[Mlir-commits] [mlir] [mlir] Extend CombineTransferReadOpTranspose pattern to handle extf ops. (PR #74754)

[llvmlistbot at llvm.org]

================
@@ -460,3 +460,33 @@ func.func @cast_f16_to_f32_write(%arg0: memref<16x16xf16>, %arg1: memref<16x16xf
   vector.transfer_write %cast, %arg3[%c0, %c0] {in_bounds = [true, true]} : vector<16x16xf32>, memref<16x16xf32>
   return
 }
+
+// -----
+
+#map1 = affine_map<(d0, d1, d2) -> (d0, d2)>
+#map2 = affine_map<(d0, d1, d2) -> (d2, d1)>
+#map3 = affine_map<(d0, d1, d2) -> (d0, d1)>
+
+//   CHECK-DAG: #[[$MAP:.+]] = affine_map<(d0, d1) -> (d1, d0)>
+// CHECK-LABEL: func @fold_transpose_into_transfer_read(
+//  CHECK-SAME:      %[[ALLOC:.+]]: memref<64x128xf16>
+//   CHECK-DAG:      %[[C0:.+]] = arith.constant 0 : index
+//   CHECK-DAG:      %[[CST:.+]] = arith.constant 0.000000e+00 : f16
+//       CHECK:      %[[READ:.+]] = vector.transfer_read %[[ALLOC]][%[[C0]], %[[C0]]], %[[CST]] {in_bounds = [true, true], permutation_map = #[[$MAP]]}
+//       CHECK:      %[[EXTF1:.+]] = arith.extf %[[READ]]
+//   CHECK-NOT:      vector.transpose
+//       CHECK:      %[[RESULT:.+]] = vector.contract
+func.func @fold_transpose_into_transfer_read(%alloc: memref<64x128xf16>, %vector: vector<32x128xf16>, %alloc2: memref<32x64xf32>) {
----------------
harsh-nod wrote:

I think that makes it a little more confusing because then it seems like we are folding the transpose and extf into the transfer read (when we are only folding the transpose into the transfer read).

https://github.com/llvm/llvm-project/pull/74754