diff options
| author | Matti Picus <matti.picus@gmail.com> | 2021-11-19 10:26:43 +0200 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2021-11-19 10:26:43 +0200 |
| commit | 0bb936ccf0ffc4798010a0e532a722255671f738 (patch) | |
| tree | 843af3b63501020e0c4c0bb3c451335ad8ee80b5 | |
| parent | 056abda14dab7fa8daf7a1ab44144aeb2250c216 (diff) | |
| parent | 9959d3de0e90f1bc140f4df7c5a9af024c7622db (diff) | |
| download | numpy-0bb936ccf0ffc4798010a0e532a722255671f738.tar.gz | |
Merge pull request #20405 from seiko2plus/issue_20356
BUG, SIMD: Fix `exp` FP stack overflow when `AVX512_SKX` is enabled
| -rw-r--r-- | numpy/core/src/umath/loops_exponent_log.dispatch.c.src | 18 |
1 files changed, 13 insertions, 5 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 95cce553a..2dd43fb85 100644 --- a/numpy/core/src/umath/loops_exponent_log.dispatch.c.src +++ b/numpy/core/src/umath/loops_exponent_log.dispatch.c.src @@ -386,7 +386,7 @@ avx512_permute_x8var_pd(__m512d t0, __m512d t1, __m512d t2, __m512d t3, * #and_masks =_mm256_and_ps, _mm512_kand# * #xor_masks =_mm256_xor_ps, _mm512_kxor# * #fmadd = _mm256_fmadd_ps, _mm512_fmadd_ps# - * #mask_to_int = _mm256_movemask_ps, # + * #mask_to_int = _mm256_movemask_ps, npyv_tobits_b32# * #full_mask= 0xFF, 0xFFFF# * #masked_store = _mm256_maskstore_ps, _mm512_mask_storeu_ps# * #cvtps_epi32 = _mm256_cvtps_epi32, # @@ -833,11 +833,19 @@ AVX512F_exp_DOUBLE(npy_double * op, op += num_lanes; num_remaining_elements -= num_lanes; } - if (overflow_mask) { + /* + * Don't count on the compiler for cast between mask and int registers. + * On gcc7 with flags -march>=nocona -O3 can cause FP stack overflow + * which may lead to putting NaN into certain HW/FP calculations. + * + * For more details, please check the comments in: + * - https://github.com/numpy/numpy/issues/20356 + */ + if (npyv_tobits_b64(overflow_mask)) { npy_set_floatstatus_overflow(); } - if (underflow_mask) { + if (npyv_tobits_b64(underflow_mask)) { npy_set_floatstatus_underflow(); } } @@ -1062,10 +1070,10 @@ AVX512F_log_DOUBLE(npy_double * op, num_remaining_elements -= num_lanes; } - if (invalid_mask) { + if (npyv_tobits_b64(invalid_mask)) { npy_set_floatstatus_invalid(); } - if (divide_by_zero_mask) { + if (npyv_tobits_b64(divide_by_zero_mask)) { npy_set_floatstatus_divbyzero(); } } |