[Mlir-commits] [mlir] [MLIR][Linalg] Add pass to convert linalg.generic back to named ops (PR #95656)

[Andrzej Warzyński]

================
@@ -79,3 +79,47 @@ func.func @op_batch_matmul(%A: tensor<2x16x8xf32>, %B: tensor<2x8x16xf32>, %Out:
 // CHECK-SAME: %[[A:.+]]: tensor<2x16x8xf32>, %[[B:.+]]: tensor<2x8x16xf32>,  %[[Out:.+]]: tensor<2x16x16xf32>) -> tensor<2x16x16xf32>
 // CHECK-NOT: linalg.generic
 // CHECK: linalg.batch_matmul ins(%[[A]], %[[B]] : tensor<2x16x8xf32>, tensor<2x8x16xf32>) outs(%[[Out]] : tensor<2x16x16xf32>) -> tensor<2x16x16xf32>
+
+// -----
+
+// This is a multi-reduction linalg.generic and cannot be lifted to matrix multiply
+#mapA = affine_map<(m, n, k1, k2) -> (m, k1, k2)>
+#mapB = affine_map<(m, n, k1, k2) -> (k2, k1, n)>
+#mapC = affine_map<(m, n, k1, k2) -> (m, n)>
+func.func @op_multi_reduction(%A: tensor<10x20x30xf32>, %B: tensor<30x20x40xf32>, 
----------------
banach-space wrote:

[nit] You could add "negative" to the test name (e.g. `@negative_op_multi_reduction`)

https://github.com/llvm/llvm-project/pull/95656