diff options
| author | Gengxin Xie <gengxin.xie@intel.com> | 2020-06-16 10:47:54 +0800 |
|---|---|---|
| committer | Gengxin Xie <gengxin.xie@intel.com> | 2020-06-16 10:47:54 +0800 |
| commit | be73f966cef19cddb8877a5d3d993d13acb44985 (patch) | |
| tree | bccf537d9d94fee10e2a271e879a3773a55711d6 | |
| parent | 334b59dfcc1e5b9ba6d28647b8002d5c7cfa2765 (diff) | |
| download | numpy-be73f966cef19cddb8877a5d3d993d13acb44985.tar.gz | |
MAINT: Remove AVX512DQ intrinsics
| -rw-r--r-- | numpy/core/src/umath/simd.inc.src | 15 |
1 files changed, 8 insertions, 7 deletions
diff --git a/numpy/core/src/umath/simd.inc.src b/numpy/core/src/umath/simd.inc.src index 207c2a1f9..7c0212984 100644 --- a/numpy/core/src/umath/simd.inc.src +++ b/numpy/core/src/umath/simd.inc.src @@ -3272,7 +3272,7 @@ AVX512F_log_DOUBLE(npy_double * op, /* call glibc when x on [1.0 - 0x1p-4, 1.0 + 0x1.09p-4] */ __mmask8 m1 = _mm512_cmp_pd_mask(x, mTo_glibc_max, _CMP_LT_OQ); __mmask8 m2 = _mm512_cmp_pd_mask(x, mTo_glibc_min, _CMP_GT_OQ); - glibc_mask = _kand_mask8(m1,m2); + glibc_mask = m1 & m2; if (glibc_mask != 0xFF) { zero_mask = _mm512_cmp_pd_mask(x, zeros_d, _CMP_EQ_OQ); @@ -3280,9 +3280,8 @@ AVX512F_log_DOUBLE(npy_double * op, negx_mask = _mm512_cmp_pd_mask(x, zeros_d, _CMP_LT_OQ); nan_mask = _mm512_cmp_pd_mask(x, x, _CMP_NEQ_UQ); - divide_by_zero_mask = _kor_mask8(divide_by_zero_mask, - _kand_mask8(zero_mask, load_mask)); - invalid_mask = _kor_mask8(invalid_mask, negx_mask); + divide_by_zero_mask = divide_by_zero_mask | (zero_mask & load_mask); + invalid_mask = invalid_mask | negx_mask; x = avx512_set_masked_lanes_pd(x, zeros_d, negx_mask); @@ -3294,7 +3293,7 @@ AVX512F_log_DOUBLE(npy_double * op, _mm512_set1_epi64(0xfff0000000000000)); denormal_mask = _mm512_cmp_epi64_mask(top12, _mm512_set1_epi64(0), _CMP_EQ_OQ); - denormal_mask = _kandn_mask8(zero_mask, denormal_mask); + denormal_mask = (~zero_mask) & denormal_mask; ix = (__m512i)_mm512_mask_mul_pd(x, denormal_mask, x, _mm512_set1_pd(0x1p52)); ix = _mm512_mask_sub_epi64(ix, denormal_mask, @@ -3311,7 +3310,8 @@ AVX512F_log_DOUBLE(npy_double * op, __m512d z = (__m512d)(_mm512_sub_epi64(ix, _mm512_and_epi64(tmp, _mm512_set1_epi64(0xfff0000000000000)))); /* c = i/64 + 1 */ - __m512d c = _mm512_fmadd_pd(_mm512_cvtepi64_pd(i), mInv64, ones_d); + __m256i i_32 = _mm512_cvtepi64_epi32(i); + __m512d c = _mm512_fmadd_pd(_mm512_cvtepi32_pd(i_32), mInv64, ones_d); /* u = 2 * (z - c) / (z + c) */ __m512d u = _mm512_div_pd(_mm512_sub_pd(z, c), _mm512_add_pd(z,c)); @@ -3340,7 +3340,8 @@ AVX512F_log_DOUBLE(npy_double * op, * k * ln2_lo + c_lo + * log(z/c) */ - __m512d k = _mm512_cvtepi64_pd(ik); + __m256i ik_32 = _mm512_cvtepi64_epi32(ik); + __m512d k = _mm512_cvtepi32_pd(ik_32); __m512d tt = _mm512_fmadd_pd(k, mLN2HI, c_hi); __m512d tt2 = _mm512_fmadd_pd(k, mLN2LO, c_lo); tt = _mm512_add_pd(tt, tt2); |