fix avxvnni dispatch

Tencent · Dec 12, 2024 · 9d22cd6 · 9d22cd6
1 parent e52daad
commit 9d22cd6
Show file tree

Hide file tree

Showing 3 changed files with 9 additions and 9 deletions.
diff --git a/src/layer/x86/convolution_3x3_winograd_int8.h b/src/layer/x86/convolution_3x3_winograd_int8.h
@@ -17,7 +17,7 @@ int conv3x3s1_winograd23_int8_avx512vnni(const Mat& bottom_blob, Mat& top_blob,
 int 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__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVX512VNNI__
 int conv3x3s1_winograd23_int8_avxvnni(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt);
 int conv3x3s1_winograd43_int8_avxvnni(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt);
 #endif
@@ -4297,7 +4297,7 @@ static int conv3x3s1_winograd23_int8(const Mat& bottom_blob, Mat& top_blob, cons
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx_vnni())
     {
         return conv3x3s1_winograd23_int8_avxvnni(bottom_blob, top_blob, AT, nT, opt);
@@ -6129,7 +6129,7 @@ static int conv3x3s1_winograd43_int8(const Mat& bottom_blob, Mat& top_blob, cons
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx_vnni())
     {
         return conv3x3s1_winograd43_int8_avxvnni(bottom_blob, top_blob, AT, nT, opt);

diff --git a/src/layer/x86/convolution_im2col_gemm_int8.h b/src/layer/x86/convolution_im2col_gemm_int8.h
@@ -16,7 +16,7 @@
 int 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__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVX512VNNI__
 int 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
 
@@ -7390,7 +7390,7 @@ static int convolution_im2col_gemm_int8(const Mat& bottom_blob, Mat& top_blob, c
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx_vnni())
     {
         return convolution_im2col_gemm_int8_avxvnni(bottom_blob, top_blob, AT, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, nT, opt);

diff --git a/src/layer/x86/lstm_int8.h b/src/layer/x86/lstm_int8.h
@@ -18,7 +18,7 @@ void lstm_dynamic_quantize_scale2int8_avx512vnni(const float* ptr, int size, flo
 void lstm_int8_avx512vnni(const Mat& bottom_blob_int8, const Mat& bottom_blob_int8_descales, Mat& top_blob, int reverse, const Mat& weight_data_tm, const Mat& weight_data_tm_int8_descales, const Mat& bias_c, const Mat& weight_hr, Mat& hidden_state, Mat& cell_state, const Option& opt);
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVX512F__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVX512F__ && !__AVXVNNI__ && !__AVX512VNNI__
 void lstm_transform_weight_int8_avxvnni(const Mat& weight_xc, const Mat& weight_xc_int8_scales, const Mat& weight_hc, const Mat& weight_hc_int8_scales, const Mat& bias_c, Mat& weight_data_tm, Mat& weight_data_tm_int8_descales, Mat& bias_c_tm, int size, int num_output, int num_directions, int hidden_size, const Option& opt);
 void lstm_dynamic_quantize_scale2int8_avxvnni(const float* ptr, int size, float scale, signed char* outptr);
 void lstm_int8_avxvnni(const Mat& bottom_blob_int8, const Mat& bottom_blob_int8_descales, Mat& top_blob, int reverse, const Mat& weight_data_tm, const Mat& weight_data_tm_int8_descales, const Mat& bias_c, const Mat& weight_hr, Mat& hidden_state, Mat& cell_state, const Option& opt);
@@ -43,7 +43,7 @@ static void lstm_transform_weight_int8(const Mat& weight_xc, const Mat& weight_x
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVX512F__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVX512F__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx_vnni())
     {
         lstm_transform_weight_int8_avxvnni(weight_xc, weight_xc_int8_scales, weight_hc, weight_hc_int8_scales, bias_c, weight_data_tm, weight_data_tm_int8_descales, bias_c_tm, size, num_output, num_directions, hidden_size, opt);
@@ -1621,7 +1621,7 @@ static void lstm_dynamic_quantize_scale2int8(const float* ptr, int size, float s
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVX512F__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVX512F__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx_vnni())
     {
         lstm_dynamic_quantize_scale2int8_avxvnni(ptr, size, scale, outptr);
@@ -1705,7 +1705,7 @@ static void lstm_int8(const Mat& bottom_blob_int8, const Mat& bottom_blob_int8_d
     }
 #endif
 
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVX512F__ && !__AVXVNNI__ && !__AVX512VNNI__
+#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX__ && !__AVX512F__ && !__AVXVNNI__ && !__AVX512VNNI__
     if (ncnn::cpu_support_x86_avx_vnni())
     {
         lstm_int8_avxvnni(bottom_blob_int8, bottom_blob_int8_descales, top_blob, reverse, weight_data_tm, weight_data_tm_int8_descales, bias_c, weight_hr, hidden_state, cell_state, opt);