diff options
| -rw-r--r-- | numpy/core/src/umath/loops_exponent_log.dispatch.c.src | 7 |
1 files changed, 5 insertions, 2 deletions
diff --git a/numpy/core/src/umath/loops_exponent_log.dispatch.c.src b/numpy/core/src/umath/loops_exponent_log.dispatch.c.src index 81b050557..e0ee7f7eb 100644 --- a/numpy/core/src/umath/loops_exponent_log.dispatch.c.src +++ b/numpy/core/src/umath/loops_exponent_log.dispatch.c.src @@ -735,6 +735,7 @@ AVX512F_exp_DOUBLE(npy_double * op, __m512d mTH_max = _mm512_set1_pd(0x1.62e42fefa39efp+9); __m512d mTH_min = _mm512_set1_pd(-0x1.74910d52d3053p+9); __m512d mTH_inf = _mm512_set1_pd(NPY_INFINITY); + __m512d mTH_ninf = _mm512_set1_pd(-NPY_INFINITY); __m512d zeros_d = _mm512_set1_pd(0.0f); __m512d ones_d = _mm512_set1_pd(1.0f); __m256i vindex = _mm256_loadu_si256((__m256i*)&indexarr[0]); @@ -751,7 +752,7 @@ AVX512F_exp_DOUBLE(npy_double * op, __mmask8 overflow_mask = avx512_get_partial_load_mask_pd(0, num_lanes); __mmask8 underflow_mask = avx512_get_partial_load_mask_pd(0, num_lanes); __mmask8 load_mask = avx512_get_full_load_mask_pd(); - __mmask8 xmin_mask, xmax_mask, inf_mask, nan_mask, nearzero_mask; + __mmask8 xmin_mask, xmax_mask, inf_mask, ninf_mask, nan_mask, nearzero_mask; while (num_remaining_elements > 0) { if (num_remaining_elements < num_lanes) { @@ -772,6 +773,7 @@ AVX512F_exp_DOUBLE(npy_double * op, xmax_mask = _mm512_cmp_pd_mask(x, mTH_max, _CMP_GT_OQ); xmin_mask = _mm512_cmp_pd_mask(x, mTH_min, _CMP_LT_OQ); inf_mask = _mm512_cmp_pd_mask(x, mTH_inf, _CMP_EQ_OQ); + ninf_mask = _mm512_cmp_pd_mask(x, mTH_ninf, _CMP_EQ_OQ); __m512i x_abs = _mm512_and_epi64(_mm512_castpd_si512(x), _mm512_set1_epi64(0x7FFFFFFFFFFFFFFF)); nearzero_mask = _mm512_cmp_pd_mask(_mm512_castsi512_pd(x_abs), @@ -779,7 +781,8 @@ AVX512F_exp_DOUBLE(npy_double * op, nearzero_mask = _mm512_kxor(nearzero_mask, nan_mask); overflow_mask = _mm512_kor(overflow_mask, _mm512_kxor(xmax_mask, inf_mask)); - underflow_mask = _mm512_kor(underflow_mask, xmin_mask); + underflow_mask = _mm512_kor(underflow_mask, + _mm512_kxor(xmin_mask, ninf_mask)); x = avx512_set_masked_lanes_pd(x, zeros_d, _mm512_kor(_mm512_kor(nan_mask, xmin_mask), _mm512_kor(xmax_mask, nearzero_mask))); |