[WaveTranform] estimate SGPR pressure before wave-transform

[cmc-rep]

This is a draft for idea-sharing

[rocm-cciapp]

PSDB Build Link: http://mlse-bdc-20dd129:8065/#/builders/10/builds/766

[cmc-rep]

@cdevadas @lalaniket8 not really asking for formal code review. Just make this PR a place for discussion:

Right now, if I measure SGPR pressure after WaveTransform for RCCL case, it is 564. If I measure lanemask + SGPR pressure using this pass on per-lane CFG before RA, it is like 138 + 4. So it is way off. Likely it means this algorithm needs improvement, and WaveTransform also needs improvement.

Finally, feel free to take this piece for the WaveTransform project.

[cdevadas]

@cdevadas @lalaniket8 not really asking for formal code review. Just make this PR a place for discussion:

Right now, if I measure SGPR pressure after WaveTransform for RCCL case, it is 564. If I measure lanemask + SGPR pressure using this pass on per-lane CFG before RA, it is like 138 + 4. So it is way off. Likely it means this algorithm needs improvement, and WaveTransform also needs improvement.

Finally, feel free to take this piece for the WaveTransform project.

Sure. It is hard to judge the pressure beforehand. This disparity is huge. I observed almost the similar difference when I incorporated the existing register pressure tracker earlier. We need to tighten the WaveTransform implementation for more optimal ACC initialization to keep them closer to the blocks they are used. Sure, we can try to use this analysis and see if it can be improved.

[cmc-rep]

I counted 550+ sgpr initialization after WaveTransform in the entry block. So maybe my estimate is pretty close if we can fix the problem in WaveTranform.