1 /* Copyright 2019 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 #ifndef TENSORFLOW_LITE_DELEGATES_GPU_COMMON_MODEL_TRANSFORMER_H_ 17 #define TENSORFLOW_LITE_DELEGATES_GPU_COMMON_MODEL_TRANSFORMER_H_ 18 19 #include <deque> 20 #include <string> 21 #include <unordered_set> 22 #include <vector> 23 24 #include "tensorflow/lite/delegates/gpu/common/model.h" 25 26 namespace tflite { 27 namespace gpu { 28 29 class TransformationReporter; 30 31 struct TransformationContext { 32 GraphFloat32* graph; 33 TransformationReporter* reporter; 34 }; 35 36 enum class TransformStatus { 37 // Transformation was not applied due to trivial conditions mismatch. 38 // 39 // This is different from DECLINED code below that provides in-depth 40 // explanation why a transformation that could have been applied but was not 41 // due to some issues. 42 SKIPPED, 43 44 // Transformation was declined, therefore, a model was not modified. 45 DECLINED, 46 47 // Transformation was applied successfully 48 APPLIED, 49 50 // Transformation may partially be applied, but left a model in an invalid 51 // state. This error should be considered unrecoverable. 52 INVALID, 53 }; 54 55 struct TransformResult { 56 TransformStatus status; 57 std::string message; 58 }; 59 60 // Class responsible for applying a transformation to a single node. 61 class NodeTransformation { 62 public: 63 virtual ~NodeTransformation() = default; 64 65 virtual TransformResult ApplyToNode(Node* node, GraphFloat32* graph) = 0; 66 }; 67 68 // Class responsible for applying a transformation to a sequence of nodes. 69 // Nodes are guaranteed to depend on each other without extra dependents being 70 // spilled. 71 class SequenceTransformation { 72 public: 73 virtual ~SequenceTransformation() = default; 74 75 // @return number of nodes in a sequence to apply this transformation. 76 virtual int ExpectedSequenceLength() const = 0; 77 78 // Applies transformations to a sequence of nodes. Transformation 79 // implementation is free manipulate with sequence nodes including adding 80 // and/or deleting nodes. if there were updates to nodes in the end and/or 81 // beginning of the sequence, then referential consistency should be 82 // maintained by updating relevant references in nodes that precede this 83 // sequence or depend on a last node of the sequence. 84 virtual TransformResult ApplyToNodesSequence( 85 const std::vector<Node*>& sequence, GraphFloat32* graph) = 0; 86 }; 87 88 // A class accumulated decisions or updates done by transformations. 89 class TransformationReporter { 90 public: 91 virtual ~TransformationReporter() = default; 92 93 virtual void DeclinedTransformation(const std::string& transformation, 94 const std::string& node_ids, 95 const std::string& message) = 0; 96 97 virtual void AppliedTransformation(const std::string& transformation, 98 const std::string& node_ids, 99 const std::string& message) = 0; 100 }; 101 102 // A class is designed to perform model transformations. 103 class ModelTransformer { 104 public: ModelTransformer(GraphFloat32 * graph,TransformationReporter * reporter)105 ModelTransformer(GraphFloat32* graph, TransformationReporter* reporter) 106 : graph_(graph), reporter_(reporter) {} 107 108 // @return false if a graph is in the broken states can not be used any more 109 bool Apply(const std::string& name, SequenceTransformation* transformation); 110 111 // @return false if a graph is in the broken states can not be used any more 112 bool Apply(const std::string& name, NodeTransformation* transformation); 113 114 private: 115 bool ApplyStartingWithNode(const std::string& name, 116 SequenceTransformation* transformation, 117 Node* begin); 118 AddNodeToProcess(Node * node)119 void AddNodeToProcess(Node* node) { 120 if (node && processed_.insert(node->id).second) { 121 to_process_.push_back(node->id); 122 } 123 } 124 125 GraphFloat32* graph_; 126 TransformationReporter* reporter_; 127 128 std::deque<NodeId> to_process_; 129 std::unordered_set<NodeId> processed_; 130 }; 131 132 class NullTransformationReporter : public TransformationReporter { 133 public: DeclinedTransformation(const std::string & transformation,const std::string & nodes_id,const std::string & message)134 void DeclinedTransformation(const std::string& transformation, 135 const std::string& nodes_id, 136 const std::string& message) override {} 137 AppliedTransformation(const std::string & transformation,const std::string & nodes_id,const std::string & message)138 void AppliedTransformation(const std::string& transformation, 139 const std::string& nodes_id, 140 const std::string& message) override {} 141 }; 142 143 } // namespace gpu 144 } // namespace tflite 145 146 #endif // TENSORFLOW_LITE_DELEGATES_GPU_COMMON_MODEL_TRANSFORMER_H_ 147