WIP: add aarch64, ppc64le, x390x wheels; change build slightly #70
Conversation
.travis.yml
Outdated
| @@ -2,7 +2,7 @@ env: | |||
| global: | |||
| - REPO_DIR=numpy | |||
| # Also see DAILY_COMMIT below | |||
| - BUILD_COMMIT=ff3df08438d570b0ccdda3f8a008278d8a4ad394 | |||
| - BUILD_COMMIT=a16dfb5b397718c5caf32c63f024d1ce3518c5a1 | |||
There was a problem hiding this comment.
might as well update this to latest HEAD
| CFLAGS="-std=c99 -fno-strict-aliasing" | ||
| # For verbosity: report where each command came from | ||
| export PS4='+(${BASH_SOURCE}:${LINENO}): ${FUNCNAME[0]:+${FUNCNAME[0]}(): }' | ||
| set -x |
There was a problem hiding this comment.
This makes the log output quite verbose: it should show for every shell command the file/line/function name used
| - git checkout %BUILD_COMMIT% | ||
| # Create _distributor_init.py | ||
| - cd .. | ||
| - python -c "import openblas_support; openblas_support.make_init('numpy/numpy')" |
There was a problem hiding this comment.
ah, there is someone using openblas_support. This should be using the openblas_support from numpy/tools and not a local copy. I think I did it this way since we were building wheels before numpy/tools/openblas_support.py existed, but now we can fix this correctly
|
Not so good. The build on both commits fails for the new architectures:
|
mattip
force-pushed
the
other-arch
branch
2 times, most recently
from
e989e3d to
c42309e
Compare
|
As there seem to be different issues going on, would it perhaps be wiser to add the architectures one at a time? Or do you think it better to tackle this as one big chunk? |
|
Thanks for the ping. Updated to latest multibuild and disabled appveyor builds for now to save CI resources |
|
I've moved the usual linux builds to manylinux2010 and added hypothesis to the testing requirements. |
|
Is there anything I could do to contribute? I got some RPi 4's lying around to test stuff, but I'm not quite sure how to approach that. |
Replaces #69. In order to build wheels for other architectures, we need to update multibuild and gfortran-install submodules, and use
OPENBLAS_VERSION="v0.3.5-605-gc815b8fb"which is 0.3.8dev, not 0.3.5.I zeroed out .appveyor.yml while this is ongoing to reduce the CI build load.
I removed the unused
openblas_support.pywhich came fromnumpy/tools. At some point we can think about usingpython numpy/tools/openblas_support.pyinstead ofgfortran-installbut that will be a separate PR.