Disable AVX256 when /vlen not supported by MSVC

google · Feb 19, 2025 · 92acc17 · 92acc17
1 parent 585faa1
commit 92acc17
Show file tree

Hide file tree

Showing 2 changed files with 18 additions and 0 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -277,6 +277,12 @@ ELSEIF(CMAKE_C_COMPILER_ID STREQUAL "Clang")
   ENDIF()
 ELSEIF(CMAKE_C_COMPILER_ID STREQUAL "MSVC")
   SET(XNNPACK_ENABLE_AVX512BF16 OFF)
+  IF(MSVC_VERSION LESS_EQUAL 1941)
+    # /vlen option not supported
+    SET(XNNPACK_ENABLE_AVX256SKX OFF)
+    SET(XNNPACK_ENABLE_AVX256VNNI OFF)
+    SET(XNNPACK_ENABLE_AVX256VNNIGFNI OFF)
+  ENDIF()
 ENDIF()
 OPTION(XNNPACK_ENABLE_HVX "Build XNNPACK with Hexagon HVX micro-kernels" ON)
 OPTION(XNNPACK_ENABLE_KLEIDIAI "Use KleidiAI GEMM microkernels for Arm" ON)

diff --git a/src/configs/gemm-config.c b/src/configs/gemm-config.c
@@ -2799,7 +2799,11 @@ static void init_qd8_f16_qc8w_gemm_config(void) {
         qd8_f16_qc8w_gemm_config.pack_weights_and_biases = NULL;  // Override the default packing function.
         qd8_f16_qc8w_gemm_config.packed_stride_weights_and_biases = NULL;  // Override the default packing function.
         qd8_f16_qc8w_gemm_config.pack_gemm_gio = (xnn_packw_gemm_gio_ukernel_fn) xnn_pack_qs8_gemm_gio_w;
+        #if XNN_ENABLE_AVX256VNNI
         qd8_f16_qc8w_gemm_config.pack_gemm_goi = (xnn_packw_gemm_goi_ukernel_fn) xnn_qs8_packw_gemm_goi_ukernel_x64c4__avx256vnni_prfm;
+        #else
+        qd8_f16_qc8w_gemm_config.pack_gemm_goi = (xnn_packw_gemm_goi_ukernel_fn) xnn_pack_qs8_gemm_goi_w;
+        #endif
         qd8_f16_qc8w_gemm_config.mr = 16;
         qd8_f16_qc8w_gemm_config.nr = 64;
         qd8_f16_qc8w_gemm_config.log2_kr = 2;
@@ -3389,7 +3393,11 @@ static void init_qd8_f32_qc8w_gemm_config(void) {
         qd8_f32_qc8w_gemm_config.pack_weights_and_biases = NULL;  // Override the default packing function.
         qd8_f32_qc8w_gemm_config.packed_stride_weights_and_biases = NULL;  // Override the default packing function.
         qd8_f32_qc8w_gemm_config.pack_gemm_gio = (xnn_packw_gemm_gio_ukernel_fn) xnn_pack_qs8_gemm_gio_w;
+        #if XNN_ENABLE_AVX256VNNI
         qd8_f32_qc8w_gemm_config.pack_gemm_goi = (xnn_packw_gemm_goi_ukernel_fn) xnn_qs8_packw_gemm_goi_ukernel_x64c4__avx256vnni_prfm;
+        #else
+        qd8_f32_qc8w_gemm_config.pack_gemm_goi = (xnn_packw_gemm_goi_ukernel_fn) xnn_pack_qs8_gemm_goi_w;
+        #endif
         qd8_f32_qc8w_gemm_config.mr = 16;
         qd8_f32_qc8w_gemm_config.nr = 64;
         qd8_f32_qc8w_gemm_config.log2_kr = 2;
@@ -4089,7 +4097,11 @@ static void init_qs8_qc8w_gemm_config(void) {
         qs8_qc8w_gemm_config.pack_weights_and_biases = NULL;  // Override the default packing function.
         qs8_qc8w_gemm_config.packed_stride_weights_and_biases = NULL;  // Override the default packing function.
         qs8_qc8w_gemm_config.pack_gemm_gio = (xnn_packw_gemm_gio_ukernel_fn) xnn_pack_qs8_gemm_gio_w;
+        #if XNN_ENABLE_AVX256VNNI
         qs8_qc8w_gemm_config.pack_gemm_goi = (xnn_packw_gemm_goi_ukernel_fn) xnn_qs8_packw_gemm_goi_ukernel_x64c4__avx256vnni_prfm;
+        #else
+        qs8_qc8w_gemm_config.pack_gemm_goi = (xnn_packw_gemm_goi_ukernel_fn) xnn_pack_qs8_gemm_goi_w;
+        #endif
         qs8_qc8w_gemm_config.pack_igemm_goki = (xnn_pack_conv_goki_w_fn) xnn_pack_qs8_conv_goki_w;
         qs8_qc8w_gemm_config.pack_igemm_kgo = (xnn_pack_conv_kgo_w_fn) xnn_pack_qs8_conv_kgo_w;
         qs8_qc8w_gemm_config.pack_deconv_goki = (xnn_pack_deconv_goki_w_fn) xnn_pack_qs8_deconv_goki_w;