Copybara import of the project:

--
17f6dab by Dillon <[email protected]>:

Add .vs to .gitignore

--
951240c by Dillon <[email protected]>:

Fix warnings on MSVC

--
585faa1 by Dillon <[email protected]>:

Disable warning on MSVC

--
92acc17 by Dillon <[email protected]>:

Disable AVX256 when /vlen not supported by MSVC

FUTURE_COPYBARA_INTEGRATE_REVIEW=#7840 from google:ds/windows 92acc17
PiperOrigin-RevId: 728495899

.gitignore

@@ -34,3 +34,4 @@ ehthumbs.db
 Thumbs.db
 *.swp
 .vscode
+.vs

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)
@@ -341,6 +347,8 @@ IF(CMAKE_C_COMPILER_ID STREQUAL "MSVC")
   # Test files have many sections, increase the limit. See
   # https://learn.microsoft.com/en-us/cpp/build/reference/bigobj-increase-number-of-sections-in-dot-obj-file.
   ADD_COMPILE_OPTIONS("/bigobj")
+  # Our float16 datatypes have constructors in C++ and not in C, which prompts a warning in MSVC
+  ADD_COMPILE_OPTIONS("/wd4190")
 ENDIF()
 
 IF(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")

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;

src/operator-run.c

@@ -2237,14 +2237,14 @@ void xnn_compute_f16_qd8_convert(
     const struct f16_qd8_convert_context context[restrict XNN_MIN_ELEMENTS(1)],
     size_t batch_index)
 {
-  return xnn_compute_f16_qx8_convert(context, xnn_f16_qd8_asymmetric_quantization_params, batch_index);
+  xnn_compute_f16_qx8_convert(context, xnn_f16_qd8_asymmetric_quantization_params, batch_index);
 }
 
 void xnn_compute_f16_qdu8_convert(
     const struct f16_qd8_convert_context context[restrict XNN_MIN_ELEMENTS(1)],
     size_t batch_index)
 {
-  return xnn_compute_f16_qx8_convert(context, xnn_f16_qdu8_asymmetric_quantization_params, batch_index);
+  xnn_compute_f16_qx8_convert(context, xnn_f16_qdu8_asymmetric_quantization_params, batch_index);
 }
 
 void xnn_compute_f32_qx8_convert(
@@ -2273,14 +2273,14 @@ void xnn_compute_f32_qd8_convert(
     const struct f32_qd8_convert_context context[restrict XNN_MIN_ELEMENTS(1)],
     size_t batch_index)
 {
-  return xnn_compute_f32_qx8_convert(context, xnn_f32_qd8_asymmetric_quantization_params, batch_index);
+  xnn_compute_f32_qx8_convert(context, xnn_f32_qd8_asymmetric_quantization_params, batch_index);
 }
 
 void xnn_compute_f32_qdu8_convert(
     const struct f32_qd8_convert_context context[restrict XNN_MIN_ELEMENTS(1)],
     size_t batch_index)
 {
-  return xnn_compute_f32_qx8_convert(context, xnn_f32_qdu8_asymmetric_quantization_params, batch_index);
+  xnn_compute_f32_qx8_convert(context, xnn_f32_qdu8_asymmetric_quantization_params, batch_index);
 }
 
 void xnn_compute_pack_lh(