Catching up on uniformity analysis

Code Generation AMDGPU
,
1

Hello,

I have been working on replacing the LegacyDivergenceAnalysis in AMDGPUUnifyDivergentExitNodes with the new uniformity analysis. Two of the tests reported a failure in SIAnnotateControlFlow because some basic blocks were reported as not uniformly reachable with the use of uniformity analysis. The tests are:

  • LLVM :: CodeGen/AMDGPU/multi-divergent-exit-region.ll
  • LLVM :: CodeGen/AMDGPU/si-unify-exit-return-unreachable.ll

Here’s the CFG of a failing kernel:

image
image1192×1064 155 KB

exit0 is reported as a not uniformly reachable basic block by LegacyDA whereas Uniformity Analysis reports it as uniformly reachable.

Thank you!

2

My initial guess would be that UniformityAnalysis doesn’t handle switch instructions

3

If UA finds uniformity where LegacyDA does not, that’s a good thing, right? The reasonable way forward will be dump both analyses and compare which values are marked divergent/uniform in each.

1 Like
4

I think you’re right. It appears to be a bug in SIAnnotateControlFlow. AMDGPUUnifyDivergentExitNodes transforms the above CFG into the following:

image
image503×576 30.3 KB

SIAnnotateControlFlow fails to annotate this CFG with exit1 and NodeBlock still living in the Stack.

5

Wild guess, but could it be that the test is simply too sensitive about the names of the blocks?

6

I doubt it, I think control flow is not correctly closed when visiting LeafBlock4 and NodeBlock6.