fix instruction extension dispatch (#5427)

src/layer/arm/convolution_packed_int8.h

@@ -12,21 +12,18 @@
 // CONDITIONS OF ANY KIND, either express or implied. See the License for the
 // specific language governing permissions and limitations under the License.
 
-#if !(__ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD)
 #if NCNN_RUNTIME_CPU && NCNN_ARM84I8MM && __aarch64__ && !__ARM_FEATURE_MATMUL_INT8
 void convolution_transform_kernel_packed_int8_i8mm(const Mat& kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h);
 void convolution_packed_int8_i8mm(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt);
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_ARM82DOT && __aarch64__ && !__ARM_FEATURE_DOTPROD
+#if NCNN_RUNTIME_CPU && NCNN_ARM82DOT && __aarch64__ && !__ARM_FEATURE_DOTPROD && !__ARM_FEATURE_MATMUL_INT8
 void convolution_transform_kernel_packed_int8_asimddp(const Mat& kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h);
 void convolution_packed_int8_asimddp(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt);
 #endif
-#endif
 
 static void convolution_transform_kernel_packed_int8(const Mat& kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
 {
-#if !(__ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD)
 #if NCNN_RUNTIME_CPU && NCNN_ARM84I8MM && __aarch64__ && !__ARM_FEATURE_MATMUL_INT8
     if (ncnn::cpu_support_arm_i8mm())
     {
@@ -35,13 +32,12 @@ static void convolution_transform_kernel_packed_int8(const Mat& kernel, Mat& ker
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_ARM82DOT && __aarch64__ && !__ARM_FEATURE_DOTPROD
+#if NCNN_RUNTIME_CPU && NCNN_ARM82DOT && __aarch64__ && !__ARM_FEATURE_DOTPROD && !__ARM_FEATURE_MATMUL_INT8
     if (ncnn::cpu_support_arm_asimddp())
     {
         convolution_transform_kernel_packed_int8_asimddp(kernel, kernel_tm, inch, outch, kernel_w, kernel_h);
         return;
     }
-#endif
 #endif
 
     const int maxk = kernel_w * kernel_h;
@@ -531,7 +527,6 @@ static void convolution_transform_kernel_packed_int8(const Mat& kernel, Mat& ker
 
 static void convolution_packed_int8(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
 {
-#if !(__ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD)
 #if NCNN_RUNTIME_CPU && NCNN_ARM84I8MM && __aarch64__ && !__ARM_FEATURE_MATMUL_INT8
     if (ncnn::cpu_support_arm_i8mm())
     {
@@ -540,13 +535,12 @@ static void convolution_packed_int8(const Mat& bottom_blob, Mat& top_blob, const
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_ARM82DOT && __aarch64__ && !__ARM_FEATURE_DOTPROD
+#if NCNN_RUNTIME_CPU && NCNN_ARM82DOT && __aarch64__ && !__ARM_FEATURE_DOTPROD && !__ARM_FEATURE_MATMUL_INT8
     if (ncnn::cpu_support_arm_asimddp())
     {
         convolution_packed_int8_asimddp(bottom_blob, top_blob, weight_data_tm, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
         return;
     }
-#endif
 #endif
 
     const int w = bottom_blob.w;

src/layer/arm/innerproduct_fp16s.h

@@ -12,23 +12,20 @@
 // CONDITIONS OF ANY KIND, either express or implied. See the License for the
 // specific language governing permissions and limitations under the License.
 
-#if !(__ARM_FEATURE_FP16_FML || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
 #if NCNN_RUNTIME_CPU && NCNN_ARM82FP16FML && __aarch64__ && !__ARM_FEATURE_FP16_FML
 void innerproduct_pack4_fp16s_neon_asimdfhm(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt);
 void innerproduct_fp16s_neon_asimdfhm(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt);
 void innerproduct_transform_kernel_fp16s_neon_asimdfhm(const Mat& weight_data, Mat& weight_data_tm, int num_input, int num_output, const Option& opt);
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_ARM82 && __aarch64__ && !__ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+#if NCNN_RUNTIME_CPU && NCNN_ARM82 && __aarch64__ && !__ARM_FEATURE_FP16_VECTOR_ARITHMETIC && !__ARM_FEATURE_FP16_FML
 void innerproduct_pack4_fp16s_neon_asimdhp(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt);
 void innerproduct_fp16s_neon_asimdhp(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt);
 void innerproduct_transform_kernel_fp16s_neon_asimdhp(const Mat& weight_data, Mat& weight_data_tm, int num_input, int num_output, const Option& opt);
 #endif
-#endif
 
