diff options
| author | Young Jin Kim <youki@microsoft.com> | 2019-09-26 13:30:38 +0300 |
|---|---|---|
| committer | Young Jin Kim <youki@microsoft.com> | 2019-09-26 13:30:38 +0300 |
| commit | 13347764fc33d3220856185c4758059a8e43cd14 (patch) | |
| tree | 018f96c6e4b58e1b50c6cdc2e5be44956f2dd046 | |
| parent | abf37c64da887d1222e5ccd9b990674e5047b133 (diff) | |
debugging linux unit tests
| -rw-r--r-- | src/PackDepthwiseConvMatrixAvx2.cc | 14 |
1 files changed, 9 insertions, 5 deletions
diff --git a/src/PackDepthwiseConvMatrixAvx2.cc b/src/PackDepthwiseConvMatrixAvx2.cc index 04c08f3..840a654 100644 --- a/src/PackDepthwiseConvMatrixAvx2.cc +++ b/src/PackDepthwiseConvMatrixAvx2.cc @@ -46,11 +46,15 @@ PackedDepthWiseConvMatrix::PackedDepthWiseConvMatrix( // Allocate packed arrays int kernel_prod_aligned = (kernel_prod + 1) / 2 * 2; - pmat_ = static_cast<int8_t *>(fbgemmAlignedAlloc(64, ((K + 31) / 32) * kernel_prod_aligned * 32 * sizeof(int8_t))); - //posix_memalign( - // (void**)&pmat_, - // 64, - // ((K + 31) / 32) * kernel_prod_aligned * 32 * sizeof(int8_t)); + //pmat_ = static_cast<int8_t *>(fbgemmAlignedAlloc(64, ((K + 31) / 32) * kernel_prod_aligned * 32 * sizeof(int8_t))); +#ifdef _MSC_VER + pmat_ = (int8_t*)_aligned_malloc(((K + 31) / 32) * kernel_prod_aligned * 32 * sizeof(int8_t), 64); +#else + posix_memalign( + (void**)&pmat_, + 64, + ((K + 31) / 32) * kernel_prod_aligned * 32 * sizeof(int8_t)); +#endif // Pack input matrix // The layout is optimized to use vpmaddubsw efficiently (see |