1 /** 2 * Copyright 2021 Huawei Technologies Co., Ltd 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17 #ifndef MINDSPORE_CCSRC_RUNTIME_FRAMEWORK_ACTOR_OUTPUT_ACTOR_H_ 18 #define MINDSPORE_CCSRC_RUNTIME_FRAMEWORK_ACTOR_OUTPUT_ACTOR_H_ 19 20 #include <vector> 21 #include <string> 22 #include <memory> 23 #include <utility> 24 #include <algorithm> 25 #include <unordered_map> 26 #include "runtime/framework/control_node_parser.h" 27 #include "runtime/framework/device_tensor_store.h" 28 #include "runtime/framework/actor/actor_common.h" 29 #include "runtime/framework/actor/abstract_actor.h" 30 #include "runtime/hardware/device_context.h" 31 #include "backend/session/anf_runtime_algorithm.h" 32 #include "ir/tensor.h" 33 34 namespace mindspore { 35 namespace runtime { 36 using mindspore::device::DeviceContext; 37 using mindspore::session::KernelWithIndex; 38 using mindspore::tensor::TensorPtr; 39 40 // The output actor is used to receive the output result of actor which represents the graph output. 41 class OutputActor : public AbstractActor { 42 public: OutputActor(std::string name,size_t loop_count,size_t outputs_num,bool need_loop_count)43 OutputActor(std::string name, size_t loop_count, size_t outputs_num, bool need_loop_count) 44 : AbstractActor(name, KernelTransformType::kOutputActor, nullptr), 45 loop_count_(loop_count), 46 current_count_(0), 47 outputs_num_(outputs_num), 48 current_outputs_num_(0), 49 need_loop_count_(need_loop_count) { 50 outputs_.resize(outputs_num); 51 output_nodes_.resize(outputs_num); 52 device_contexts_.resize(outputs_num); 53 } 54 ~OutputActor() override = default; 55 56 void Init() override; 57 58 // The output actor collects loop count when receive the input control of loop count actor. 59 void CollectLoopCount(size_t loop_count, OpContext<DeviceTensor> *const context); 60 61 // The output actor collects output result when receive the data of actor. 62 void CollectOutput(const AnfNodePtr &output_node, size_t output_index, size_t output_position, 63 OpContext<DeviceTensor> *const context); 64 65 // The graph output need be set new device address every step or loop, to avoid that the device address 66 // context of tensor be rewritten in the next step or next loop. 67 void UpdateOutputDeviceAddress(); 68 outputs()69 std::vector<TensorPtr> &outputs() { return outputs_; } 70 71 private: 72 friend class GraphScheduler; 73 74 // The loop count is constant, the current count is increased after each step running finished. 75 // Collect the output result in the last loop which is represented by "loop_count_ - current_count_ == 1". 76 size_t loop_count_; 77 size_t current_count_; 78 79 // The outputs. 80 std::vector<TensorPtr> outputs_; 81 std::vector<KernelWithIndex> output_nodes_; 82 size_t outputs_num_; 83 size_t current_outputs_num_; 84 bool need_loop_count_; 85 }; 86 87 using OutputActorPtr = std::shared_ptr<OutputActor>; 88 } // namespace runtime 89 } // namespace mindspore 90 91 #endif // MINDSPORE_CCSRC_RUNTIME_FRAMEWORK_ACTOR_OUTPUT_ACTOR_H_ 92