[Mlir-commits] [mlir] [MLIR][XeGPU] Extend unrolling support for scatter ops with chunk_size (PR #144447)
Chao Chen
llvmlistbot at llvm.org
Tue Jun 17 09:40:25 PDT 2025
================
@@ -462,10 +492,32 @@ struct UnrollLoadGatherOp : public UnrollPattern<xegpu::LoadGatherOp> {
SmallVector<Value> convertedTdescs = pack(
op.getTensorDesc(), convertedTdescTypes, *targetShape, loc, rewriter);
- SmallVector<Type> convertedMaskTypes =
- getUnrolledTypes(maskTy, *targetShape);
- SmallVector<Value> convertedMasks =
- pack(op.getMask(), convertedMaskTypes, *targetShape, loc, rewriter);
+ SmallVector<Type> convertedMaskTypes;
+ SmallVector<Value> convertedMasks;
+
+ if (originalChunkSize > 1) {
+ targetMaskShape.pop_back();
+ convertedMaskTypes = getUnrolledTypes(maskTy, targetMaskShape);
+ SmallVector<Value> convertedMasks1D = pack(
----------------
chencha3 wrote:
nit: convertedMasks1D --> convertedMasks
https://github.com/llvm/llvm-project/pull/144447
More information about the Mlir-commits
mailing list