 static void innerproduct_pack4_fp16s_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt)
 {
-#if !(__ARM_FEATURE_FP16_FML || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
 #if NCNN_RUNTIME_CPU && NCNN_ARM82FP16FML && __aarch64__ && !__ARM_FEATURE_FP16_FML
     if (ncnn::cpu_support_arm_asimdfhm())
     {
@@ -37,13 +34,12 @@ static void innerproduct_pack4_fp16s_neon(const Mat& bottom_blob, Mat& top_blob,
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_ARM82 && __aarch64__ && !__ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+#if NCNN_RUNTIME_CPU && NCNN_ARM82 && __aarch64__ && !__ARM_FEATURE_FP16_VECTOR_ARITHMETIC && !__ARM_FEATURE_FP16_FML
     if (ncnn::cpu_support_arm_asimdhp())
     {
         innerproduct_pack4_fp16s_neon_asimdhp(bottom_blob, top_blob, weight_data_fp16, bias_data, activation_type, activation_params, opt);
         return;
     }
-#endif
 #endif
 
     const int num_input = bottom_blob.w * bottom_blob.elempack;
@@ -294,7 +290,6 @@ static void innerproduct_pack4_fp16s_neon(const Mat& bottom_blob, Mat& top_blob,
 
 static void innerproduct_fp16s_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt)
 {
-#if !(__ARM_FEATURE_FP16_FML || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
 #if NCNN_RUNTIME_CPU && NCNN_ARM82FP16FML && __aarch64__ && !__ARM_FEATURE_FP16_FML
     if (ncnn::cpu_support_arm_asimdfhm())
     {
@@ -303,13 +298,12 @@ static void innerproduct_fp16s_neon(const Mat& bottom_blob, Mat& top_blob, const
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_ARM82 && __aarch64__ && !__ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+#if NCNN_RUNTIME_CPU && NCNN_ARM82 && __aarch64__ && !__ARM_FEATURE_FP16_VECTOR_ARITHMETIC && !__ARM_FEATURE_FP16_FML
     if (ncnn::cpu_support_arm_asimdhp())
     {
         innerproduct_fp16s_neon_asimdhp(bottom_blob, top_blob, weight_data_fp16, bias_data, activation_type, activation_params, opt);
         return;
     }
-#endif
 #endif
 
     const int num_input = bottom_blob.w * bottom_blob.elempack;
@@ -516,7 +510,6 @@ static void innerproduct_fp16s_neon(const Mat& bottom_blob, Mat& top_blob, const
 
 static void innerproduct_transform_kernel_fp16s_neon(const Mat& weight_data, Mat& weight_data_tm, int num_input, int num_output, const Option& opt)
 {
-#if !(__ARM_FEATURE_FP16_FML || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
 #if NCNN_RUNTIME_CPU && NCNN_ARM82FP16FML && __aarch64__ && !__ARM_FEATURE_FP16_FML
     if (ncnn::cpu_support_arm_asimdfhm())
     {
@@ -525,13 +518,12 @@ static void innerproduct_transform_kernel_fp16s_neon(const Mat& weight_data, Mat
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_ARM82 && __aarch64__ && !__ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+#if NCNN_RUNTIME_CPU && NCNN_ARM82 && __aarch64__ && !__ARM_FEATURE_FP16_VECTOR_ARITHMETIC && !__ARM_FEATURE_FP16_FML
     if (ncnn::cpu_support_arm_asimdhp())
     {
         innerproduct_transform_kernel_fp16s_neon_asimdhp(weight_data, weight_data_tm, num_input, num_output, opt);
         return;
     }
-#endif
 #endif
 
     int out_elempack = 1;

src/layer/arm/innerproduct_gemm_fp16s.h

@@ -12,19 +12,16 @@
 // CONDITIONS OF ANY KIND, either express or implied. See the License for the
 // specific language governing permissions and limitations under the License.
 
