Check cpuinfo only when needed.

[QuLogic]

The cpuinfo module does not support s390x and fails to import, but we don't really need it since we've set CFLAGS. Move the import later so it's only used when actually checking CPU instruction support.

TODO:

• [N/A] Unit tests and/or doctests in docstrings
• tox -e py37 passes locally
• tox -e py27 passes locally
• [N/A] Docstrings and API docs for any new/modified user-facing classes and functions
• Changes documented in docs/release.rst
• tox -e docs passes locally
• AppVeyor and Travis CI passes
• Test coverage to 100% (Coveralls passes)

[alimanfoo]

Thanks a lot for this. It looks good. Only thought is that if user has set CFLAGS then sources for SSE2 and AVX2 modules from c-blosc will never be included. So, I.e., user might give CFLAGS="-mavx2" but would not get AVX2 modules compiled because sources would be excluded. Hope that makes sense.

Not sure what is best solution, maybe would have been better rather than have variables like DISABLE_NUMCODECS_AVX2 which only allows to turn something off, we should have ENABLE_NUMCODECS_AVX2 which could be set to 1 to force on or set to 0 to force off. Thoughts welcome.

[QuLogic]

Since those files require setting -DSHUFFLE_SSE2_ENABLED and -DSHUFFLE_AVX2_ENABLED, maybe check if the user's set that in CFLAGS?

It's impossible to just look for a compiler option (e.g., -msse2), because it could be implied by any number of -march flags.

[jakirkham]

Would it be possible to raise this with the developers of cpuinfo?

ref: https://github.com/workhorsy/py-cpuinfo

[QuLogic]

It's known upstream: workhorsy/py-cpuinfo#69

[jakirkham]

It seems like the developer has done some work on it (has a branch) and probably needs more people from the community with access to that architecture to engage with him and try it out. Would you mind engaging in that issue directly and sharing a bit more about what in that branch doesn’t work?

[jakirkham]

It looks like upstream addressed this issue. So I've submitted PR ( #202 ), which merely updates cpuinfo.py to a commit with the needed feature. Can you please check if that is acceptable @QuLogic? If so, I'll go ahead and merge that PR.