[Mlir-commits] [mlir] [mlir][tensor] Introduce `FoldTensorCastUnPackOp` (PR #121393)

[Andrzej Warzyński]

================
@@ -2814,6 +2815,26 @@ func.func @fold_cast_pack_dynamic_tile_size(
 
 // -----
 
+// CHECK-LABEL:   func.func @fold_cast_unpack_dynamic_tile_size(
+// CHECK-SAME:      %[[SRC:.*]]: tensor<1x1x8x1xi32>,
+// CHECK-SAME:      %[[DEST:.*]]: tensor<7x?xi32>) -> tensor<7x?xi32> {
+// CHECK:           %[[RES:.*]] = tensor.unpack %[[SRC]] inner_dims_pos = [0, 1] inner_tiles = [8, 1] into %[[DEST]] {some_attr} : tensor<1x1x8x1xi32> -> tensor<7x?xi32>
+// CHECK:           return %[[RES]] : tensor<7x?xi32>
+func.func @fold_cast_unpack_dynamic_tile_size(
+  %src: tensor<1x1x8x1xi32>,
+  %res: tensor<7x?xi32>) -> tensor<7x?xi32> {
+
+    %cast = tensor.cast %src : tensor<1x1x8x1xi32> to tensor<1x1x?x1xi32>
+    %c8 = arith.constant 8 : index
+    %unpack = tensor.unpack %cast
+      inner_dims_pos = [0, 1]
+      inner_tiles = [%c8, 1]
+      into %res {some_attr} : tensor<1x1x?x1xi32> -> tensor<7x?xi32>
----------------
banach-space wrote:

I've actually followed the test above https://github.com/llvm/llvm-project/blob/45e874e39030bc622ea43fbcfc4fcdd1dd404353/mlir/test/Dialect/Tensor/canonicalize.mlir#L2791-L2813

😂 Let me unify this.

https://github.com/llvm/llvm-project/pull/121393