1 /* Copyright 2018 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_CORE_GRAPPLER_OPTIMIZERS_IMPLEMENTATION_SELECTOR_H_ 17 #define TENSORFLOW_CORE_GRAPPLER_OPTIMIZERS_IMPLEMENTATION_SELECTOR_H_ 18 19 #include <string> 20 21 #include "tensorflow/core/framework/op.h" 22 #include "tensorflow/core/grappler/costs/graph_properties.h" 23 #include "tensorflow/core/grappler/grappler_item.h" 24 #include "tensorflow/core/grappler/op_types.h" 25 #include "tensorflow/core/grappler/optimizers/custom_graph_optimizer.h" 26 #include "tensorflow/core/grappler/optimizers/custom_graph_optimizer_registry.h" 27 #include "tensorflow/core/grappler/optimizers/function_api_info.h" 28 #include "tensorflow/core/grappler/utils/graph_view.h" 29 #include "tensorflow/core/lib/core/errors.h" 30 #include "tensorflow/core/lib/core/stringpiece.h" 31 #include "tensorflow/core/lib/strings/strcat.h" 32 #include "tensorflow/core/util/device_name_utils.h" 33 34 namespace tensorflow { 35 namespace grappler { 36 37 // Motivation: To achieve the same high level functionality, the underlying 38 // implementations sometimes are different for various devices where the 39 // function runs. In order to achieve the correct result and best performance, 40 // the proper implementation needs to be picked dynamically. 41 // 42 // Currently there are two approaches to do this. 43 // (1) Utilize case op and dynamacically change the branch index. 44 // (2) Swap function implementation, it will be deprecated. 45 // 46 // Idea for approach 1. 47 // This transformation rewrites the DeviceIndex op with a Const op with value 48 // of the index of the device the associcated Case op runs. 49 // Example: 50 // def plus_one_gpu(x): return x + 1.0 51 // def plus_one_reference_implementation(x): return x + 1.0 52 // input = tf.constant(2.0, dtype=tf.float32) 53 // cpu_fn = lambda:plus_one_reference_implementation(input) 54 // gpu_fn = lambda:plus_one_gpu(input) 55 // control_flow_ops.execute_fn_for_device( 56 // {"CPU": cpu_fn, "GPU":gpu_fn)}, default_fn=cpu_fn) 57 // 58 // Idea for approach 2. 59 // This transformation replaces function calls by the appropriate function 60 // definition based on properties of the runtime system. For instance, 61 // we may choose one implementation over another if we have a GPU with 62 // enough memory available. 63 // 64 // It is a way for the programmer to specify alternative implementations 65 // of the same functionality in the graph, and let TensorFlow pick the 66 // most appropriate one at runtime. 67 // 68 // For instance, the python code might specify: 69 // @Defun(tf.float32, 70 // api_implements='plus_one', 71 // api_preferred_device='GPU') 72 // def plus_one_gpu(x): return x + 1.0 73 // 74 // @Defun(tf.float32, 75 // api_implements='plus_one') 76 // def plus_one_reference_implementation(x): return x + 1.0 77 // input = tf.constant(2.0, dtype=tf.float32) 78 // 79 // z = plus_one_reference_implementation(input) 80 // z = plus_one_gpu(input) 81 // print(sess.run(z)) 82 // 83 84 // At runtime, we will select either `plus_one_gpu` or 85 // `plus_one_reference_implementation` based on the availability of the GPU. 86 // 87 // Available annotations: 88 // - api_implements(string): all functions mapping to the same 89 // string can be interchanged. For now, all functions must have the same 90 // signature and overloads are not allowed. Defuns within defuns are 91 // allowed. 92 // - api_preferred_device(string): sets which device is preferred. 93 class ImplementationSelector : public CustomGraphOptimizer { 94 public: 95 ImplementationSelector() = default; 96 ~ImplementationSelector() override = default; Init(const tensorflow::RewriterConfig_CustomGraphOptimizer * config)97 Status Init( 98 const tensorflow::RewriterConfig_CustomGraphOptimizer* config) override { 99 return Status::OK(); 100 } name()101 string name() const override { 102 return "implementation_selector"; 103 } 104 UsesFunctionLibrary()105 bool UsesFunctionLibrary() const override { return false; } 106 107 // This call is not thread-safe. 108 Status Optimize(Cluster* cluster, const GrapplerItem& item, 109 GraphDef* optimized_graph) override; 110 111 // Does not take any feedback. Feedback(Cluster * cluster,const GrapplerItem & item,const GraphDef & optimized_graph,double result)112 void Feedback(Cluster* cluster, const GrapplerItem& item, 113 const GraphDef& optimized_graph, double result) override {} 114 115 private: 116 Status LoadFunctions(const GraphDef& graph); 117 Status MaybeOptimizeFunctionCall(utils::MutableNodeView* node_view) const; 118 119 // Finds all call sites for functions, then replace with the appropriate 120 // implementation. 121 // There are two ways of calling functions: 122 // 1. By specifying an op name as a function name, and 123 // 2. Via the functional interface, where the function name appears as an 124 // Attr. 125 // 126 // There may be multiple call sites for a given function. The function body 127 // may call into another function, so a function might have to be duplicated. 128 // For simplicity, we do not change function bodies. Also, we do not change 129 // gradients. 130 Status SelectImplementation(GraphDef* graph) const; 131 132 // Rewrites the DeviceIndex op with a Const op with value of the index of the 133 // device the associcated Case op runs. 134 135 // This function first looks up all the DeviceIndex ops. 136 // Then for each of these ops, it finds the device of the 137 // associated Case op that takes the DeviceIndex op as the input, and 138 // caculates the index of the device in the device list of DeviceIndex op. 139 // Lastly, it rewrites the DeviceIndex op with a Const op and sets the value 140 // to be the index. 141 // 142 // Example input nodes: 143 // node { 144 // name: "x" 145 // op: "DeviceIndex" 146 // device: "/device:CPU:0" 147 // attr { 148 // key: "device_names" 149 // value { 150 // list { 151 // s: "CPU" 152 // s: "TPU_REPLICATED_CORE" 153 // s: "GPU" 154 // } 155 // } 156 // } 157 // } 158 // node { 159 // name: "case" 160 // op: "Case" 161 // input: "x" 162 // device: "/device:GPU:0" 163 // ... 164 // } 165 // Example output nodes: 166 // 167 // name: "x" 168 // op: "Const" 169 // device: "/device:CPU:0" 170 // attr { 171 // key: "dtype" 172 // value { 173 // type: DT_INT32 174 // } 175 // } 176 // attr { 177 // key: "value" 178 // value { 179 // tensor { 180 // dtype: DT_INT32 181 // int_val: 2 182 // } 183 // } 184 // } 185 // node { 186 // name: "case" 187 // op: "Case" 188 // input: "x" 189 // device: "/device:GPU:0" 190 // ... 191 // } 192 Status SelectDeviceIndex(GraphDef* graph) const; 193 194 std::unique_ptr<FunctionLibraryApiInfo> lib_info_; 195 196 TF_DISALLOW_COPY_AND_ASSIGN(ImplementationSelector); 197 }; 198 199 } // namespace grappler 200 } // namespace tensorflow 201 202 #endif // TENSORFLOW_CORE_GRAPPLER_OPTIMIZERS_IMPLEMENTATION_SELECTOR_H_ 203