From be6c42823c61dcc4606e55e4a2bd8849df1a5755 Mon Sep 17 00:00:00 2001 From: nihui Date: Tue, 26 Sep 2023 08:59:21 +0000 Subject: [PATCH] apply code-format changes --- src/layer/x86/convolution_x86.cpp | 88 +++++++++++++++---------------- 1 file changed, 44 insertions(+), 44 deletions(-) diff --git a/src/layer/x86/convolution_x86.cpp b/src/layer/x86/convolution_x86.cpp index 5f2df2b82fb..9a06ebceeb8 100644 --- a/src/layer/x86/convolution_x86.cpp +++ b/src/layer/x86/convolution_x86.cpp @@ -1233,33 +1233,33 @@ int Convolution_x86::create_pipeline_int8_x86(const Option& opt) const int maxk = kernel_w * kernel_h; const int num_input = weight_data_size / maxk / num_output; -// int elempack = 1; -// int out_elempack_int32 = 1; -// #if __SSE2__ -// if (opt.use_packing_layout) -// { -// elempack = num_input % 8 == 0 ? 8 : 1; -// out_elempack_int32 = num_output % 4 == 0 ? 4 : 1; -// } -// #endif // __SSE2__ - -// if (elempack == 8 && out_elempack_int32 == 4 && opt.use_winograd_convolution && opt.use_winograd43_convolution && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1) -// { -// #if __SSE2__ -// conv3x3s1_winograd43_transform_kernel_pack8to4_int8_sse(weight_data, weight_winograd43_data, num_input, num_output, opt); -// #endif // __SSE2__ -// } -// else if (elempack == 8 && out_elempack_int32 == 1 && opt.use_winograd_convolution && opt.use_winograd43_convolution && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1) -// { -// #if __SSE2__ -// conv3x3s1_winograd43_transform_kernel_pack8to1_int8_sse(weight_data, weight_winograd43_data, num_input, num_output, opt); -// #endif // __SSE2__ -// } -// else if (elempack == 1 && out_elempack_int32 == 1 && opt.use_winograd_convolution && opt.use_winograd23_convolution && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1 && num_input >= 16 && num_output >= 16) -// { -// conv3x3s1_winograd23_transform_kernel_int8_sse(weight_data, weight_winograd23_data, num_input, num_output, opt); -// // conv3x3s1_winograd43_transform_kernel_int8_sse(weight_data, weight_winograd43_data, num_input, num_output, opt); -// } + // int elempack = 1; + // int out_elempack_int32 = 1; + // #if __SSE2__ + // if (opt.use_packing_layout) + // { + // elempack = num_input % 8 == 0 ? 8 : 1; + // out_elempack_int32 = num_output % 4 == 0 ? 4 : 1; + // } + // #endif // __SSE2__ + + // if (elempack == 8 && out_elempack_int32 == 4 && opt.use_winograd_convolution && opt.use_winograd43_convolution && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1) + // { + // #if __SSE2__ + // conv3x3s1_winograd43_transform_kernel_pack8to4_int8_sse(weight_data, weight_winograd43_data, num_input, num_output, opt); + // #endif // __SSE2__ + // } + // else if (elempack == 8 && out_elempack_int32 == 1 && opt.use_winograd_convolution && opt.use_winograd43_convolution && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1) + // { + // #if __SSE2__ + // conv3x3s1_winograd43_transform_kernel_pack8to1_int8_sse(weight_data, weight_winograd43_data, num_input, num_output, opt); + // #endif // __SSE2__ + // } + // else if (elempack == 1 && out_elempack_int32 == 1 && opt.use_winograd_convolution && opt.use_winograd23_convolution && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1 && num_input >= 16 && num_output >= 16) + // { + // conv3x3s1_winograd23_transform_kernel_int8_sse(weight_data, weight_winograd23_data, num_input, num_output, opt); + // // conv3x3s1_winograd43_transform_kernel_int8_sse(weight_data, weight_winograd43_data, num_input, num_output, opt); + // } if (opt.use_winograd_convolution && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1) { // conv3x3s1_winograd23_transform_kernel_int8(weight_data, weight_data_tm, num_input, num_output, opt); @@ -1367,23 +1367,23 @@ int Convolution_x86::forward_int8_x86(const Mat& bottom_blob, Mat& top_blob, con NCNN_LOGE("opt.num_threads %d changed, convolution gemm will use load-time value %d", opt.num_threads, nT); } -// if (elempack == 8 && out_elempack_int32 == 4 && opt.use_winograd_convolution && opt.use_winograd43_convolution && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1) -// { -// #if __SSE2__ -// conv3x3s1_winograd43_pack8to4_int8_sse(bottom_blob_bordered, top_blob_int32, weight_winograd43_data, opt); -// #endif // __SSE2__ -// } -// else if (elempack == 8 && out_elempack_int32 == 1 && opt.use_winograd_convolution && opt.use_winograd43_convolution && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1) -// { -// #if __SSE2__ -// conv3x3s1_winograd43_pack8to1_int8_sse(bottom_blob_bordered, top_blob_int32, weight_winograd43_data, opt); -// #endif // __SSE2__ -// } -// else if (elempack == 1 && out_elempack_int32 == 1 && opt.use_winograd_convolution && opt.use_winograd23_convolution && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1 && num_input >= 16 && num_output >= 16) -// { -// conv3x3s1_winograd23_int8_sse(bottom_blob_bordered, top_blob_int32, weight_winograd23_data, opt); -// // conv3x3s1_winograd43_int8_sse(bottom_blob_bordered, top_blob_int32, weight_winograd43_data, opt); -// } + // if (elempack == 8 && out_elempack_int32 == 4 && opt.use_winograd_convolution && opt.use_winograd43_convolution && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1) + // { + // #if __SSE2__ + // conv3x3s1_winograd43_pack8to4_int8_sse(bottom_blob_bordered, top_blob_int32, weight_winograd43_data, opt); + // #endif // __SSE2__ + // } + // else if (elempack == 8 && out_elempack_int32 == 1 && opt.use_winograd_convolution && opt.use_winograd43_convolution && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1) + // { + // #if __SSE2__ + // conv3x3s1_winograd43_pack8to1_int8_sse(bottom_blob_bordered, top_blob_int32, weight_winograd43_data, opt); + // #endif // __SSE2__ + // } + // else if (elempack == 1 && out_elempack_int32 == 1 && opt.use_winograd_convolution && opt.use_winograd23_convolution && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1 && num_input >= 16 && num_output >= 16) + // { + // conv3x3s1_winograd23_int8_sse(bottom_blob_bordered, top_blob_int32, weight_winograd23_data, opt); + // // conv3x3s1_winograd43_int8_sse(bottom_blob_bordered, top_blob_int32, weight_winograd43_data, opt); + // } if (opt.use_winograd_convolution && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1) { // conv3x3s1_winograd23_int8(bottom_blob_bordered, top_blob_int32, weight_data_tm, _nT, opt);