You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

subgraph_executor.h 4.8 kB

4 years ago
4 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120
  1. /**
  2. * Copyright 2019-2020 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. #ifndef GE_HYBRID_EXECUTOR_EXECUTOR_SUBGRAPH_EXECUTOR_H_
  17. #define GE_HYBRID_EXECUTOR_EXECUTOR_SUBGRAPH_EXECUTOR_H_
  18. #include <vector>
  19. #include "common/blocking_queue.h"
  20. #include "common/thread_pool.h"
  21. #include "hybrid/executor/subgraph_context.h"
  22. #include "hybrid/executor/node_state.h"
  23. #include "hybrid/executor/hybrid_execution_context.h"
  24. #include "hybrid/executor/worker/shape_inference_engine.h"
  25. #include "hybrid/model/graph_item.h"
  26. #include "hybrid/node_executor/task_context.h"
  27. namespace ge {
  28. namespace hybrid {
  29. // Executor for executing a subgraph
  30. class SubgraphExecutor {
  31. public:
  32. SubgraphExecutor(const GraphItem *graph_item, GraphExecutionContext *context, bool force_infer_shape = false);
  33. ~SubgraphExecutor();
  34. Status InitForPartialExecution(const std::vector<TensorValue> &inputs,
  35. const std::vector<ConstGeTensorDescPtr> &input_desc);
  36. Status PartialExecuteAsync(int task_group);
  37. /**
  38. * Execute subgraph async, output tensor address(not data) and output tensor descriptions are
  39. * valid after this method returned
  40. * @param inputs input tensors
  41. * @param input_desc input tensor descriptions
  42. * @return SUCCESS on success, error code otherwise
  43. */
  44. Status ExecuteAsync(const std::vector<TensorValue> &inputs,
  45. const std::vector<ConstGeTensorDescPtr> &input_desc);
  46. /**
  47. * Execute subgraph async, output tensor address(not data) and output tensor descriptions are
  48. * valid after this method returned
  49. * @param inputs input tensors
  50. * @param input_desc input tensor descriptions
  51. * @return SUCCESS on success, error code otherwise
  52. */
  53. Status ExecuteAsync(const std::vector<TensorValue> &inputs,
  54. const std::vector<ConstGeTensorDescPtr> &input_desc,
  55. const std::vector<TensorValue> &outputs);
  56. /**
  57. * Execute subgraph async, output tensor address(not data) and output tensor descriptions are
  58. * valid after this method returned
  59. * @param task_context instance of TaskContext
  60. * @return SUCCESS on success, error code otherwise
  61. */
  62. Status ExecuteAsync(TaskContext &task_context);
  63. /**
  64. * Synchronize all tasks in the subgraph. output tensor data are valid after this method returned
  65. * @return SUCCESS on success, error code otherwise
  66. */
  67. Status Synchronize();
  68. /**
  69. * Get output tensors
  70. * @param outputs output tensors
  71. * @return SUCCESS on success, error code otherwise
  72. */
  73. Status GetOutputs(std::vector<TensorValue> &outputs);
  74. /**
  75. * Get output tensors and output tensor descriptions
  76. * @param outputs output tensors
  77. * @param output_desc output tensor descriptions
  78. * @return SUCCESS on success, error code otherwise
  79. */
  80. Status GetOutputs(std::vector<TensorValue> &outputs, std::vector<ConstGeTensorDescPtr> &output_desc);
  81. private:
  82. Status PrepareForExecution(GraphExecutionContext *ctx, NodeState &node_state);
  83. Status EnableOutputZeroCopy(const std::vector<TensorValue> &outputs);
  84. Status InferShape(ShapeInferenceEngine *shape_inference_engine, NodeState &node_state) const;
  85. Status Init(const std::vector<TensorValue> &inputs,
  86. const std::vector<ConstGeTensorDescPtr> &input_desc);
  87. Status InitInputsForUnknownShape(const std::vector<TensorValue> &inputs,
  88. const std::vector<ConstGeTensorDescPtr> &input_desc);
  89. Status InitInputsForKnownShape(const std::vector<TensorValue> &inputs);
  90. Status ExecuteAsyncForKnownShape(const std::vector<TensorValue> &inputs);
  91. Status ScheduleTasks(int group = -1);
  92. Status PrepareNodes(int group = -1);
  93. Status LaunchTasks();
  94. Status SetOutputsToParentNode(TaskContext &task_context);
  95. const GraphItem *graph_item_;
  96. GraphExecutionContext *context_;
  97. std::unique_ptr<SubgraphContext> subgraph_context_;
  98. bool force_infer_shape_;
  99. ThreadPool pre_run_pool_;
  100. BlockingQueue<NodeState *> ready_queue_;
  101. std::unique_ptr<ShapeInferenceEngine> shape_inference_engine_;
  102. std::shared_ptr<TaskContext> known_shape_task_context_;
  103. };
  104. } // namespace hybrid
  105. } // namespace ge
  106. #endif // GE_HYBRID_EXECUTOR_EXECUTOR_SUBGRAPH_EXECUTOR_H_

图引擎模块(GE)是MindSpore的一个子模块,其代码由C++实现,位于前端模块ME和底层硬件之间,起到承接作用。图引擎模块以ME下发的图作为输入,然后进行一系列的深度图优化操作,最后输出一张可以在底层硬件上高效运行的图。GE针对昇腾AI处理器的硬件结构特点,做了特定的优化工作,以此来充分发挥出昇腾AI处理器的强大算力。在进行模型训练/推理时,GE会被自动调用而用户并不感知。GE主要由GE API和GE Core两部分组成,详细的架构图如下所示