[SYCL] Improve sycl-post-link performance with -split=kernel

[sarnex]

Right now we are computing a new callgraph in every call to extractCallGraph. extractCallGraph is called every time we do a module split, so for -split=kernel, that would be once per kernel. For modules with many kernels, this can take a very long time. We only need to compute this once because the input IR doesn't seem to change between splits.

This improves performance of sycl-post-link from ~45min to ~7min for an example with 13k kernels

Signed-off-by: Sarnie, Nick [email protected]

[sarnex]

/summary:run

[asudarsa]

Hi @kbobrovs, can you please take a look? Thanks.

[asudarsa]

It will be great if we can add a test here. Not sure if we have any existing test that track 'compile time'.

[AlexeySachkov]

It will be great if we can add a test here. Not sure if we have any existing test that track 'compile time'.

I suppose that this is a non-functional change, which shouldn't require a test. I also doubt that we will be able to write proper 'compile time' test, or at least it should go into a separate suite of performance tests. The problem is that such tests can't be launched in a regular environment or they will be prone to fluctuations caused by many factors, like: having more than one test launched in parallel, having a misconfigured machine which uses tubro boost or operates on a non-fixed CPU frequency, etc.

[sarnex]

I agree with Alexey, it would be very difficult to write a unit test that isn't super flaky. I'll reach out @asudarsa offline to see if I can connect with some quality engineers to see if we have a testing environment for this

[kbobrovs]

Thanks, good catch!

[AlexeySachkov]

CUDA testing seems to fail due to infra issues, but since this patch is effectively a non-functional change, I will go ahead and merge it

[sarnex]

Thanks! @AlexeySachkov