[Mlir-commits] [mlir] [mlir][Linalg]: Optimize any structured linalg operation in transform::PromoteOp to avoid unnecessary copies (PR #69876)

[Nicolas Vasilache]

================
@@ -54,7 +54,8 @@ func.func @matmul_f32(%A: memref<?xi8>, %M: index, %N: index, %K: index) {
 
 //       CHECK:         linalg.copy ins(%[[vA]] : memref<?x?xf32, strided<[?, 1], offset: ?>>) outs(%[[partialA]] : memref<?x?xf32, strided<[?, 1], offset: ?>>)
 //       CHECK:         linalg.copy ins(%[[vB]] : memref<?x?xf32, strided<[?, 1], offset: ?>>) outs(%[[partialB]] : memref<?x?xf32, strided<[?, 1], offset: ?>>)
-//       CHECK:         linalg.copy ins(%[[vC]] : memref<?x?xf32, strided<[?, 1], offset: ?>>) outs(%[[partialC]] : memref<?x?xf32, strided<[?, 1], offset: ?>>)
+//       CHECK-NOT:     linalg.copy ins(%[[vC]] : memref<?x?xf32, strided<[?, 1], offset: ?>>) outs(%[[partialC]] : memref<?x?xf32, strided<[?, 1], offset: ?>>)
----------------
nicolasvasilache wrote:

This seems wrong to me .. if you don't copy in the local buffer, your local buffer contains garbage and you run linalg.matmul on it, the conditions seem off somewhere.

https://github.com/llvm/llvm-project/pull/69876