-#if !(__ARM_FEATURE_FP16_FML || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
 #if NCNN_RUNTIME_CPU && NCNN_ARM82FP16FML && __aarch64__ && !__ARM_FEATURE_FP16_FML
 void innerproduct_gemm_fp16s_neon_asimdfhm(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt);
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_ARM82 && __aarch64__ && !__ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+#if NCNN_RUNTIME_CPU && NCNN_ARM82 && __aarch64__ && !__ARM_FEATURE_FP16_VECTOR_ARITHMETIC && !__ARM_FEATURE_FP16_FML
 void innerproduct_gemm_fp16s_neon_asimdhp(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt);
 #endif
-#endif
 
 static void innerproduct_gemm_fp16s_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt)
 {
-#if !(__ARM_FEATURE_FP16_FML || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
 #if NCNN_RUNTIME_CPU && NCNN_ARM82FP16FML && __aarch64__ && !__ARM_FEATURE_FP16_FML
     if (ncnn::cpu_support_arm_asimdfhm())
     {
@@ -33,13 +30,12 @@ static void innerproduct_gemm_fp16s_neon(const Mat& bottom_blob, Mat& top_blob,
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_ARM82 && __aarch64__ && !__ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+#if NCNN_RUNTIME_CPU && NCNN_ARM82 && __aarch64__ && !__ARM_FEATURE_FP16_VECTOR_ARITHMETIC && !__ARM_FEATURE_FP16_FML
     if (ncnn::cpu_support_arm_asimdhp())
     {
         innerproduct_gemm_fp16s_neon_asimdhp(bottom_blob, top_blob, weight_data_fp16, bias_data, activation_type, activation_params, opt);
         return;
     }
-#endif
 #endif
 
     const int num_input = bottom_blob.w;

src/layer/x86/cast_bf16.h

@@ -17,7 +17,7 @@ void cast_fp32_to_bf16_sse_avx512bf16(const Mat& bottom_blob, Mat& top_blob, con
 void cast_bf16_to_fp32_sse_avx512bf16(const Mat& bottom_blob, Mat& top_blob, const Option& opt);
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVX512BF16__
 void cast_fp32_to_bf16_sse_avx2(const Mat& bottom_blob, Mat& top_blob, const Option& opt);
 void cast_bf16_to_fp32_sse_avx2(const Mat& bottom_blob, Mat& top_blob, const Option& opt);
 #endif
@@ -32,7 +32,7 @@ static void cast_fp32_to_bf16_sse(const Mat& bottom_blob, Mat& top_blob, const O
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVX512BF16__
     if (ncnn::cpu_support_x86_avx2())
     {
         cast_fp32_to_bf16_sse_avx2(bottom_blob, top_blob, opt);
@@ -104,7 +104,7 @@ static void cast_bf16_to_fp32_sse(const Mat& bottom_blob, Mat& top_blob, const O
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVX512BF16__
     if (ncnn::cpu_support_x86_avx2())
     {
         cast_bf16_to_fp32_sse_avx2(bottom_blob, top_blob, opt);

src/layer/x86/convolution_3x3_winograd_int8.h

@@ -12,29 +12,27 @@
 // CONDITIONS OF ANY KIND, either express or implied. See the License for the
 // specific language governing permissions and limitations under the License.
 
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
 #if NCNN_RUNTIME_CPU && NCNN_AVX512VNNI && __AVX512F__ && !__AVX512VNNI__
 void conv3x3s1_winograd23_int8_avx512vnni(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt);
 void conv3x3s1_winograd43_int8_avx512vnni(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt);
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
 void conv3x3s1_winograd23_int8_avxvnni(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt);
 void conv3x3s1_winograd43_int8_avxvnni(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt);
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
 void conv3x3s1_winograd23_transform_kernel_int8_avx2(const Mat& kernel, Mat& AT, int inch, int outch, const Option& opt);
 void conv3x3s1_winograd23_int8_avx2(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt);
 void conv3x3s1_winograd43_transform_kernel_int8_avx2(const Mat& kernel, Mat& AT, int inch, int outch, const Option& opt);
 void conv3x3s1_winograd43_int8_avx2(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt);
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__
+#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
 void conv3x3s1_winograd23_int8_xop(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt);
 void conv3x3s1_winograd43_int8_xop(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt);
 #endif
-#endif
 
 static void pack_A_tile_int8(const Mat& A, Mat& AT, int batch, int max_ii, int max_kk)
 {
@@ -3430,14 +3428,12 @@ static inline void conv3x3s1_winograd23_transform_kernel_tile_int8(const Mat& ke
 
