Fix #938: Call win32 APIs directly

[mdboom]

Description

Instead of using pywin32, just calls the win32 APIs directly using Cython extern.

closes#938

This has a measurable impact on import time of about 9% (import cuda.bindings.driver in a fresh interpreter), mainly by not spending time importing win32api:

• Before: 92.1 ms
• After: 84.6 ms

It also improves "time to first call" by about 10% (since the first call resolves all of the dynamic function pointers and makes many win32 API calls):

• Before: 103.2 ms
• After: 93.4 ms

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[mdboom]

/ok to test

[copy-pr-bot]

/ok to test

@mdboom, there was an error processing your request: E1

See the following link for more information: https://docs.gha-runners.nvidia.com/cpr/e/1/

[mdboom]

/ok to test

[copy-pr-bot]

/ok to test

@mdboom, there was an error processing your request: E1

See the following link for more information: https://docs.gha-runners.nvidia.com/cpr/e/1/

[mdboom]

/ok to test 7828876

[rwgk]

Wonderful!

It'd be ideal to also reduce the code duplication (get_cuda_version(), cdef extern from "windows.h":) in a set of follow-on PRs. I believe it'll be pretty straightforward after this PR and the associated codegen PRs are merged.

[mdboom]

It'd be ideal to also reduce the code duplication (get_cuda_version(), cdef extern from "windows.h":) in a set of follow-on PRs. I believe it'll be pretty straightforward after this PR and the associated codegen PRs are merged.

The files that are generated by cybind are including the externs in a template, so they at least aren't copy-pasted. That is, all of the templates for each library have a ${snippet_windows_externs} that inserts all of this boilerplate (including the get_cuda_version function).

It might have been nicer to cimportcuda.bindings._lib.windll instead, but that would add a build-time dependency on cuda_bindings that not all of the cybind-generated packages have (though I might be misunderstanding that part). So I thought it better to make them "standalone" by having cybind put this snippet in all of them. See the cybind PR I also filed for more details.

[mdboom]

/ok to test 8c7ea2e

[rwgk]

Wow

CI / Test linux-64 / py3.10, 13.0.0, wheels, GPU l4 (push) Failing after 4m

> assert delay_ms - generous_tolerance <= elapsed_time_ms < delay_ms + generous_tolerance
E assert 520.4592895507812 < (500.0 + 20)

I haven't seen that flake for a while.

Just rerun and ignore. The timing being off by a small margin certainly isn't due to a problem in cuda-bindings.

[rwgk]

Tests passed, except for that one flake. I think a rerun of that test will resolve the flake.

[leofang]

Let's mergenot merge until the internal MR is approved, since it involves other teams as well.

[leofang]

FYI, I noticed the cuFile module is not changed?

[mdboom]

FYI, I noticed the cuFile module is not changed?

Ah, I think that's just an oversight on my part. I will regenerate that as well.

[mdboom]

/ok to test e0e868a

[mdboom]

/ok to test 7a46173

[mdboom]

/ok to test 488bdc2

[mdboom]

/ok to test 673974c

[copy-pr-bot]

/ok to test 673974c

@mdboom, there was an error processing your request: E2

See the following link for more information: https://docs.gha-runners.nvidia.com/cpr/e/2/

[mdboom]

/ok to test dcce4f5

[mdboom]

@kkraus14: I merged main into here for one final test, but otherwise no changes since your last "approved" review.

[leofang]

Thanks a lot, @mdboom! Be sure to backport this PR.

[github-actions]

Doc Preview CI
Preview removed because the pull request was closed or merged.