[Mlir-commits] [mlir] [mlir][linalg] Vectorize directly to a named contraction (PR #147296)

[Andrzej Warzyński]

================
@@ -0,0 +1,400 @@
+// RUN: mlir-opt %s -transform-interpreter -split-input-file | FileCheck %s
+
+func.func @matmul(%A: tensor<8x4xf32>, %B: tensor<4x16xf32>,
+    %C: tensor<8x16xf32>) -> tensor<8x16xf32> {
+  %0 = linalg.matmul
+    ins(%A, %B : tensor<8x4xf32>, tensor<4x16xf32>)
+    outs(%C: tensor<8x16xf32>) -> tensor<8x16xf32>
+  return %0 : tensor<8x16xf32>
+}
+
+// CHECK: #[[$MAP_A:.+]] = affine_map<(d0, d1, d2) -> (d0, d2)>
+// CHECK: #[[$MAP_B:.+]] = affine_map<(d0, d1, d2) -> (d2, d1)>
+// CHECK: #[[$MAP_C:.+]] = affine_map<(d0, d1, d2) -> (d0, d1)>
+// CHECK-LABEL: func.func @matmul(
+// CHECK-SAME:    %[[A:.*]]: tensor<8x4xf32>, %[[B:.*]]: tensor<4x16xf32>,
+// CHECK-SAME:    %[[C:.*]]: tensor<8x16xf32>)
+//      CHECK: %[[LOAD_A:.*]] = vector.transfer_read %[[A]]{{.*}}: tensor<8x4xf32>, vector<8x4xf32>
+//      CHECK: %[[LOAD_B:.*]] = vector.transfer_read %[[B]]{{.*}}: tensor<4x16xf32>, vector<4x16xf32>
+//      CHECK: %[[LOAD_C:.*]] = vector.transfer_read %[[C]]{{.*}}: tensor<8x16xf32>, vector<8x16xf32>
+//      CHECK: %[[CONTRACT:.*]] = vector.contract
+// CHECK-SAME:   indexing_maps = [#[[$MAP_A]], #[[$MAP_B]], #[[$MAP_C]]]
+// CHECK-SAME:   kind = #vector.kind<add>
+// CHECK-SAME:   %[[LOAD_A]], %[[LOAD_B]], %[[LOAD_C]]
+// CHECK: vector.transfer_write %[[CONTRACT]], %[[C]]{{.*}}: vector<8x16xf32>, tensor<8x16xf32>
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
+    %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
+    transform.structured.vectorize %0 {create_named_contraction} : !transform.any_op
+    transform.yield
+  }
+}
+
+// -----
+
+func.func @matmul_dynamic(%A: tensor<?x?xf32>, %B: tensor<?x?xf32>,
+    %C: tensor<?x?xf32>) -> tensor<?x?xf32> {
+  %0 = linalg.matmul
+    ins(%A, %B : tensor<?x?xf32>, tensor<?x?xf32>)
+    outs(%C: tensor<?x?xf32>) -> tensor<?x?xf32>
+  return %0 : tensor<?x?xf32>
+}
+
+// CHECK: #[[$MAP_A:.+]] = affine_map<(d0, d1, d2) -> (d0, d2)>
+// CHECK: #[[$MAP_B:.+]] = affine_map<(d0, d1, d2) -> (d2, d1)>
+// CHECK: #[[$MAP_C:.+]] = affine_map<(d0, d1, d2) -> (d0, d1)>
+// CHECK-LABEL: func.func @matmul_dynamic(
+// CHECK-SAME:    %[[A:.*]]: tensor<?x?xf32>, %[[B:.*]]: tensor<?x?xf32>,
+// CHECK-SAME:    %[[C:.*]]: tensor<?x?xf32>)
+//      CHECK: %[[LOAD_A:.*]] = vector.mask{{.*}}{ vector.transfer_read %[[A]]{{.*}}: tensor<?x?xf32>, vector<8x4xf32>
+//      CHECK: %[[LOAD_B:.*]] = vector.mask{{.*}}{ vector.transfer_read %[[B]]{{.*}}: tensor<?x?xf32>, vector<4x16xf32>