 static void conv3x3s1_winograd23_transform_kernel_int8(const Mat& kernel, Mat& AT, int inch, int outch, const Option& opt)
 {
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx2())
     {
         conv3x3s1_winograd23_transform_kernel_int8_avx2(kernel, AT, inch, outch, opt);
         return;
     }
-#endif
 #endif
 
     const int M = outch;
@@ -4430,7 +4426,6 @@ static inline void conv3x3s1_winograd23_transform_output_tile_int8(const Mat& to
 
 static void conv3x3s1_winograd23_int8(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt)
 {
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
 #if NCNN_RUNTIME_CPU && NCNN_AVX512VNNI && __AVX512F__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx512_vnni())
     {
@@ -4439,7 +4434,7 @@ static void conv3x3s1_winograd23_int8(const Mat& bottom_blob, Mat& top_blob, con
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx_vnni())
     {
         conv3x3s1_winograd23_int8_avxvnni(bottom_blob, top_blob, AT, nT, opt);
@@ -4447,7 +4442,7 @@ static void conv3x3s1_winograd23_int8(const Mat& bottom_blob, Mat& top_blob, con
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx2())
     {
         conv3x3s1_winograd23_int8_avx2(bottom_blob, top_blob, AT, nT, opt);
@@ -4455,13 +4450,12 @@ static void conv3x3s1_winograd23_int8(const Mat& bottom_blob, Mat& top_blob, con
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__
+#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_xop())
     {
         conv3x3s1_winograd23_int8_xop(bottom_blob, top_blob, AT, nT, opt);
         return;
     }
-#endif
 #endif
 
     int outw = top_blob.w;
@@ -4642,14 +4636,12 @@ static inline void conv3x3s1_winograd43_transform_kernel_tile_int8(const Mat& ke
 
 static void conv3x3s1_winograd43_transform_kernel_int8(const Mat& kernel, Mat& AT, int inch, int outch, const Option& opt)
 {
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx2())
     {
         conv3x3s1_winograd43_transform_kernel_int8_avx2(kernel, AT, inch, outch, opt);
         return;
     }
-#endif
 #endif
 
     const int M = outch;
@@ -6260,7 +6252,6 @@ static inline void conv3x3s1_winograd43_transform_output_tile_int8(const Mat& to
 
 static void conv3x3s1_winograd43_int8(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt)
 {
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
 #if NCNN_RUNTIME_CPU && NCNN_AVX512VNNI && __AVX512F__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx512_vnni())
     {
@@ -6269,7 +6260,7 @@ static void conv3x3s1_winograd43_int8(const Mat& bottom_blob, Mat& top_blob, con
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx_vnni())
     {
         conv3x3s1_winograd43_int8_avxvnni(bottom_blob, top_blob, AT, nT, opt);
@@ -6277,7 +6268,7 @@ static void conv3x3s1_winograd43_int8(const Mat& bottom_blob, Mat& top_blob, con
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx2())
     {
         conv3x3s1_winograd43_int8_avx2(bottom_blob, top_blob, AT, nT, opt);
@@ -6285,13 +6276,12 @@ static void conv3x3s1_winograd43_int8(const Mat& bottom_blob, Mat& top_blob, con
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__
+#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_xop())
     {
         conv3x3s1_winograd43_int8_xop(bottom_blob, top_blob, AT, nT, opt);
         return;
     }
-#endif
 #endif
 
     int outw = top_blob.w;

src/layer/x86/convolution_im2col_gemm_int8.h

@@ -12,24 +12,22 @@
 // CONDITIONS OF ANY KIND, either express or implied. See the License for the
 // specific language governing permissions and limitations under the License.
 
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
 #if NCNN_RUNTIME_CPU && NCNN_AVX512VNNI && __AVX512F__ && !__AVX512VNNI__
 void convolution_im2col_gemm_int8_avx512vnni(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int nT, const Option& opt);
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
 void convolution_im2col_gemm_int8_avxvnni(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int nT, const Option& opt);
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
 void convolution_im2col_gemm_transform_kernel_int8_avx2(const Mat& kernel, Mat& AT, int inch, int outch, int kernel_w, int kernel_h, const Option& opt);
 void convolution_im2col_gemm_int8_avx2(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int nT, const Option& opt);
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__
+#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
 void convolution_im2col_gemm_int8_xop(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int nT, const Option& opt);
 #endif
-#endif
 
 static void convolution_im2col_pack_A_tile_int8(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk)
 {
@@ -7476,14 +7474,12 @@ static void convolution_im2col_input_tile_int8(const Mat& bottom_blob, Mat& B, i
 
 static void convolution_im2col_gemm_transform_kernel_int8(const Mat& kernel, Mat& AT, int inch, int outch, int kernel_w, int kernel_h, const Option& opt)
 {
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx2())
     {
         convolution_im2col_gemm_transform_kernel_int8_avx2(kernel, AT, inch, outch, kernel_w, kernel_h, opt);
         return;
     }
-#endif
 #endif
 
