Repository navigation
discrepancy of results for dpctl.tensor.sin and dpt.tensor.expm1 for f4 and f8 data types on gpu #1439
Description
For the following cases the results are different for float32 and float64 on gpu.
sin
>>> dpt.sin(dpt.asarray(-0.,dtype='f4',device='gpu'))
usm_ndarray(0., dtype=float32)
>>> dpt.sin(dpt.asarray(-0.,dtype='f8',device='gpu'))
usm_ndarray(-0.)
expm1
>>> dpt.expm1(dpt.asarray(-0.,dtype='f4',device='gpu'))
usm_ndarray(-0., dtype=float32)
>>> dpt.expm1(dpt.asarray(-0.,dtype='f8',device='gpu'))
usm_ndarray(0.)
Metadata
Metadata
Assignees
Labels
No labels
Activity
[-]Array API and `dpctl.tensor` discrepancy for special cases of a few functions[/-][+]discrepancy of results for `dpctl.tensor.sin` and `dpt.tensor.expm1` for `f4` and `f8` data types on `gpu`[/+]ndgrigorian commentedon Oct 15, 2023
The cause in both cases is
std::sin, which does not preserve the sign bit of-0.on GPU.Anywhere that
std::sinis used, logic will need to be added to account for signed 0.Resolves gh-1439
Resolves gh-1439
Merge pull request #1477 from IntelPython/resolve-gh-1439