summaryrefslogtreecommitdiff
path: root/iqk_mul_mat.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'iqk_mul_mat.cpp')
-rw-r--r--iqk_mul_mat.cpp4
1 files changed, 1 insertions, 3 deletions
diff --git a/iqk_mul_mat.cpp b/iqk_mul_mat.cpp
index c204b22c..f38163d5 100644
--- a/iqk_mul_mat.cpp
+++ b/iqk_mul_mat.cpp
@@ -1383,9 +1383,7 @@ static void mul_mat_iq1bn_q8_K64(int n, const void * vx, size_t bx, const DataIn
#if defined __AVX512VNNI__ && defined __AVX512VL__
auto dot = _mm256_dpbusd_epi32(_mm256_dpbusd_epi32(_mm256_setzero_si256(), m1_8, dot1), m1_8, dot2);
#else
- dot1 = _mm256_madd_epi16(m1_16, _mm256_maddubs_epi16(m1_8, dot1));
- dot2 = _mm256_madd_epi16(m1_16, _mm256_maddubs_epi16(m1_8, dot2));
- auto dot = _mm256_add_epi32(_mm256_add_epi32(dot1, dot2));
+ auto dot = _mm256_madd_epi16(m1_16, _mm256_add_epi16(_mm256_maddubs_epi16(m1_8, dot1), _mm256_maddubs_epi16(m1_8, dot2)));
#endif
accd[iy] = _mm256_fmadd_ps(_mm256_set1_ps(q8.scale(iy, i)), _mm256_cvtepi32_ps(dot), accd[iy]);
}