[llvm] [offload] Use pinned memory for KLE (PR #213767)

[Joseph Huber via llvm-commits]

================
@@ -2840,6 +2840,10 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
     return true;
   }
 
+  /// Transfers out of registered pinned memory take the one-step path in
+  /// dataSubmitImpl instead of staging through an intermediate buffer.
+  bool hasFastTransferWithPinnedMemory() const override { return true; }
----------------
jhuber6 wrote:

What exactly controls this? I'd assume in pretty much any case a direct copying would be faster.

https://github.com/llvm/llvm-project/pull/213767