[PVC] resmpl_12_224 got >15% perf regression with agama 1146.48

[chengjunlu]

This issue is reported to IGC originally. The root cause is related to the Triton auto-grf retrying threshold.
Copied the issue to Triton XPU repo to track.

Comments from [Kwasniewski, Patryk]:

The exact flow is this:

1. Old driver/RemoveUnusedIdImplicitArguments=0 - the kernel is compiled with 1472 spills, which triggers a recompilation on application level (a separate IGC call) with -cl-intel-256-GRF-per-thread. This compiles without spills and this version is used for execution.
2. New driver/RemoveUnusedIdImplicitArguments=1 - the kernel is compiled with 896 spills, the recompilation is not triggered, and this version is used.

if anything, this is an improvement: it shows that change works well, reducing spills from 1472 to 896. The issue is that now "better" kernel does not triggers recompilation with large GRF mode.

[anmyachev]

@chengjunlu @whitneywhtsang do you remember why we chose the value 1000 to restart the compilation with a large GRF mode? We could reduce it to 800 for example.

[whitneywhtsang]

@chengjunlu @whitneywhtsang do you remember why we chose the value 1000 to restart the compilation with a large GRF mode? We could reduce it to 800 for example.

We probably could, I don't recall we did a lot of experiment to pick 1000.

[chengjunlu]

@anmyachev
I don't remember why the 1000 was chosen neigher.

Do you think it make sense to align the threshold with the common Triton Config maxnreg ?

1. We always try to use larger number GRF per thread to eliminate the register spilling.
2. And the maximum number GRF per threads is controlled by maxnreg. The valid options are: 128, 256 and 512 for CRI.

It can give us a knob to config the recompiling threshold.

[anmyachev]

@anmyachev I don't remember why the 1000 was chosen neigher.

Do you think it make sense to align the threshold with the common Triton Config maxnreg ?

1. We always try to use larger number GRF per thread to eliminate the register spilling.
2. And the maximum number GRF per threads is controlled by maxnreg. The valid options are: 128, 256 and 512 for CRI.

It can give us a knob to config the recompiling threshold.

yes, it makes sense.