[Openmp-commits] [openmp] [amdgpu] Implement D2D memcpy as HSA call (PR #69955)

[Jon Chesterfield via Openmp-commits]

================
@@ -2250,14 +2250,37 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
                                           PinnedMemoryManager);
   }
 
-  /// Exchange data between two devices within the plugin. This function is not
-  /// supported in this plugin.
+  /// Exchange data between two devices within the plugin.
   Error dataExchangeImpl(const void *SrcPtr, GenericDeviceTy &DstGenericDevice,
                          void *DstPtr, int64_t Size,
                          AsyncInfoWrapperTy &AsyncInfoWrapper) override {
----------------
JonChesterfield wrote:

The memory copies are handwritten in each direction so it's not a matter of pulling a signal from a stream and calling an existing function. It's implementable but not trivial.

I'd like to land this blocking version first in order to establish that the underling HSA calls behave as they're supposed to via the internal testing infrastructure, before adding the additional complication to wire it into the signals.

Especially so as I am far from convinced that signals work robustly between GPUs, given they're on different HSA queues.

https://github.com/llvm/llvm-project/pull/69955