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_context.h 2.5 kB

4 years ago
4 years ago
4 years ago
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970
  1. /**
  2. * Copyright 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_ITERATION_CONTEXT_H_
  17. #define GE_HYBRID_EXECUTOR_ITERATION_CONTEXT_H_
  18. #include <vector>
  19. #include "mmpa/mmpa_api.h"
  20. #include "hybrid/common/tensor_value.h"
  21. #include "hybrid/executor/hybrid_execution_context.h"
  22. #include "hybrid/executor/node_state.h"
  23. #include "hybrid/executor/node_done_manager.h"
  24. #include "hybrid/model/graph_item.h"
  25. #include "hybrid/model/node_item.h"
  26. namespace ge {
  27. namespace hybrid {
  28. class SubgraphContext {
  29. public:
  30. explicit SubgraphContext(const GraphItem *graph_item, GraphExecutionContext *execution_context);
  31. ~SubgraphContext();
  32. Status Init();
  33. void SetGroup(int group);
  34. void ResetContext(const NodePtr &node);
  35. void Reset();
  36. NodeStatePtr GetOrCreateNodeState(const NodeItem *node_item);
  37. void OnError(Status error);
  38. Status SetInput(const NodeItem &node_item, int input_index, const TensorValue &tensor);
  39. Status SetOutput(const NodeItem &node_item, int output_index, const TensorValue &tensor);
  40. Status SetInput(int index, const TensorValue &tensor);
  41. Status GetInput(int index, TensorValue &tensor);
  42. Status GetOutputs(std::vector<TensorValue> &outputs);
  43. Status Await(const NodePtr &node);
  44. void NodeDone(const NodePtr &node);
  45. private:
  46. NodeStatePtr CreateNodeState(const NodeItem *node_item);
  47. FrameStatePtr GetOrCreateFrameState(const NodeItem &node_item); // no lock
  48. friend class TaskContext;
  49. const GraphItem *graph_item_;
  50. GraphExecutionContext *execution_context_;
  51. mmRWLock_t rw_lock_;
  52. std::vector<TensorValue> all_inputs_;
  53. std::vector<TensorValue> all_outputs_;
  54. NodeDoneManager node_done_manager_;
  55. std::unordered_map<const NodeItem *, NodeStatePtr> node_states_;
  56. std::unordered_map<int64_t, FrameStatePtr> frame_states_;
  57. int group_ = -1;
  58. };
  59. } // namespace hybrid
  60. } // namespace ge
  61. #endif // GE_HYBRID_EXECUTOR_ITERATION_CONTEXT_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