Skip to content

Commit

Permalink
w
Browse files Browse the repository at this point in the history
  • Loading branch information
@nihui
nihui committed Jan 15, 2025
1 parent 1b80186 commit de71af2
Show file tree
Hide file tree
Showing 2 changed files with 34 additions and 34 deletions.
66 changes: 33 additions & 33 deletions src/layer/x86/convolution_im2col_gemm_int8.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -33,11 +33,11 @@ void unpack_output_tile_int32_avx2(const Mat& topT, Mat& top_blob, int i, int ma
#endif

// gemm_x86.h
#if __AVX512F__
#if NCNN_RUNTIME_CPU && __AVX512F__
namespace Gemm_x86_avx512_utility {
#elif __FMA__
#elif NCNN_RUNTIME_CPU && __FMA__
namespace Gemm_x86_fma_utility {
#elif __AVX__
#elif NCNN_RUNTIME_CPU && __AVX__
namespace Gemm_x86_avx_utility {
#else
namespace Gemm_x86_utility {
Expand All @@ -50,11 +50,11 @@ static void convolution_im2col_pack_A_tile_int8(const Mat& A, Mat& AT, int i, in
{
// A = (pa, maxk, inch/pa), outch

#if __AVX512F__
#if NCNN_RUNTIME_CPU && __AVX512F__
Gemm_x86_avx512_utility::pack_A_tile_int8(A, AT, i, max_ii, k, max_kk);
#elif __FMA__
#elif NCNN_RUNTIME_CPU && __FMA__
Gemm_x86_fma_utility::pack_A_tile_int8(A, AT, i, max_ii, k, max_kk);
#elif __AVX__
#elif NCNN_RUNTIME_CPU && __AVX__
Gemm_x86_avx_utility::pack_A_tile_int8(A, AT, i, max_ii, k, max_kk);
#else
Gemm_x86_utility::pack_A_tile_int8(A, AT, i, max_ii, k, max_kk);
Expand All @@ -65,11 +65,11 @@ static void convolution_gemm_transB_packed_tile_int8(const Mat& AT_tile, const M
{
// NCNN_LOGE("convolution_gemm_transB_packed_tile_int8 %d %d %d %d %d %d", i, max_ii, j, max_jj, k, max_kk);

#if __AVX512F__
#if NCNN_RUNTIME_CPU && __AVX512F__
Gemm_x86_avx512_utility::gemm_transB_packed_tile_int8(AT_tile, BT_tile, topT_tile, i, max_ii, j, max_jj, k, max_kk);
#elif __FMA__
#elif NCNN_RUNTIME_CPU && __FMA__
Gemm_x86_fma_utility::gemm_transB_packed_tile_int8(AT_tile, BT_tile, topT_tile, i, max_ii, j, max_jj, k, max_kk);
#elif __AVX__
#elif NCNN_RUNTIME_CPU && __AVX__
Gemm_x86_avx_utility::gemm_transB_packed_tile_int8(AT_tile, BT_tile, topT_tile, i, max_ii, j, max_jj, k, max_kk);
#else
Gemm_x86_utility::gemm_transB_packed_tile_int8(AT_tile, BT_tile, topT_tile, i, max_ii, j, max_jj, k, max_kk);
Expand Down Expand Up @@ -1604,22 +1604,22 @@ static void convolution_im2col_input_tile_int8_impl(const Mat& bottom_blob, Mat&
_mm_storeu_si128((__m128i*)offsets2, _vindex2);
_mm_storeu_si128((__m128i*)offsets3, _vindex3);

pp[0] = offsets0[0];
pp[1] = offsets2[0];
pp[2] = offsets0[1];
pp[3] = offsets2[1];
pp[4] = offsets0[2];
pp[5] = offsets2[2];
pp[6] = offsets0[3];
pp[7] = offsets2[3];
pp[8] = offsets1[0];
pp[9] = offsets3[0];
pp[10] = offsets1[1];
pp[11] = offsets3[1];
pp[12] = offsets1[2];
pp[13] = offsets3[2];
pp[14] = offsets1[3];
pp[15] = offsets3[3];
pp[0] = ((const signed char*)bottom_blob)[offsets0[0]];
pp[1] = ((const signed char*)bottom_blob)[offsets2[0]];
pp[2] = ((const signed char*)bottom_blob)[offsets0[1]];
pp[3] = ((const signed char*)bottom_blob)[offsets2[1]];
pp[4] = ((const signed char*)bottom_blob)[offsets0[2]];
pp[5] = ((const signed char*)bottom_blob)[offsets2[2]];
pp[6] = ((const signed char*)bottom_blob)[offsets0[3]];
pp[7] = ((const signed char*)bottom_blob)[offsets2[3]];
pp[8] = ((const signed char*)bottom_blob)[offsets1[0]];
pp[9] = ((const signed char*)bottom_blob)[offsets3[0]];
pp[10] = ((const signed char*)bottom_blob)[offsets1[1]];
pp[11] = ((const signed char*)bottom_blob)[offsets3[1]];
pp[12] = ((const signed char*)bottom_blob)[offsets1[2]];
pp[13] = ((const signed char*)bottom_blob)[offsets3[2]];
pp[14] = ((const signed char*)bottom_blob)[offsets1[3]];
pp[15] = ((const signed char*)bottom_blob)[offsets3[3]];

#endif // __AVX2__

Expand Down Expand Up @@ -1651,14 +1651,14 @@ static void convolution_im2col_input_tile_int8_impl(const Mat& bottom_blob, Mat&
_mm_storeu_si128((__m128i*)offsets0, _vindex0);
_mm_storeu_si128((__m128i*)offsets1, _vindex1);

pp[0] = offsets0[0];
pp[1] = offsets0[1];
pp[2] = offsets0[2];
pp[3] = offsets0[3];
pp[4] = offsets1[0];
pp[5] = offsets1[1];
pp[6] = offsets1[2];
pp[7] = offsets1[3];
pp[0] = ((const signed char*)bottom_blob)[offsets0[0]];
pp[1] = ((const signed char*)bottom_blob)[offsets0[1]];
pp[2] = ((const signed char*)bottom_blob)[offsets0[2]];
pp[3] = ((const signed char*)bottom_blob)[offsets0[3]];
pp[4] = ((const signed char*)bottom_blob)[offsets1[0]];
pp[5] = ((const signed char*)bottom_blob)[offsets1[1]];
pp[6] = ((const signed char*)bottom_blob)[offsets1[2]];
pp[7] = ((const signed char*)bottom_blob)[offsets1[3]];

#endif // __AVX2__

Expand Down
2 changes: 1 addition & 1 deletion src/layer/x86/gemm_x86.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -41,7 +41,7 @@ class Gemm_x86 : public Gemm
Mat CT_data;
};

// expose some gemm internal routines
// expose some gemm internal routines for convolution uses
namespace Gemm_x86_utility {
#if NCNN_INT8
void pack_A_tile_int8(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk);
Expand Down

0 comments on commit de71af2

Please sign in to comment.