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.

hybrid_model_async_executor.h 2.7 kB

4 years ago
4 years ago
4 years ago
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788
  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_MODEL_HYBRID_MODEL_ASYNC_EXECUTOR_H_
  17. #define GE_HYBRID_EXECUTOR_MODEL_HYBRID_MODEL_ASYNC_EXECUTOR_H_
  18. #include <atomic>
  19. #include <mutex>
  20. #include <future>
  21. #include "external/ge/ge_api_error_codes.h"
  22. #include "external/ge/ge_api_types.h"
  23. #include "graph/load/new_model_manager/data_inputer.h"
  24. #include "hybrid/executor/hybrid_model_executor.h"
  25. #include "runtime/stream.h"
  26. namespace ge {
  27. namespace hybrid {
  28. class HybridModel;
  29. class HybridModelAsyncExecutor {
  30. public:
  31. explicit HybridModelAsyncExecutor(HybridModel *model);
  32. ~HybridModelAsyncExecutor();
  33. Status Init();
  34. Status Execute(const vector<GeTensor> &inputs, vector<GeTensor> &outputs);
  35. Status Start(const std::shared_ptr<ModelListener> &listener);
  36. void SetDeviceId(uint32_t device_id);
  37. void SetModelId(uint32_t model_id);
  38. Status Stop();
  39. Status EnqueueData(const std::shared_ptr<InputDataWrapper> &data);
  40. private:
  41. Status InitInputTensors();
  42. Status RunInternal();
  43. Status SyncVarData();
  44. Status HandleResult(Status exec_ret,
  45. uint32_t data_id,
  46. HybridModelExecutor::ExecuteArgs &args,
  47. OutputData *output_data);
  48. Status CopyOutputs(HybridModelExecutor::ExecuteArgs &args,
  49. OutputData *output_data,
  50. std::vector<ge::OutputTensorInfo> &outputs);
  51. Status OnComputeDone(uint32_t data_index, uint32_t result_code, std::vector<ge::OutputTensorInfo> &outputs);
  52. Status PreRun(InputData &current_data);
  53. Status CopyInputData(const InputData &current_data);
  54. std::mutex mu_;
  55. HybridModel *model_;
  56. uint32_t device_id_ = 0U;
  57. uint32_t model_id_ = 0U;
  58. std::atomic_bool run_flag_;
  59. std::unique_ptr<DataInputer> data_inputer_;
  60. std::unique_ptr<HybridModelExecutor> executor_;
  61. std::future<Status> future_;
  62. uint64_t iterator_count_ = 0;
  63. rtStream_t stream_ = nullptr;
  64. std::map<uint32_t, TensorValue> input_tensors_;
  65. std::shared_ptr<ModelListener> listener_;
  66. };
  67. } // namespace hybrid
  68. } // namespace ge
  69. #endif // GE_HYBRID_EXECUTOR_MODEL_HYBRID_MODEL_ASYNC_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