[Mlir-commits] [mlir] [mlir][AMDGPU] Allow non-contiguous destination memrefs for gather_to_lds (PR #152559)

[Alan Li]

================
@@ -518,9 +518,6 @@ LogicalResult GatherToLDSOp::verify() {
   MemRefType srcType = cast<MemRefType>(getSrc().getType());
   MemRefType dstType = cast<MemRefType>(getDst().getType());
 
-  if (!dstType.areTrailingDimsContiguous(dstType.getRank()))
-    return emitOpError("destination types must be contiguous");
-
----------------
lialan wrote:

Sorry I was wrong. If we consider the subgroup ID is 1dimensional, then the semantic of the op is to gather data into a consecutive memory address. In that case, I meant, we should at least do `dstType.areTrailingDimsContiguous(1)` because we want to make sure the last trailing dim is contiguous.

https://github.com/llvm/llvm-project/pull/152559