[PATCH] D138443: [AMDGPU] Fix GCNSubtarget::getMinNumVGPRs, add unit test to check consistency between GCNSubtarget's getMinNumVGPRs, getMaxNumVGPRs and getOccupancyWithNumVGPRs.

[Jay Foad via Phabricator via llvm-commits]

foad added a comment.

I have no objection to committing this.

The only thing I'd like to try cleaning up in future is that getMinNumVGPRs relies on calling getNumWavesPerEUWithNumVGPRs with a NumVGPRs value greater than getAddressableNumVGPRs and still getting sensible answers, which seems to me like a case that getNumWavesPerEUWithNumVGPRs should not have to handle. I.e. I think we should be able to do this:

  --- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
  +++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
  @@ -1020,6 +1020,7 @@ unsigned getAddressableNumVGPRs(const MCSubtargetInfo *STI) {
   
   unsigned getNumWavesWithNumVGPRs(const MCSubtargetInfo *STI,
                                    unsigned NumVGPRs) {
  +  assert(NumVGPRs <= getAddressableNumVGPRs(STI));
     NumVGPRs = std::min(std::max(NumVGPRs, 1u), getAddressableNumVGPRs(STI));
     return std::min(getTotalNumVGPRs(STI) /
                         unsigned(alignTo(NumVGPRs, getVGPRAllocGranule(STI))),

but currently we can't.



================
Comment at: llvm/unittests/Target/AMDGPU/AMDGPUUnitTests.cpp:100
+
+  std::map<std::string, SmallVector<std::string>> TablePerCPUs;
+  for (auto CPUName : CPUs) {
----------------
vpykhtin wrote:
> arsenm wrote:
> > vpykhtin wrote:
> > > arsenm wrote:
> > > > StringMap?
> > > I need std::string as a key, StringMap uses StringRef but it doesn't own string data. Its not a big deal here because its only used on test fail or when printing.
> > DenseMap
> DenseMap doesn't work with std::string, it's not specialized for it.
`StringMap<std:string>`?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D138443/new/

https://reviews.llvm.org/D138443