fix dispatch

Tencent · Dec 16, 2024 · cdba870 · cdba870
1 parent 0d0abf2
commit cdba870
Show file tree

Hide file tree

Showing 2 changed files with 46 additions and 25 deletions.
diff --git a/.github/workflows/test-coverage.yml b/.github/workflows/test-coverage.yml
@@ -62,7 +62,16 @@ jobs:
         files: build/lcov.info
 
   linux-gcc-x64-sde:
+    name: ${{ matrix.cpu }}
     runs-on: ubuntu-24.04
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - { cpu: hsw, AVX2: ON, AVXVNNI: OFF, AVXVNNIINT8: OFF, AVXNECONVERT: OFF, AVX512: OFF, AVX512VNNI: OFF, AVX512BF16: OFF, AVX512FP16: OFF }
+          - { cpu: arl, AVX2: ON, AVXVNNI: ON,  AVXVNNIINT8: OFF, AVXNECONVERT: OFF, AVX512: OFF, AVX512VNNI: OFF, AVX512BF16: OFF, AVX512FP16: OFF }
+          - { cpu: arl, AVX2: ON, AVXVNNI: ON,  AVXVNNIINT8: ON,  AVXNECONVERT: ON,  AVX512: OFF, AVX512VNNI: OFF, AVX512BF16: OFF, AVX512FP16: OFF }
+          - { cpu: spr, AVX2: ON, AVXVNNI: OFF, AVXVNNIINT8: OFF, AVXNECONVERT: OFF, AVX512: ON,  AVX512VNNI: ON,  AVX512BF16: ON,  AVX512FP16: ON  }
     steps:
     - uses: actions/checkout@v4
     - name: update
@@ -79,12 +88,24 @@ jobs:
         CXX: g++-14
       run: |
         mkdir build && cd build
-        cmake -DCMAKE_BUILD_TYPE=debug -DNCNN_COVERAGE=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_AVX2=ON -DNCNN_AVX512=ON -DNCNN_AVX512VNNI=ON -DNCNN_AVX512BF16=ON -DNCNN_AVX512FP16=ON -DNCNN_XOP=OFF -DNCNN_AVXVNNI=OFF -DNCNN_AVXNECONVERT=OFF -DNCNN_OPENMP=OFF -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..
+        cmake -DCMAKE_BUILD_TYPE=debug -DNCNN_COVERAGE=ON -DNCNN_RUNTIME_CPU=OFF \
+            -DNCNN_AVX=ON \
+            -DNCNN_F16C=ON \
+            -DNCNN_XOP=OFF \
+            -DNCNN_AVX2=${{ matrix.AVX2 }} \
+            -DNCNN_AVXVNNI=${{ matrix.AVXVNNI }} \
+            -DNCNN_AVXVNNIINT8=${{ matrix.AVXVNNIINT8 }} \
+            -DNCNN_AVXNECONVERT=${{ matrix.AVXNECONVERT }} \
+            -DNCNN_AVX512=${{ matrix.AVX2 }} \
+            -DNCNN_AVX512VNNI=${{ matrix.AVX512VNNI }} \
+            -DNCNN_AVX512BF16=${{ matrix.AVX512BF16 }} \
+            -DNCNN_AVX512FP16=${{ matrix.AVX512FP16 }} \
+            -DNCNN_OPENMP=OFF -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..
         cmake --build . -j $(nproc)
     - name: test
       run: |
         cd build
-        TESTS_EXECUTABLE_LOADER=$SDE_PATH/sde64 TESTS_EXECUTABLE_LOADER_ARGUMENTS="-spr;--" ctest --output-on-failure -j 2
+        TESTS_EXECUTABLE_LOADER=$SDE_PATH/sde64 TESTS_EXECUTABLE_LOADER_ARGUMENTS="-${{ matrix.cpu }};--" ctest --output-on-failure -j $(nproc)
     - name: lcov-collect
       run: |
         cd build

diff --git a/src/layer/x86/gemm_int8.h b/src/layer/x86/gemm_int8.h
@@ -36,7 +36,7 @@ void transpose_pack_B_tile_fp32_to_int8_avxvnniint8(const Mat& B, Mat& BT, int j
 void gemm_transB_packed_tile_int8_avxvnniint8(const Mat& AT_tile, const Mat& BT_tile, Mat& topT_tile, int i, int max_ii, int j, int max_jj, int k, int max_kk);
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
 void pack_A_tile_int8_avxvnni(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk);
 void transpose_pack_A_tile_int8_avxvnni(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk);
 void pack_B_tile_int8_avxvnni(const Mat& B, Mat& BT, int j, int max_jj, int k, int max_kk);
@@ -48,7 +48,7 @@ void transpose_pack_B_tile_fp32_to_int8_avxvnni(const Mat& B, Mat& BT, int j, in
 void gemm_transB_packed_tile_int8_avxvnni(const Mat& AT_tile, const Mat& BT_tile, Mat& topT_tile, int i, int max_ii, int j, int max_jj, int k, int max_kk);
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
 void pack_A_tile_int8_avx2(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk);
 void transpose_pack_A_tile_int8_avx2(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk);
 void pack_B_tile_int8_avx2(const Mat& B, Mat& BT, int j, int max_jj, int k, int max_kk);
@@ -61,7 +61,7 @@ void unpack_output_tile_int32_to_fp32_avx2(const Mat& topT, const Mat& C, Mat& t
 void gemm_transB_packed_tile_int8_avx2(const Mat& AT_tile, const Mat& BT_tile, Mat& topT_tile, int i, int max_ii, int j, int max_jj, int k, int max_kk);
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__ && !__AVX2__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
 void gemm_transB_packed_tile_int8_xop(const Mat& AT_tile, const Mat& BT_tile, Mat& topT_tile, int i, int max_ii, int j, int max_jj, int k, int max_kk);
 #endif
 
