[Mlir-commits] [mlir] [mlir][tosa] Forward concat insert_slice destination into DPS provider (PR #183490)

[Hsiangkai Wang]

================
@@ -454,11 +456,84 @@ struct ConcatConverter : public OpConversionPattern<tosa::ConcatOp> {
   }
 };
 
+// Forward the destination tensor of concat generated tensor.insert_slice ops
+// into single-use destination-style tensor producers. This avoids creating a
+// producer on a temporary tensor that is immediately copied into the concat
+// result tensor.
----------------
Hsiangkai wrote:

Give an example  to show the effect of applying the pattern. Something like,

Before:
`
  %0 = tensor.empty() : tensor<4xf32>
  %1 = linalg.fill ins(%cst : f32) outs(%0 : tensor<4xf32>) -> tensor<4xf32>
  %2 = tensor.empty() : tensor<8xf32>
  %inserted_slice = tensor.insert_slice %1 into %2[0] [4] [1] : tensor<4xf32> into tensor<8xf32>
  %inserted_slice_1 = tensor.insert_slice %arg0 into %inserted_slice[4] [4] [1] : tensor<4xf32> into tensor<8xf32>
`

After:
`
  %2 = tensor.empty() : tensor<8xf32>
  %ext = tensor.extract_slice %2 ... : tensor<4xf32>
  %1 = linalg.fill ins(%cst : f32) outs(%ext : tensor<4xf32>) -> tensor<4xf32>
  %inserted_slice = tensor.insert_slice %1 into %2[0] [4] [1] : tensor<4xf32> into tensor<8xf32>
  %inserted_slice_1 = tensor.insert_slice %arg0 into %inserted_slice[4] [4] [1] : tensor<4xf32> into tensor<8xf32>
`

https://github.com/llvm/llvm-project/pull/183490