[Mlir-commits] [mlir] [MLIR][Linalg] Pattern to fold AddOp to accumulation via contraction op's dest (PR #110514)

[Rolf Morel]

================
@@ -0,0 +1,291 @@
+// RUN: mlir-opt %s -transform-interpreter -cse -split-input-file | FileCheck %s
+
+!type = tensor<2048x2048xf32>
+func.func @fold_add_on_two_matmuls(%arg0: !type, %arg1: !type) -> !type {
+  %0 = arith.constant dense<1.111111e+00> : !type
+  %cst = arith.constant 0.000000e+00 : f32
+  %1 = tensor.empty() : !type
+  %2 = linalg.fill ins(%cst : f32) outs(%1 : !type) -> !type
+  %3 = linalg.matmul ins(%arg0, %0 : !type, !type) outs(%2 : !type) -> !type
+  %4 = linalg.matmul ins(%arg1, %0 : !type, !type) outs(%2 : !type) -> !type
+  %5 = linalg.add ins(%3, %4 : !type, !type) outs(%1 : !type) -> !type
+  return %5 : !type
+}
+
+// CHECK-LABEL: func.func @fold_add_on_two_matmuls
+// CHECK: %[[ACC:.+]] = linalg.matmul
----------------
rolfmorel wrote:

Have converted one test case to now include full details regarding arguments. I have left out types as only arguments get shuffled around and their shapes ought to be same before the rewrite already.

https://github.com/llvm/llvm-project/pull/110514