1 | /* Copyright 2022 The TensorFlow Authors. All Rights Reserved. |
2 | |
3 | Licensed under the Apache License, Version 2.0 (the "License"); |
4 | you may not use this file except in compliance with the License. |
5 | You may obtain a copy of the License at |
6 | |
7 | http://www.apache.org/licenses/LICENSE-2.0 |
8 | |
9 | Unless required by applicable law or agreed to in writing, software |
10 | distributed under the License is distributed on an "AS IS" BASIS, |
11 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
12 | See the License for the specific language governing permissions and |
13 | limitations under the License. |
14 | ==============================================================================*/ |
15 | |
16 | #include "llvm/ADT/ArrayRef.h" |
17 | #include "llvm/ADT/STLExtras.h" |
18 | #include "llvm/ADT/SmallVector.h" |
19 | #include "mlir/Dialect/Func/IR/FuncOps.h" // from @llvm-project |
20 | #include "mlir/IR/Attributes.h" // from @llvm-project |
21 | #include "mlir/IR/Builders.h" // from @llvm-project |
22 | #include "mlir/IR/BuiltinOps.h" // from @llvm-project |
23 | #include "mlir/IR/BuiltinTypes.h" // from @llvm-project |
24 | #include "mlir/IR/Diagnostics.h" // from @llvm-project |
25 | #include "mlir/IR/Operation.h" // from @llvm-project |
26 | #include "mlir/IR/TypeUtilities.h" // from @llvm-project |
27 | #include "mlir/IR/Types.h" // from @llvm-project |
28 | #include "mlir/Transforms/Passes.h" // from @llvm-project |
29 | #include "tensorflow/compiler/mlir/tensorflow/ir/tf_types.h" |
30 | #include "tensorflow/compiler/mlir/tensorflow/utils/convert_tensor.h" |
31 | #include "tensorflow/dtensor/cc/constants.h" |
32 | #include "tensorflow/dtensor/mlir/dtensor_mlir_passes.h" |
33 | |
34 | namespace tensorflow { |
35 | namespace dtensor { |
36 | |
37 | namespace { |
38 | #define GEN_PASS_DEF_DTENSORANNOTATEGLOBALSHAPE |
39 | #include "tensorflow/dtensor/mlir/dtensor_passes.h.inc" |
40 | |
41 | // Sets `_global_shape` attributes to argument/return values of `function`. |
42 | void AnnotateFunctionArgRetvalGlobalShapes(mlir::func::FuncOp function, |
43 | mlir::OpBuilder* builder) { |
44 | for (const auto& argument_type_and_index : |
45 | llvm::enumerate(function.getArgumentTypes())) { |
46 | const int index = argument_type_and_index.index(); |
47 | const auto& argument_type = argument_type_and_index.value(); |
48 | // Extract TensorType from element of resource type to allow setting proper |
49 | // global shape of resource types. |
50 | if (auto resource_type = mlir::getElementTypeOrSelf(argument_type) |
51 | .dyn_cast<mlir::TF::ResourceType>()) { |
52 | auto subtype = resource_type.getSubtypes(); |
53 | if (subtype.size() == 1) { |
54 | // subtype returns a Array of TensorType -- if it contains more than one |
55 | // Tensor type, we give up extracting the single TensorType inside the |
56 | // subtype. |
57 | function.setArgAttr(index, kGlobalShapeDialectAttr, |
58 | ConvertTypeToTensorShapeAttr(subtype[0])); |
59 | } |
60 | } else { |
61 | function.setArgAttr(index, kGlobalShapeDialectAttr, |
62 | ConvertTypeToTensorShapeAttr(argument_type)); |
63 | } |
64 | } |
65 | |
66 | for (const auto& retval_type_and_index : |
67 | llvm::enumerate(function.getFunctionType().getResults())) { |
68 | const int index = retval_type_and_index.index(); |
69 | const auto& retval_type = retval_type_and_index.value(); |
70 | function.setResultAttr(index, kGlobalShapeDialectAttr, |
71 | ConvertTypeToTensorShapeAttr(retval_type)); |
72 | } |
73 | } |
74 | |
75 | // Sets `_global_shape` attribute of an `op` with array of ShapeAttr of |
76 | // `outputs. |
77 | void AnnotateOperationGlobalShape(mlir::Operation* op, |
78 | mlir::OpBuilder* builder) { |
79 | llvm::SmallVector<mlir::Attribute, 4> op_global_shape; |
80 | op_global_shape.reserve(op->getNumResults()); |
81 | |
82 | for (const auto& result_type : op->getResultTypes()) |
83 | op_global_shape.emplace_back(ConvertTypeToTensorShapeAttr(result_type)); |
84 | |
85 | op->setAttr(kGlobalShape, builder->getArrayAttr(op_global_shape)); |
86 | } |
87 | |
88 | // Pass that annotates function argument/return values and all operation with |
89 | // `_global_shape` attribute. This will be used during SPMD expansion to |
90 | // preserve original global shape of operations in graph after shape has been |
91 | // modified to local shape. |
92 | struct DTensorAnnotateGlobalShape |
93 | : public impl::DTensorAnnotateGlobalShapeBase<DTensorAnnotateGlobalShape> { |
94 | void runOnOperation() override { |
95 | mlir::MLIRContext& context = getContext(); |
96 | mlir::OpBuilder builder(&context); |
97 | |
98 | auto module = getOperation(); |
99 | module.walk([&](mlir::func::FuncOp function) { |
100 | if (function.empty()) return; |
101 | |
102 | auto* terminator = function.getBody().front().getTerminator(); |
103 | AnnotateFunctionArgRetvalGlobalShapes(function, &builder); |
104 | function.getBody().walk([&](mlir::Operation* op) { |
105 | if (op == terminator) return; |
106 | |
107 | AnnotateOperationGlobalShape(op, &builder); |
108 | }); |
109 | }); |
110 | } |
111 | }; |
112 | |
113 | } // namespace |
114 | |
115 | std::unique_ptr<mlir::OperationPass<mlir::ModuleOp>> |
116 | CreateDTensorAnnotateGlobalShape() { |
117 | return std::make_unique<DTensorAnnotateGlobalShape>(); |
118 | } |
119 | |
120 | } // namespace dtensor |
121 | } // namespace tensorflow |
122 | |