[Mlir-commits] [mlir] [MLIR][Vector] Implement TransferOpReduceRank as MaskableOpRewritePattern (PR #92426)

[Hugo Trachino]

================
@@ -187,3 +187,49 @@ module attributes {transform.with_named_sequence} {
     transform.yield
   }
 }
+
+// -----
+
+
+//       CHECK:   #[[MAP:.*]] = affine_map<(d0, d1, d2, d3) -> (d1, 0, d3)>
+//       CHECK:   func.func @transfer_read_reduce_rank_scalable(
+//  CHECK-SAME:       %[[ARG_0:.*]]: memref<?x?x?x?xf32>) -> vector<8x[4]x2x3xf32> {
+//       CHECK:     %[[C0:.*]] = arith.constant 0 : index
+//       CHECK:     %[[TFR:.*]] = vector.transfer_read %arg0[%[[C0]], %[[C0]], %[[C0]], %[[C0]]]{{.*}} permutation_map = #[[MAP]]} : memref<?x?x?x?xf32>, vector<[4]x2x3xf32>
+//       CHECK:     %[[BC:.*]] = vector.broadcast %[[TFR]] : vector<[4]x2x3xf32> to vector<8x[4]x2x3xf32>
+//       CHECK:     return %[[BC]] : vector<8x[4]x2x3xf32>
+func.func @transfer_read_reduce_rank_scalable(%mem: memref<?x?x?x?xf32>) -> vector<8x[4]x2x3xf32> {
----------------
nujaa wrote:

There is a transposition in the permutation map of `permutation_with_mask_xfer_read_scalable`. It would hence also trigger `TransferReadPermutationLowering` which IIRC would fail without https://github.com/llvm/llvm-project/pull/91987 . Changes were not completely independent, this one is a bit more _unit_ test-like

https://github.com/llvm/llvm-project/pull/92426