[Openmp-commits] [llvm] [openmp] [OpenMP][OMPX] Add ballot_sync (PR #91297)

[Shilei Tian via Openmp-commits]

================
@@ -57,6 +59,12 @@ int32_t shuffleDown(uint64_t Mask, int32_t Var, uint32_t LaneDelta,
   return __builtin_amdgcn_ds_bpermute(Index << 2, Var);
 }
 
+uint32_t ballotSync(uint32_t Mask, int32_t Pred) {
----------------
shiltian wrote:

The wavefront size for AMDGPU can be 32 or 64, which means the mask here should be `uint32_t` or `uint64_t` correspondingly. Do we want to unify them with `uint64_t` and provide a full mark variable, or provide `_32` and `_64` variant? @jdoerfert 

https://github.com/llvm/llvm-project/pull/91297