[Mlir-commits] [mlir] [MLIR][Transform] FuseOp: accept transform params, add use_forall argument (PR #161883)

[Tuomas Kärnä]

================
@@ -665,24 +759,69 @@ transform::FuseOp::apply(transform::TransformRewriter &rewriter,
 }
 
 LogicalResult transform::FuseOp::verify() {
-  SmallVector<int64_t> permutation =
-      extractFromIntegerArrayAttr<int64_t>(getTileInterchange());
-  auto sequence = llvm::to_vector(llvm::seq<int64_t>(0, permutation.size()));
-  if (!std::is_permutation(sequence.begin(), sequence.end(),
-                           permutation.begin(), permutation.end())) {
-    return emitOpError() << "expects interchange to be a permutation, found "
-                         << getTileInterchange();
+  ArrayRef<int64_t> permutation = getStaticTileInterchange();
+  if (!llvm::any_of(permutation,
+                    [](int64_t v) { return ShapedType::isDynamic(v); })) {
----------------
tkarna wrote:

> what does interchange do when loop type is forall?

As one would expect, it changes the order of the iteration variables:

```cpp
%0 = scf.forall (%arg3, %arg4) = (0, 0) to (2048, 1024) step (32, 64) shared_outs(%arg5 = %arg2) -> (tensor<1024x2048xf32>) {
  %slice_0 = tensor.extract_slice %arg0[%arg4, 0] [64, 512] [1, 1] : tensor<1024x512xf16> to tensor<64x512xf16>
  %slice_1 = tensor.extract_slice %arg1[0, %arg3] [512, 32] [1, 1] : tensor<512x2048xf16> to tensor<512x32xf16>
  ...
```
with `interchange = [1, 0]` turns into
```cpp
%0 = scf.forall (%arg3, %arg4) = (0, 0) to (1024, 2048) step (64, 32) shared_outs(%arg5 = %arg2) -> (tensor<1024x2048xf32>) {
  %slice_0 = tensor.extract_slice %arg0[%arg3, 0] [64, 512] [1, 1] : tensor<1024x512xf16> to tensor<64x512xf16>
  %slice_1 = tensor.extract_slice %arg1[0, %arg4] [512, 32] [1, 1] : tensor<512x2048xf16> to tensor<512x32xf16>
  ...
```

https://github.com/llvm/llvm-project/pull/161883