f

Author: nihui

src/layer/x86/gemm_bf16s.h

@@ -2515,7 +2515,7 @@ static void gemm_transB_packed_tile_bf16s(const Mat& AT_tile, const Mat& BT_tile
             for (; kk < max_kk; kk++)
             {
                 __m512 _pA0 = bfloat2float_avx512(_mm256_loadu_si256((const __m256i*)pA));
-                __m512 _pB0 = bfloat2float_avx512(_mm256_castps_si256(_mm256_set1_ps(((const float*)pB)[0])));
+                __m512 _pB0 = bfloat2float_avx512(_mm256_set1_epi32(((const int*)pB)[0]));
                 __m512 _pB1 = _mm512_permute_ps(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
 
                 _sum0 = _mm512_fmadd_ps(_pA0, _pB0, _sum0);
@@ -3228,7 +3228,7 @@ static void gemm_transB_packed_tile_bf16s(const Mat& AT_tile, const Mat& BT_tile
 #endif // __AVX512BF16__
             for (; kk < max_kk; kk++)
             {
-                __m512 _pA0 = bfloat2float_avx512(_mm256_set1_ps(((const float*)pA)[0]));
+                __m512 _pA0 = bfloat2float_avx512(_mm256_set1_epi32(((const int*)pA)[0]));
                 __m512 _pB0 = bfloat2float_avx512(_mm256_loadu_si256((const __m256i*)pB));
                 __m512 _pB1 = _mm512_permute_ps(_pB0, _MM_SHUFFLE(0, 3, 2, 1));