Linux memory fix

1 files changed, 7 insertions, 15 deletions

diff --git a/src/PackDepthwiseConvMatrixAvx2.cc b/src/PackDepthwiseConvMatrixAvx2.cc
index 840a654..a84c469 100644
--- a/src/PackDepthwiseConvMatrixAvx2.cc
+++ b/src/PackDepthwiseConvMatrixAvx2.cc
@@ -46,15 +46,7 @@ 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)));
-#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
+  pmat_ = static_cast<int8_t *>(fbgemmAlignedAlloc(64, ((K + 31) / 32) * kernel_prod_aligned * 32 * sizeof(int8_t)));
 
   // Pack input matrix
   // The layout is optimized to use vpmaddubsw efficiently (see
@@ -109,7 +101,7 @@ PackedDepthWiseConvMatrix::PackedDepthWiseConvMatrix(
   // (12, 8), (12, 9), (12, 10), zero, ..., (15, 8), (15, 9), (15, 10), zero
   // (28, 8), (28, 9), (28, 10), zero, ..., (31, 8), (31, 9), (31, 10), zero
   for (int k1 = 0; k1 < K; k1 += 32) {
-    __m256i* b_v = new __m256i[kernel_prod];
+    __m256i* b_v = static_cast<__m256i*>(ALIGNED_MALLOC(kernel_prod * sizeof(__m256i), 64));
     int remainder = K - k1;
     if (remainder < 32) {
       __m256i mask_v = _mm256_loadu_si256(
@@ -126,7 +118,7 @@ PackedDepthWiseConvMatrix::PackedDepthWiseConvMatrix(
     }
 
     // Interleave 2 SIMD registers
-    __m256i* b_interleaved_epi16 = new __m256i[kernel_prod_aligned];
+    __m256i* b_interleaved_epi16 = static_cast<__m256i*>(ALIGNED_MALLOC(kernel_prod_aligned * sizeof(__m256i), 64));
     __m256i zero_v = _mm256_setzero_si256();
     for (int i = 0; i < kernel_prod_aligned / 2; ++i) {
       if (2 * i + 1 >= kernel_prod) {
@@ -142,7 +134,7 @@ PackedDepthWiseConvMatrix::PackedDepthWiseConvMatrix(
     }
 
     // Interleave 4 SIMD registers
-    __m256i* b_interleaved_epi32 = new __m256i[kernel_prod_aligned];
+    __m256i* b_interleaved_epi32 = static_cast<__m256i*>(ALIGNED_MALLOC(kernel_prod_aligned * sizeof(__m256i), 64));
     for (int i = 0; i < kernel_prod_aligned / 4; ++i) {
       b_interleaved_epi32[4 * i] = _mm256_unpacklo_epi16(
           b_interleaved_epi16[4 * i], b_interleaved_epi16[4 * i + 2]);
@@ -164,9 +156,9 @@ PackedDepthWiseConvMatrix::PackedDepthWiseConvMatrix(
           b_interleaved_epi32[i]);
     }
 
-    delete[] b_v;
-    delete[] b_interleaved_epi16;
-    delete[] b_interleaved_epi32;
+    FREE(b_v);
+    FREE(b_interleaved_epi16);
+    FREE(b_interleaved_epi32);
   }
   FREE(smat_transposed);
 }