[llvm] [Offload] Introduce ATTACH map-type support for pointer attachment. (PR #149036)

Abhinav Gaba via llvm-commits llvm-commits at lists.llvm.org
Wed Aug 13 13:25:31 PDT 2025


================
@@ -2538,6 +2538,10 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
                                           getAgent(), (uint64_t)Size);
   }
 
+  Error dataFence(__tgt_async_info *Async) override {
----------------
abhinavgaba wrote:

Sounds good! We can pull the plugin changes out of this PR.

But could you clarify what you mean by "implement them first"? 

>From what I gather, amd/cuda/x86_64 devices don't do out-of-order tranfers within a single queue. So the no-op/empty implementation is the final implementation for them.

Intel uses out-of-order queues in the downstream level_zero plugin, but that's not upstreamed yet. I know there's some effort to add out-of-order queue support as part of the libOffload changes being done by CodePlay, but I'm not sure if they are meant for execution only, or also for data-transfers. @adurang might be more familiar with that aspect.

https://github.com/llvm/llvm-project/pull/149036


More information about the llvm-commits mailing list