[Mlir-commits] [mlir] [MLIR][NVVM] Add TMA linear prefetch Op (PR #141211)

Durgadoss R llvmlistbot at llvm.org
Fri May 23 08:05:15 PDT 2025


================
@@ -2344,6 +2344,49 @@ def NVVM_PrefetchTensorMapOp : NVVM_Op<"prefetch.tensormap",
   }];
 }
 
+def NVVM_CpAsyncBulkPrefetchOp : NVVM_Op<"cp.async.bulk.prefetch"> {
----------------
durga4github wrote:

Yes Guray. I believe combining it with the `tensor` variant will complicate the Op and its maintenance.
The tensor variants have more additions towards Blackwell (in terms of modes etc.) but this non-tensor variant stays the same. So, keeping them separate will help them evolve in their own ways.




https://github.com/llvm/llvm-project/pull/141211


More information about the Mlir-commits mailing list