Tweak scheduling parameter of elementwise operations

oleksandr-pavlyk commented 2 months ago

For contiguous inputs, bump local-work-group size from 64 to 128 work-items.

This change is guided by performance study on Newton root finding example rich in elementwise operations.

With this change, unitrace states that 311 invocations of the kernel took 2805824666 ns, before that, with 64 workitems, the time was 3475091844 ns.

                                                      "dpctl::tensor::kernels::multiply::multiply_inplace_contig_kernel<std::complex<float>, std::complex<float>, 4u, 2u>[SIM
D16 {15625; 1; 1} {128; 1; 1}]",          311,           2805824666,    25.380501,              9021944,              8915416,              9143437

[x] 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?
[x] 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[bot] commented 2 months ago

Deleted rendered PR docs from intelpython.github.com/dpctl, latest should be updated shortly. :crossed_fingers:

coveralls commented 2 months ago

coverage: 88.21%. remained the same when pulling a9062030965dd881190a2f83a2307b911e870917 on increase-lws-for-elementwise-operations into 7757857466c2fcfb92e8f8e1ed38e90b35c42327 on master.

github-actions[bot] commented 2 months ago

Array API standard conformance tests for dpctl=0.17.0dev0=py310h15de555_302 ran successfully. Passed: 870 Failed: 8 Skipped: 92

IntelPython / dpctl

Tweak scheduling parameter of elementwise operations #1651