[mlir][gpu] Handle async in gpu.launch_func lowering.
authorChristian Sigg <csigg@google.com>
Thu, 29 Oct 2020 16:58:48 +0000 (17:58 +0100)
committerChristian Sigg <csigg@google.com>
Thu, 29 Oct 2020 21:16:42 +0000 (22:16 +0100)
commitfce99e5f739eddce3e52a9c4967fc8a993c5772d
treeae27a4b415ddf51ebb3b022b33c5d92cf5060183
parent6c2ad4cf8758335e0896ba0540b21bbf6d239fbd
[mlir][gpu] Handle async in gpu.launch_func lowering.

For the synchronous case, destroy the stream after synchronization.

Sneak in a unrelated change to report why the gpu.wait conversion pattern didn't match.

Reviewed By: herhut

Differential Revision: https://reviews.llvm.org/D89933
mlir/lib/Conversion/GPUCommon/ConvertLaunchFuncToRuntimeCalls.cpp