     // NCNN_LOGE("convolution_im2col_gemm_transform_kernel");
@@ -7558,24 +7554,23 @@ static void convolution_im2col_gemm_transform_kernel_int8(const Mat& kernel, Mat
 
 static void convolution_im2col_gemm_int8(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int nT, const Option& opt)
 {
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
 #if NCNN_RUNTIME_CPU && NCNN_AVX512VNNI && __AVX512F__ && !__AVX512VNNI__
-    if (ncnn::cpu_support_x86_avx512vnni())
+    if (ncnn::cpu_support_x86_avx512_vnni())
     {
         convolution_im2col_gemm_int8_avx512vnni(bottom_blob, top_blob, AT, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, nT, opt);
         return;
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__
-    if (ncnn::cpu_support_x86_avxvnni())
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
+    if (ncnn::cpu_support_x86_avx_vnni())
     {
         convolution_im2col_gemm_int8_avxvnni(bottom_blob, top_blob, AT, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, nT, opt);
         return;
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx2())
     {
         convolution_im2col_gemm_int8_avx2(bottom_blob, top_blob, AT, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, nT, opt);
@@ -7583,13 +7578,12 @@ static void convolution_im2col_gemm_int8(const Mat& bottom_blob, Mat& top_blob,
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__
+#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_xop())
     {
         convolution_im2col_gemm_int8_xop(bottom_blob, top_blob, AT, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, nT, opt);
         return;
     }
-#endif
 #endif
 
     const int maxk = kernel_w * kernel_h;

src/layer/x86/convolution_packed_int8.h

@@ -12,31 +12,26 @@
 // CONDITIONS OF ANY KIND, either express or implied. See the License for the
 // specific language governing permissions and limitations under the License.
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
-void convolution_transform_kernel_packed_int8_avx2(const Mat& kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h);
-#endif
-
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
 #if NCNN_RUNTIME_CPU && NCNN_AVX512VNNI && __AVX512F__ && !__AVX512VNNI__
 void convolution_packed_int8_avx512vnni(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt);
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
 void convolution_packed_int8_avxvnni(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt);
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
+void convolution_transform_kernel_packed_int8_avx2(const Mat& kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h);
 void convolution_packed_int8_avx2(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt);
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__
+#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
 void convolution_packed_int8_xop(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt);
 #endif
-#endif
 
 static void convolution_transform_kernel_packed_int8(const Mat& kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
 {
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx2())
     {
         convolution_transform_kernel_packed_int8_avx2(kernel, kernel_tm, inch, outch, kernel_w, kernel_h);
@@ -880,7 +875,6 @@ static void convolution_transform_kernel_packed_int8(const Mat& kernel, Mat& ker
 
 static void convolution_packed_int8(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
 {
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
 #if NCNN_RUNTIME_CPU && NCNN_AVX512VNNI && __AVX512F__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx512_vnni())
     {
@@ -889,7 +883,7 @@ static void convolution_packed_int8(const Mat& bottom_blob, Mat& top_blob, const
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx_vnni())
     {
         convolution_packed_int8_avxvnni(bottom_blob, top_blob, weight_data_tm, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
@@ -897,7 +891,7 @@ static void convolution_packed_int8(const Mat& bottom_blob, Mat& top_blob, const
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
+#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx2())
     {
         convolution_packed_int8_avx2(bottom_blob, top_blob, weight_data_tm, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
@@ -905,13 +899,12 @@ static void convolution_packed_int8(const Mat& bottom_blob, Mat& top_blob, const
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__
+#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__ && !__AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_xop())
     {
         convolution_packed_int8_xop(bottom_blob, top_blob, weight_data_tm, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
         return;
     }
-#endif
 #endif
 
     const int w = bottom_blob.w;