[Mlir-commits] [mlir] [mlir][vector] Drop innermost unit dims on transfer_write. (PR #78554)

[Han-Chung Wang]

================
@@ -76,3 +76,24 @@ func.func @contiguous_inner_most_dim_out_of_bounds_2d(%arg0: memref<1x1xf32>) ->
 //  CHECK-NOT:   memref.subview
 //      CHECK:   %[[READ:.+]] = vector.transfer_read %[[SRC]]
 //      CHECK:   return %[[READ]] : vector<4x8xf32>
+
+// -----
+
+func.func @drop_inner_most_dim_for_transfer_write(%arg0: memref<1x512x16x1xf32, strided<[8192, 16, 1, 1], offset: ?>>, %arg1: vector<1x16x16x1xf32>, %arg2: index) {
+  %c0 = arith.constant 0 : index
+  vector.transfer_write %arg1, %arg0[%c0, %arg2, %c0, %c0]
+    {in_bounds = [true, true, true, true]}
+    : vector<1x16x16x1xf32>, memref<1x512x16x1xf32, strided<[8192, 16, 1, 1], offset: ?>>
+  return
+}
+// CHECK:      func.func @drop_inner_most_dim_for_transfer_write
+// CHECK-SAME:   %[[DEST:[a-zA-Z0-9]+]]
----------------
hanhanW wrote:

This is actually intended. It's a trick that I learned from other contributors. In this context, we don't have to check `tensor<xxxxxx>` things, so it makes life much easier.

https://github.com/llvm/llvm-project/pull/78554