[Mlir-commits] [mlir] [mlir] Add direct vectorization lowering for `tensor.pack` ops (PR #78660)

[Han-Chung Wang]

================
@@ -1393,6 +1400,182 @@ vectorizeAsLinalgGeneric(RewriterBase &rewriter, VectorizationState &state,
   return success();
 }
 
+/// Given a tensor::PackOp, return the permutation from the "tiled"
+/// shape to the "packed" shape, defined as the following:
+/// The "packed" shape is the same as the `dest` shape of the pack op.
+/// The "tiled" shape is a permutation of the `dest` shape such that
+/// each outer dimension is in the original `source` order, and the
+/// inner_tile dimensions immediately follow their corresponding outer
+/// dimension.
+/// i.e. for the following tensor.pack:
+/// ```mlir
+/// %pack = tensor.pack %0 padding_value(%1)
+///   outer_dims_perm = [0, 2, 1]
+///   inner_dims_pos = [2, 1]
+///   inner_tiles = [16, 2]
+///   into %2 : tensor<32x8x16> -> tensor<32x1x4x16x2>
+/// ```
+/// The "packed" shape is `32x1x4x16x2`
+/// The "tiled" shape is `32x(4x2)x(1x16)`
+static SmallVector<int64_t>
+getTiledShapeToPackedShapePerm(tensor::PackOp packOp) {
----------------
hanhanW wrote:

I think this is doing something similar to a snippet of logic in lowerPack, which also needs the permutation for transpose op. It would be good if we can refactor it to `Tensor/Utils/Utils.[cpp|h]` and use it in both places.

https://github.com/llvm/llvm-project/blob/a31a60074717fc40887cfe132b77eec93bedd307/mlir/lib/Dialect/Linalg/Transforms/Transforms.cpp#L235-L265

https://github.com/llvm/llvm-project/pull/78660