@@ -109,15 +109,15 @@ static void pack_A_tile_int8(const Mat& A, Mat& AT, int i, int max_ii, int k, in
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx_vnni())
     {
         pack_A_tile_int8_avxvnni(A, AT, i, max_ii, k, max_kk);
         return;
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx2())
     {
         pack_A_tile_int8_avx2(A, AT, i, max_ii, k, max_kk);
@@ -457,15 +457,15 @@ static void transpose_pack_A_tile_int8(const Mat& A, Mat& AT, int i, int max_ii,
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx_vnni())
     {
         transpose_pack_A_tile_int8_avxvnni(A, AT, i, max_ii, k, max_kk);
         return;
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx2())
     {
         transpose_pack_A_tile_int8_avx2(A, AT, i, max_ii, k, max_kk);
@@ -797,15 +797,15 @@ static void pack_B_tile_int8(const Mat& B, Mat& BT, int j, int max_jj, int k, in
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx_vnni())
     {
         pack_B_tile_int8_avxvnni(B, BT, j, max_jj, k, max_kk);
         return;
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx2())
     {
         pack_B_tile_int8_avx2(B, BT, j, max_jj, k, max_kk);
@@ -1159,15 +1159,15 @@ static void transpose_pack_B_tile_int8(const Mat& B, Mat& BT, int j, int max_jj,
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx_vnni())
     {
         transpose_pack_B_tile_int8_avxvnni(B, BT, j, max_jj, k, max_kk);
         return;
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx2())
     {
         transpose_pack_B_tile_int8_avx2(B, BT, j, max_jj, k, max_kk);
@@ -1608,15 +1608,15 @@ static void pack_A_tile_fp32_to_int8(const Mat& A, Mat& AT, int i, int max_ii, i
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx_vnni())
     {
         pack_A_tile_fp32_to_int8_avxvnni(A, AT, i, max_ii, k, max_kk, scales);
         return;
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx2())
     {
         pack_A_tile_fp32_to_int8_avx2(A, AT, i, max_ii, k, max_kk, scales);
@@ -3373,15 +3373,15 @@ static void transpose_pack_A_tile_fp32_to_int8(const Mat& A, Mat& AT, int i, int
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx_vnni())
     {
         transpose_pack_A_tile_fp32_to_int8_avxvnni(A, AT, i, max_ii, k, max_kk, scales);
         return;
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx2())
     {
         transpose_pack_A_tile_fp32_to_int8_avx2(A, AT, i, max_ii, k, max_kk, scales);
@@ -4993,15 +4993,15 @@ static void pack_B_tile_fp32_to_int8(const Mat& B, Mat& BT, int j, int max_jj, i
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx_vnni())
     {
         pack_B_tile_fp32_to_int8_avxvnni(B, BT, j, max_jj, k, max_kk, scale);
         return;
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx2())
     {
         pack_B_tile_fp32_to_int8_avx2(B, BT, j, max_jj, k, max_kk, scale);
@@ -6055,15 +6055,15 @@ static void transpose_pack_B_tile_fp32_to_int8(const Mat& B, Mat& BT, int j, int
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx_vnni())
     {
         transpose_pack_B_tile_fp32_to_int8_avxvnni(B, BT, j, max_jj, k, max_kk, scale);
         return;
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx2())
     {
         transpose_pack_B_tile_fp32_to_int8_avx2(B, BT, j, max_jj, k, max_kk, scale);
@@ -7433,7 +7433,7 @@ static void transpose_pack_B_tile_fp32_to_int8(const Mat& B, Mat& BT, int j, int
 
 static void unpack_output_tile_int32_to_fp32(const Mat& topT, const Mat& C, Mat& top_blob, int broadcast_type_C, int i, int max_ii, int j, int max_jj, const Mat& descales, float alpha, float beta, int output_transpose)
 {
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx2())
     {
         unpack_output_tile_int32_to_fp32_avx2(topT, C, top_blob, broadcast_type_C, i, max_ii, j, max_jj, descales, alpha, beta, output_transpose);
@@ -12453,23 +12453,23 @@ static void gemm_transB_packed_tile_int8(const Mat& AT_tile, const Mat& BT_tile,
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx_vnni())
     {
         gemm_transB_packed_tile_int8_avxvnni(AT_tile, BT_tile, topT_tile, i, max_ii, j, max_jj, k, max_kk);
         return;
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx2())
     {
         gemm_transB_packed_tile_int8_avx2(AT_tile, BT_tile, topT_tile, i, max_ii, j, max_jj, k, max_kk);
         return;
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__ && !__AVX2__ && !__AVXVNNI__ && !__AVXVNNIINT8__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_xop())
     {
         gemm_transB_packed_tile_int8_xop(AT_tile, BT_tile, topT_tile, i, max_ii, j, max_jj, k, max_kk);