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

[Joseph Huber via llvm-commits]

================
@@ -1027,6 +1027,22 @@ struct GenericDeviceTy : public DeviceAllocatorTy {
   virtual Error queryAsyncImpl(__tgt_async_info &AsyncInfo, bool ReleaseQueue,
                                bool *IsQueueWorkCompleted) = 0;
 
+  /// Indicate whether the plugin transfers data faster when the host side of
+  /// the transfer is pinned memory. If a plugin returns true, the kernel
+  /// launch environment is staged in a pinned host buffer before it is
+  /// submitted. Plugins may benefit for different reasons: some pick a cheaper
+  /// copy path for buffers they know are pinned, others rely on the driver
+  /// only issuing a true asynchronous transfer out of page-locked memory.
+  virtual bool hasFastTransferWithPinnedMemory() const { return false; }
+
+  /// Allocate a pinned host buffer to stage a kernel launch environment. The
+  /// caller owns it until it registers it with
+  /// AsyncInfoWrapperTy::freeAllocationAfterSynchronization, which releases it
+  /// once the transfer reading it has completed. Returns nullptr if staging is
+  /// unavailable, in which case the caller must submit the launch environment
+  /// from ordinary host memory.
+  KernelLaunchEnvironmentTy *getPinnedLaunchEnvBuffer();
----------------
jhuber6 wrote:

Isn't this basically reinventing implicit arguments? What's stopping us from just tacking this to the end of the implicit arguments, unless we need CUDA to handle this as well?

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