[Mlir-commits] [mlir] [mlir][memref] Update tests to use memref.assume_alignment properly. (PR #142358)
Han-Chung Wang
llvmlistbot at llvm.org
Mon Jun 2 03:50:28 PDT 2025
================
@@ -226,13 +226,13 @@ func.func @main() {
scf.for %arg12 = %17 to %c128 step %c4 {
%19 = arith.muli %18, %c4 : index
%20 = vector.load %accShmemPtr[%arg12, %19] : memref<128x128xf32, 3>, vector<4xf32>
- vector.store %20, %matrixD[%arg12, %19] : memref<128x128xf32>, vector<4xf32>
+ vector.store %20, %align_matrixD[%arg12, %19] : memref<128x128xf32>, vector<4xf32>
}
gpu.terminator
}
// Step 5. Copy D2H
- %5 = gpu.memcpy async [%token] %matrixDHost, %matrixD : memref<128x128xf32>, memref<128x128xf32>
+ %5 = gpu.memcpy async [%token] %matrixDHost, %align_matrixD : memref<128x128xf32>, memref<128x128xf32>
----------------
hanhanW wrote:
I'm not sure if we want to use `align_matrixD` here or not, so I move the declaration outside the GPU kernel block. I'm happy to move it back, if my change is not correct.
https://github.com/llvm/llvm-project/pull/142358
More information about the Mlir-commits
mailing list