Improve compatibility with 32bit architectures

[mrocklin]

This replaces np.foo64 with np.foo_ and vice versa for foo in {float, int}

See #20

• Tests added / passed
• Passes flake8 dask
• Fully documented, including docs/source/changelog.rst for all changes
  and one of the docs/source/*-api.rst files for new API

cc @QuLogic

[pitrou]

Two things:

• I wouldn't change np.float64 (it doesn't have anything with the architecture being 32- or 64-bit)
• I would use np.intp for integers

See https://docs.scipy.org/doc/numpy-1.13.0/user/basics.types.html#array-types-and-conversions-between-types for reference

[mrocklin]

Fixed

…

On Thu, Nov 30, 2017 at 11:49 AM, Antoine Pitrou ***@***.***> wrote: Two things: - I wouldn't change np.float64 (it doesn't have anything with the architecture between 32- or 64-bit) - I would use np.intp for integers — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#2937 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AASszAQtKaTfPQrKPwD1KhO23E-hDhoyks5s7tyigaJpZM4QwxgB> .

[pitrou]

LGTM. @jakirkham, any concerns?

[jakirkham]

Not really. Looks pretty clean to me. 👍 Happy to see this was relatively straightforward.

Should we be adding 32-bit to Travis or another CI to avoid regressions?

Side note: Guessing the xref in the OP was suppose to be issue ( #2507 ) (unless I'm missing something).

[mrocklin]

Side note: Guessing the xref in the OP was suppose to be issue ( #2507 ) (unless I'm missing something).

Yup. Thanks.

[mrocklin]

I noticed some Windows errors in Appveyor, so this might not be complete yet.

[pitrou]

I noticed some Windows errors in Appveyor, so this might not be complete yet.

Probably because np.int_ is a 32-bit int on 64-bit Windows (Windows 64 follows the LLP64 model).

[jakirkham]

Probably true. It would be nice if Windows was less weird with integers. 😞

[mrocklin]

Tests pass. Merging this soon if there are no further concerns.

[jakirkham]

Thanks @mrocklin.

[QuLogic]

OK, this fixes all the new failures from 0.16.0, but not the repartition ones originally from #2507.