-
-
Notifications
You must be signed in to change notification settings - Fork 39
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.
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
TEST: 1.5.x + blas variants #130
Conversation
Hi! This is the friendly automated conda-forge-linting service. I just wanted to let you know that I linted all conda-recipes in your PR ( |
The picture is not exactly rosy. The only silver lining is that the failures for linux+blis resp. osx+mkl are only one failed test each. I mentioned in conda/conda-build#3947 that IMO, foundational packages like numpy/scipy should be tested across all available blas variants, but due to the problem described in that very issue, I currently don't see an easy way to do it. Note: I checked most of these boxes for all python versions (cpython 3.6-3.8, pypy 3.6), and they always showed the same behaviour - hence the simplified matrix.
linux + x86 + blis: 1 failed test (TestEigBanded.test_dgbtrf)
linux + ppc64le + openblas: 100 failed tests
osx + blis: Import error (see conda-forge/blis-feedstock#15) osx + mkl: 1 failed test (TestProjections.test_iterative_refinements_dense)
|
Updated the matrix with the netlib builds, which are all green, fortunately. |
Try again for blis after conda-forge/blis-feedstock#16. |
@isuruf
0.7 builds for linux & osx have existed for a while already... |
You need to update the version in https://github.com/conda-forge/blas-feedstock/blob/master/recipe/meta.yaml#L50 |
So after bumping blis to 0.7.0 and openblas to 0.3.10 (see conda-forge/blas-feedstock#45), the picture changed a bit, in some ways unexpectedly. In short:
linux + x86 + blis + cpython 3.8: 6 failed tests
linux + x86 + openblas: 1 failed test (cluster/tests/test_vq.py::TestKMean::test_krandinit)
linux + ppc64le + cpython 3.6: 60 failed tests
linux + ppc64le + cpython 3.7: 50 failed tests
linux + ppc64le + cpython 3.8: 78 failed tests
linux + ppc64le + pypy: 164 failed tests
osx + mkl + cpython: 1 failed test (TestProjections.test_iterative_refinements_dense)
osx + mkl + pypy: 3 failed tests
|
@martin-frbg you might be interested in the openblas ppc64le errors here. Also, apparently openblas 0.3.10 gained a segfault relative to 0.3.9 on linux x86, at least as exercised by the scipy test suite. |
Thanks for the pointer. Can you give me a bit of info about build options, hardware and compiler used ? The x86 segfault is most surprising (few changes in 0.3.10 except AVX512), ppc64 feels like a minefield with each gcc version bringing its own optimizer effects and defects to bear on power8/9. |
@martin-frbg
This is really more the domain of @isuruf, but if you mean how openblas was built, then the log is here. There are a lot of compilers flying around there (
|
Nothing immediately suspicious with those build options and the BLAS/LAPACK tests I regularly use. Not sure if I can run the scipy build today, but most likely candidate for the segfault in SVD would appear to be OpenBLAS PR 2516 (new AVX2 STRSM kernel) IFF the scipy test runs in single precision at all. |
@martin-frbg, looks like falling back to nehalem kernel fixed it. |
Thanks, that would suggest that the new Haswell STRSM kernel gets some corner case wrong. (Reverting this should be straightforward if you want to patch your build). Stil trying to get a scipy build going on ppc64le at unicamp.br |
Yep, will do that.
Let us know if you need help. |
For those that might have subscribed to this thread, I'm continuing this sanity check in #153. PS. I also fixed some incorrect totals for the number of CI runs. |
Opened scipy/scipy#13325 for |
Hi! This is the friendly automated conda-forge-linting service. I was trying to look for recipes to lint for you, but it appears we have a merge conflict. Please ping the 'conda-forge/core' team (using the @ notation in a comment) if you believe this is a bug. |
31de4bb
to
f9724af
Compare
Hi! This is the friendly automated conda-forge-linting service. I just wanted to let you know that I linted all conda-recipes in your PR ( |
f9724af
to
b385d8d
Compare
Updating to blis 0.8.0From 13 failures & 6 timeouts out of 80 combinations, there are now 13 (different) failures & 2 timeouts out of 80 combinations. The good news:
The bad news:
Details
osx + mkl + cpython: 1 failed test
osx + mkl + pypy: 3 failed tests
win + blis: 424-429 failed tests
win + openblas + cpython: 1 failed test
|
…da-forge-pinning 2021.01.26.08.41.30
b385d8d
to
170382d
Compare
No more 1.5 patch releases expected, time to close this PR. Work continues in #153 |
Checklist
0
(if the version changed)conda-smithy
(Use the phrase@conda-forge-admin, please rerender
in a comment in this PR for automated rerendering)Based on #129, but trying to see what happens with various different blas variants (especially for ppc64le).