[Mlir-commits] [mlir] [mlir][xegpu] Allow `xegpu.insert_prefetch` to override Load op's tile size. (PR #215419)

[Tuomas Kärnä]

================
@@ -156,18 +157,29 @@ def InsertPrefetchOp : Op<Transform_Dialect, "xegpu.insert_prefetch", [
     The load op must reside within the `scf.for` loop. Number of prefetch steps
     is set by the `nb_prefetch` argument (default value is 1). Returns a handle
     to the created `xegpu.create_nd_desc` op.
+
+    By default the prefetched tile shape is taken from the load op's tensor
----------------
tkarna wrote:

Having a transform op would give more flexibility to address load/prefetch patterns that upstream passes do you support yet. As such this extension looks good to me. In addition, if there will be an insert-prefetch cpp pattern in the future it would be good to expose that as a transform op as well.

https://github.com/llvm/llvm-project/pull/215419