[Openmp-commits] [clang] [llvm] [mlir] [openmp] [OpenMP][offload] Cross-team reductions with variable number of teams (PR #195102)

[Johannes Doerfert via Openmp-commits]

================
@@ -61,6 +65,78 @@ static uint32_t gpu_irregular_simd_reduce(void *reduce_data,
   return (logical_lane_id == 0);
 }
 
+// Reduction within a block on the GPU.
+//
+// Template parameters:
+// - checkLiveness: Whether to check the liveness of the lanes. This is only
+//                  useful if gpu_block_reduce is called in a context where
+//                  L2 parallel regions are possible.
----------------
jdoerfert wrote:

L2 parallel regions are sequentialized, no? That should be the trivial case of num_threads == 1 handled in nvptx_parallel_reduce_nowait. Am I missing something?

https://github.com/llvm/llvm-project/pull/195102