+//      CHECK: %[[LOAD_C:.*]] = vector.mask{{.*}}{ vector.transfer_read %[[C]]{{.*}}: tensor<?x?xf32>, vector<8x16xf32>
+//      CHECK: %[[CONTRACT:.*]] = vector.contract
+// CHECK-SAME:   indexing_maps = [#[[$MAP_A]], #[[$MAP_B]], #[[$MAP_C]]]
+// CHECK-SAME:   kind = #vector.kind<add>
+// CHECK-SAME:   %[[LOAD_A]], %[[LOAD_B]], %[[LOAD_C]]
+// CHECK: vector.mask{{.*}}{ vector.transfer_write %[[CONTRACT]], %[[C]]{{.*}}: vector<8x16xf32>, tensor<?x?xf32>
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
+    %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
+    transform.structured.vectorize %0 vector_sizes [8, 16, 4]
+      {create_named_contraction} : !transform.any_op
+    transform.yield
+  }
+}
+
+// -----
+
+func.func @matmul_dynamic_memref(%A: memref<?x?xf32>, %B: memref<?x?xf32>,
+    %C: memref<?x?xf32>) {
+  linalg.matmul
+    ins(%A, %B : memref<?x?xf32>, memref<?x?xf32>)
+    outs(%C: memref<?x?xf32>)
+  return
+}
+
+// CHECK: #[[$MAP_A:.+]] = affine_map<(d0, d1, d2) -> (d0, d2)>
+// CHECK: #[[$MAP_B:.+]] = affine_map<(d0, d1, d2) -> (d2, d1)>
+// CHECK: #[[$MAP_C:.+]] = affine_map<(d0, d1, d2) -> (d0, d1)>
+// CHECK-LABEL: func.func @matmul_dynamic_memref(
+// CHECK-SAME:    %[[A:.*]]: memref<?x?xf32>, %[[B:.*]]: memref<?x?xf32>,
+// CHECK-SAME:    %[[C:.*]]: memref<?x?xf32>)
+//      CHECK: %[[LOAD_A:.*]] = vector.mask{{.*}}{ vector.transfer_read %[[A]]{{.*}}: memref<?x?xf32>, vector<8x4xf32>
+//      CHECK: %[[LOAD_B:.*]] = vector.mask{{.*}}{ vector.transfer_read %[[B]]{{.*}}: memref<?x?xf32>, vector<4x16xf32>
+//      CHECK: %[[LOAD_C:.*]] = vector.mask{{.*}}{ vector.transfer_read %[[C]]{{.*}}: memref<?x?xf32>, vector<8x16xf32>
+//      CHECK: %[[CONTRACT:.*]] = vector.contract
+// CHECK-SAME:   indexing_maps = [#[[$MAP_A]], #[[$MAP_B]], #[[$MAP_C]]]
+// CHECK-SAME:   kind = #vector.kind<add>
+// CHECK-SAME:   %[[LOAD_A]], %[[LOAD_B]], %[[LOAD_C]]
+// CHECK: vector.mask{{.*}}{ vector.transfer_write %[[CONTRACT]], %[[C]]{{.*}}: vector<8x16xf32>, memref<?x?xf32>
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
+    %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
+    transform.structured.vectorize %0 vector_sizes [8, 16, 4]
+      {create_named_contraction} : !transform.any_op
+    transform.yield
+  }
+}
+
+// -----
+
+func.func @matmul_scalable(%A: tensor<?x?xf32>, %B: tensor<?x?xf32>,
----------------
banach-space wrote:

[nit] Given that this is a variation of `@matmul_dynamic` ...
```suggestion
func.func @matmul_dynamic_scalable(%A: tensor<?x?xf32>, %B: tensor<?x?xf32>,
```

https://github.com/llvm/llvm-project/pull/147296