Skip to content

REV: Manual revert of float16 svml use #31178

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
charris merged 1 commit into from
Apr 7, 2026
+1 −45
Merged

REV: Manual revert of float16 svml use #31178

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 1 addition & 5 deletions numpy/_core/src/umath/loops_half.dispatch.c.src
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,7 @@
#define NPY__SVML_IS_ENABLED 0
#endif

#if NPY__SVML_IS_ENABLED && !defined(NPY_HAVE_AVX512_SPR)
#if NPY__SVML_IS_ENABLED

typedef __m256i npyvh_f16;
#define npyv_cvt_f16_f32 _mm512_cvtph_ps
Expand Down Expand Up @@ -80,11 +80,7 @@ NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(HALF_@func@)
if (!is_mem_overlap(src, steps[0], dst, steps[1], len) &&
(steps[0] == sizeof(npy_half)) &&
(steps[1] == sizeof(npy_half))) {
#ifdef NPY_HAVE_AVX512_SPR
__svml_@intrin@s32(src, dst, len);
#else
avx512_@intrin@_f16(src, dst, len);
#endif
return;
}
#endif // NPY__SVML_IS_ENABLED
Expand Down
40 changes: 0 additions & 40 deletions numpy/_core/tests/test_umath_accuracy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,8 +21,6 @@
IS_AVX = __cpu_features__.get('AVX512F', False) or \
(__cpu_features__.get('FMA3', False) and __cpu_features__.get('AVX2', False))

IS_AVX512FP16 = __cpu_features__.get('AVX512FP16', False)

# only run on linux with AVX, also avoid old glibc (numpy/numpy#20448).
runtest = (sys.platform.startswith('linux')
and IS_AVX and not _glibc_older_than("2.17"))
Expand Down Expand Up @@ -84,8 +82,6 @@ def test_validate_transcendentals(self):
maxulperr = data_subset['ulperr'].max()
assert_array_max_ulp(npfunc(inval), outval, maxulperr)

@pytest.mark.skipif(IS_AVX512FP16,
reason="SVML FP16 have slightly higher ULP errors")
@pytest.mark.parametrize("ufunc", UNARY_OBJECT_UFUNCS)
def test_validate_fp16_transcendentals(self, ufunc):
with np.errstate(all='ignore'):
Expand All @@ -94,39 +90,3 @@ def test_validate_fp16_transcendentals(self, ufunc):
datafp32 = datafp16.astype(np.float32)
assert_array_max_ulp(ufunc(datafp16), ufunc(datafp32),
maxulp=1, dtype=np.float16)

@pytest.mark.skipif(not IS_AVX512FP16,
reason="lower ULP only apply for SVML FP16")
def test_validate_svml_fp16(self):
max_ulp_err = {
"arccos": 2.54,
"arccosh": 2.09,
"arcsin": 3.06,
"arcsinh": 1.51,
"arctan": 2.61,
"arctanh": 1.88,
"cbrt": 1.57,
"cos": 1.43,
"cosh": 1.33,
"exp2": 1.33,
"exp": 1.27,
"expm1": 0.53,
"log": 1.80,
"log10": 1.27,
"log1p": 1.88,
"log2": 1.80,
"sin": 1.88,
"sinh": 2.05,
"tan": 2.26,
"tanh": 3.00,
}

with np.errstate(all='ignore'):
arr = np.arange(65536, dtype=np.int16)
datafp16 = np.frombuffer(arr.tobytes(), dtype=np.float16)
datafp32 = datafp16.astype(np.float32)
for func in max_ulp_err:
ufunc = getattr(np, func)
ulp = np.ceil(max_ulp_err[func])
assert_array_max_ulp(ufunc(datafp16), ufunc(datafp32),
maxulp=ulp, dtype=np.float16)
Loading