diff options
author | Erik Schnetter <schnetter@gmail.com> | 2015-10-29 15:41:32 -0400 |
---|---|---|
committer | Erik Schnetter <schnetter@gmail.com> | 2015-10-29 15:41:32 -0400 |
commit | 71359cfa476930dbd30af743135fe6a243a7b1a5 (patch) | |
tree | 68b67b60959f385fcf785b5c3fe6fda24c2f9a0e | |
parent | 2e41e71ffb6d12325f9f0bee5dc79539dfc02cb5 (diff) | |
download | vecmathlib-71359cfa476930dbd30af743135fe6a243a7b1a5.zip vecmathlib-71359cfa476930dbd30af743135fe6a243a7b1a5.tar.gz |
Correct type name __m256i
-rw-r--r-- | vec_avx_fp16_16.h | 28 |
1 files changed, 14 insertions, 14 deletions
diff --git a/vec_avx_fp16_16.h b/vec_avx_fp16_16.h index 6af27e5..ddc7f18 100644 --- a/vec_avx_fp16_16.h +++ b/vec_avx_fp16_16.h @@ -282,13 +282,13 @@ template <> struct intvec<fp16, 16> : floatprops<fp16> { intvec lsr(intvec n) const { #ifdef __AVX2__ // TODO: Use permute instead of shift/mask? - _mm256i mlo = _mm256_set1_epi32(U(0x0000ffff)); - _mm256i vlo = _mm256_and_si256(mlo, v); - _mm256i vhi = v; - _mm256i clo = _mm256_and_si256(mlo, n); - _mm256i chi = _mm256_and_si256(mlo, _mm256_srli_epi32(n, 16)); - _mm256i rlo = _mm256_srlv_epi32(vlo, clo); - _mm256i rhi = _mm256_andnot_si256(mlo, _mm256_srlv_epi32(vhi, chi)); + __m256i mlo = _mm256_set1_epi32(U(0x0000ffff)); + __m256i vlo = _mm256_and_si256(mlo, v); + __m256i vhi = v; + __m256i clo = _mm256_and_si256(mlo, n); + __m256i chi = _mm256_and_si256(mlo, _mm256_srli_epi32(n, 16)); + __m256i rlo = _mm256_srlv_epi32(vlo, clo); + __m256i rhi = _mm256_andnot_si256(mlo, _mm256_srlv_epi32(vhi, chi)); return _mm256_or_si256(rhi, rlo); #else intvec r; @@ -313,13 +313,13 @@ template <> struct intvec<fp16, 16> : floatprops<fp16> { intvec operator<<(intvec n) const { #ifdef __AVX2__ // TODO: Use permute instead of shift/mask? - _mm256i mlo = _mm256_set1_epi32(U(0x0000ffff)); - _mm256i vlo = v; - _mm256i vhi = _mm256_andnot_si256(mlo, v; - _mm256i clo = _mm256_and_si256(mlo, n); - _mm256i chi = _mm256_and_si256(mlo, _mm256_srli_epi32(n, 16)); - _mm256i rlo = _mm256_and_si256(mlo, _mm256_sllv_epi32(vlo, clo)); - _mm256i rhi = _mm256_sllv_epi32(vhi, chi); + __m256i mlo = _mm256_set1_epi32(U(0x0000ffff)); + __m256i vlo = v; + __m256i vhi = _mm256_andnot_si256(mlo, v; + __m256i clo = _mm256_and_si256(mlo, n); + __m256i chi = _mm256_and_si256(mlo, _mm256_srli_epi32(n, 16)); + __m256i rlo = _mm256_and_si256(mlo, _mm256_sllv_epi32(vlo, clo)); + __m256i rhi = _mm256_sllv_epi32(vhi, chi); return _mm256_or_si256(rhi, rlo); #else intvec r; |