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

[Andrzej Warzyński]

================
@@ -0,0 +1,49 @@
+// RUN: mlir-opt %s -linalg-generalize-named-ops | mlir-opt --linalg-specialize-generic-ops | FileCheck %s
+
+func.func @matmul(%arg0: tensor<?x?xf32>, %arg1: tensor<?x?xf32>, %arg2: tensor<?x?xf32>) -> tensor<?x?xf32> {
+  %0 = linalg.matmul ins(%arg0, %arg1 : tensor<?x?xf32>, tensor<?x?xf32>) outs(%arg2 : tensor<?x?xf32>) -> tensor<?x?xf32>
+  return %0 : tensor<?x?xf32>
+}
+
+// CHECK-LABEL: @matmul
+// CHECK-SAME: %[[A:.+]]: tensor<?x?xf32>, %[[B:.+]]: tensor<?x?xf32>, %[[Out:.+]]: tensor<?x?xf32>) -> tensor<?x?xf32>
+// CHECK-NOT: linalg.generic
+// CHECK: linalg.matmul ins(%[[A]], %[[B]] : tensor<?x?xf32>, tensor<?x?xf32>) outs(%[[Out]] : tensor<?x?xf32>) -> tensor<?x?xf32>
+
+// -----
+
+func.func @add(%arg0: tensor<?x?xf32>, %arg1: tensor<?x?xf32>, %arg2: tensor<?x?xf32>) -> tensor<?x?xf32> {
+  %0 = linalg.add ins(%arg0, %arg1 : tensor<?x?xf32>, tensor<?x?xf32>) outs(%arg2 : tensor<?x?xf32>) -> tensor<?x?xf32>
+  return %0 : tensor<?x?xf32>
+}
+
+// CHECK-LABEL: add
+// CHECK-SAME: %[[A:.+]]: tensor<?x?xf32>, %[[B:.+]]: tensor<?x?xf32>,  %[[Out:.+]]: tensor<?x?xf32>) -> tensor<?x?xf32>
+// CHECK-NOT: linalg.generic
+// CHECK: linalg.add ins(%[[A]], %[[B]] : tensor<?x?xf32>, tensor<?x?xf32>) outs(%[[Out]] : tensor<?x?xf32>) -> tensor<?x?xf32>
+
+// -----
+
+func.func @exp(%arg: memref<7x14x21xf32>, %out: memref<7x14x21xf32>) {
----------------
banach-space wrote:

This is the first test to use static shapes. Is the split into static/dynamic shapes significant here? if not, I suggest duplicating the first test so that there's static and dynamic variant (as a sanity check). Following from that, just stick with dynamic shapes. 

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