[Mlir-commits] [mlir] [mlir] `im2col` & `l2cache` on cp.async.bulk.tensor.shared.cluster.global` (PR #72967)

[llvmlistbot at llvm.org]

================
@@ -1404,20 +1404,34 @@ def NVVM_CpAsyncBulkTensorGlobalToSharedClusterOp :
   AttrSizedOperandSegments]>,
   Arguments<(ins  LLVM_PointerShared:$dstMem,
                   LLVM_AnyPointer:$tmaDescriptor,
-                  LLVM_PointerShared:$mbar,
-                  Optional<I16>:$multicastMask,                  
                   Variadic<I32>:$coordinates,
+                  LLVM_PointerShared:$mbar,                  
+                  Variadic<I16>:$im2colOffsets,
----------------
durga4github wrote:

ok, I was thinking we may need to provide an empty-vector like construct to take care of 0-sized im2col.
Not needing to provide that is even better !

+1. This change looks good to me.

https://github.com/llvm/llvm-project/pull/72967