[Mlir-commits] [mlir] [MLIR][XeGPU][TransformOps] Add insert_prefetch op (PR #167356)

[Sang Ik Lee]

================
@@ -252,3 +252,79 @@ module attributes {transform.with_named_sequence} {
     transform.yield
   }
 }
+
+// -----
+
+// CHECK-LABEL: @insert_prefetch_dpas_a
+func.func @insert_prefetch_dpas_a(%arg0: memref<4096x4096xf16>, %arg1: memref<4096x4096xf16>, %arg2: memref<4096x4096xf16>) {
+  %c32 = arith.constant 32 : index
+  %c4096 = arith.constant 4096 : index
+  %c0 = arith.constant 0 : index
+  %0 = xegpu.create_nd_tdesc %arg2 : memref<4096x4096xf16> -> !xegpu.tensor_desc<256x256xf16>
+  %1 = xegpu.load_nd %0[%c0, %c0]  : !xegpu.tensor_desc<256x256xf16> -> vector<256x256xf16>
+  // CHECK: xegpu.create_nd_tdesc %arg0
+  // CHECK: xegpu.create_nd_tdesc %arg1
+  // CHECK: %[[V0:.+]] = xegpu.create_nd_tdesc %arg0
+  // CHECK-SAME: !xegpu.tensor_desc<256x32xf16
+  // CHECK: xegpu.prefetch_nd %[[V0]]
----------------
silee2 wrote:

I think the tests should check for generated offsets for prefetch_nd ops.
Especially for the in loop prefetch_nd that needs offsets adjusted based on `nb_prefetch`

https://github.com/llvm/llvm-project/pull/167356