8000 AVX test failures for 32 bit manylinux1 wheels · Issue #17174 · numpy/numpy · GitHub
[go: up one dir, main page]
Skip to content

AVX test failures for 32 bit manylinux1 wheels #17174

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

Jump to bottom
Closed
charris opened this issue Aug 27, 2020 · 7 comments · Fixed by #16247
Closed

AVX test failures for 32 bit manylinux1 wheels #17174

charris opened this issue Aug 27, 2020 · 7 comments · Fixed by #16247
Assignees
@seiko2plus
Labels
00 - Bug component: numpy.ufunc component: SIMD Issues in SIMD (fast instruction sets) code or machinery

Comments

@charris
Copy link
Member
charris commented Aug 27, 2020

See test here. The problem is in TestAVXUfuncs.test_avx_based_ufunc. Note that the manylinux2010 wheels are fine. One possible solution not providing manylinux1 wheels for NumPy 1.20.

TestAVXUfuncs.test_avx_based_ufunc

Reproducing code example:

import numpy as np
<< your code here >>

Error message:

Numpy/Python version information:
@mattip
Copy link
Member
mattip commented Aug 27, 2020

this is the flaky 'reciprocal' ufunc on 32-bits we have seen in other places. like this issue about msse2 flags, this one that seems to be a problem with older gcc

@mattip
Copy link
Member
mattip commented Aug 27, 2020

I think dropping manylinux1 wheels for linux-32-bit is an option. Should we run it past the mailing list?

@charris
Copy link
Member Author
charris commented Aug 27, 2020

Should we run it past the mailing list?

Yes, it was one of the options I was considering for 1.20. The other was dropping Python 3.6, at least from the wheels. I think moving to newer compilers is something we need to do when we can.

@charris
Copy link
Member Author
charris commented Aug 27, 2020

I'll make a post.

@seiko2plus
Copy link
Member

@mattip,

I think dropping manylinux1 wheels for linux-32-bit is an option

No wait, please assign me to this issue, I will fix it.

@charris charris assigned charris and seiko2plus and unassigned charris Aug 28, 2020
@charris
Copy link
Member Author
charris commented Aug 28, 2020 •
edited
Loading

No wait, please assign me to this issue, I will fix it.

Done. Note that I (we?) were contemplating dropping manylinux1 even without the 32 bit problems.

@seiko2plus
Copy link
Member
seiko2plus commented Aug 31, 2020 •
edited
Loading

@charris, this issue will be resolved by #16247, I will reduce the amount of work so it can only cover (sqrt, absolute, square, reciprocal) to speed up the progress since NEON and SSE2 require a lot of emulations.

@seiko2plus seiko2plus mentioned this issue Sep 24, 2020
Merged
11 tasks
@mattip mattip added component: SIMD Issues in SIMD (fast instruction sets) code or machinery component: numpy.ufunc labels Sep 25, 2020
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees

@seiko2plus seiko2plus

Labels
00 - Bug component: numpy.ufunc component: SIMD Issues in SIMD (fast instruction sets) code or machinery
Projects
None yet
Milestone
No milestone
Development

Successfully merging a pull request may close this issue.

ENH:Umath Replace raw SIMD of unary float point(32-64) with NPYV - g0 seiko2plus/numpy
3 participants
@charris @mattip @seiko2plus
0