Skip to content

test_reference_numerics_large fails on PPC #85978

@Flamefire

Description

@Flamefire

🐛 Describe the bug

test_reference_numerics_large_cos_cpu_bfloat16
test_reference_numerics_large_cos_cpu_float32
test_reference_numerics_large_sin_cpu_bfloat16
test_reference_numerics_large_sin_cpu_float32

All fail on POWER9 (at least) in this test introduced with #76626 into 1.12

Reproducer:

$ python -c 'import torch; print(torch.tensor([1.0000e+20,1.0000e+20,1.0000e+20,1.0000e+20,1.0000e+20,1.0000e+20,1.0000e+20,1.0000e+20], dtype=torch.float32).sin())'
tensor([-inf, -inf, -inf, -inf, -inf, -inf, -inf, -inf])

Not related to a similar earlier failure reported at #60259 as that is for complex values while this is for plain values.

Versions

The inf results are also seem on 1.10 & 1.11 but the test was introduced in 1.12

Metadata

Metadata

Assignees

No one assigned

    Labels

    module: POWERIssues specific to the POWER/ppc architecturetriagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate module

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions