Boolean reduction performance improvements

[ndgrigorian]

This PR makes changes to boolean reductions which align with #1364

Namely, the traversal pattern of work groups in boolean reductions has been changed to be fastest over the iteration dimension, rather than the reduction dimension, and a specialized kernel for reductions over axis 0 in matrices has been added.

The original contiguous boolean reduction kernel has also been renamed to make the difference more apparent.

• Have you provided a meaningful PR description?
• Have you added a test, reproducer or referred to an issue with a reproducer?
• Have you tested your changes locally for CPU and GPU devices?
• Have you made sure that new changes do not introduce compiler warnings?
• Have you checked performance impact of proposed changes?
• If this PR is a work in progress, are you opening the PR as a draft?

[github-actions]

View rendered docs @ https://intelpython.github.io/dpctl/pulls/1401/index.html

[github-actions]

Array API standard conformance tests for dpctl=0.14.6dev5=py310ha25a700_4 ran successfully.
Passed: 916
Failed: 84
Skipped: 119

[ndgrigorian]

Using the same example as in #1364, the performance benefits are clear:

Before:

In [5]: x = dpt.reshape(dpt.asarray(1, dtype="f4")/dpt.square(\
   ...:                            dpt.arange(1, 1282200*128 + 1, dtype="f4")), (1282200, 128))

In [6]: %time y = dpt.all(x, axis=0)
CPU times: user 481 ms, sys: 349 ms, total: 830 ms
Wall time: 763 ms

In [7]: %time y = dpt.all(x, axis=0)
CPU times: user 232 ms, sys: 325 ms, total: 556 ms
Wall time: 601 ms

In [8]: %time y = dpt.all(x, axis=0)
CPU times: user 316 ms, sys: 235 ms, total: 551 ms
Wall time: 599 ms

In [9]: %time y = dpt.any(x, axis=0)
CPU times: user 454 ms, sys: 261 ms, total: 715 ms
Wall time: 774 ms

In [10]: %time y = dpt.any(x, axis=0)
CPU times: user 284 ms, sys: 308 ms, total: 592 ms
Wall time: 639 ms

In [11]: %time y = dpt.any(x, axis=0)
CPU times: user 280 ms, sys: 325 ms, total: 605 ms
Wall time: 654 ms

after:

In [3]: x = dpt.reshape(dpt.asarray(1, dtype="f4")/dpt.square(\
   ...:                            dpt.arange(1, 1282200*128 + 1, dtype="f4")), (1282200, 128))

In [4]: %time y = dpt.all(x, axis=0)
CPU times: user 210 ms, sys: 17.3 ms, total: 227 ms
Wall time: 198 ms

In [5]: %time y = dpt.all(x, axis=0)
CPU times: user 8.63 ms, sys: 36.6 ms, total: 45.3 ms
Wall time: 50.3 ms

In [6]: %time y = dpt.all(x, axis=0)
CPU times: user 15 ms, sys: 35.2 ms, total: 50.3 ms
Wall time: 51.1 ms

In [7]: %time y = dpt.any(x, axis=0)
CPU times: user 81 ms, sys: 19.2 ms, total: 100 ms
Wall time: 108 ms

In [8]: %time y = dpt.any(x, axis=0)
CPU times: user 18.6 ms, sys: 25.4 ms, total: 44 ms
Wall time: 46.6 ms

In [9]: %time y = dpt.any(x, axis=0)
CPU times: user 7.31 ms, sys: 35.7 ms, total: 43 ms
Wall time: 45.5 ms

[oleksandr-pavlyk]

Please add "numpy<1.26" restriction to pip install commands in "generate-coverage" and "sycl-nightly" workflows. It looks good to go in, but I'd prefer a green CI

[coveralls]

coverage: 85.774%. remained the same when pulling 351232a on boolean-reduction-performance into 83fff33 on master.

[ndgrigorian]

Please add "numpy<1.26" restriction to pip install commands in "generate-coverage" and "sycl-nightly" workflows. It looks good to go in, but I'd prefer a green CI

It's been added and fixes the CI. I'll look into properly solving the problem in a separate PR.

[github-actions]

Array API standard conformance tests for dpctl=0.14.6dev5=py310ha25a700_5 ran successfully.
Passed: 916
Failed: 84
Skipped: 119