=     -D DATATYPE=QD8 -D SDOT=0 -o src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-2x8c8-minmax-wasmusdot.c &