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.

model_v2_executor.h 3.0 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081
  1. /**
  2. * Copyright (c) Huawei Technologies Co., Ltd. 2022. All rights reserved.
  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 AIR_CXX_RUNTIME_V2_CORE_MODEL_V_2_EXECUTOR_H_
  17. #define AIR_CXX_RUNTIME_V2_CORE_MODEL_V_2_EXECUTOR_H_
  18. #include <memory>
  19. #include "graph/compute_graph.h"
  20. #include "graph/ge_error_codes.h"
  21. #include "model_desc.h"
  22. #include "runtime/stream.h"
  23. #include "exe_graph/runtime/tensor.h"
  24. #include "common/ge_visibility.h"
  25. #include "exe_graph_resource_guard.h"
  26. #include "exe_graph_executor.h"
  27. #include "subscriber/executor_subscribers_scheduler.h"
  28. namespace gert {
  29. enum class ExecutorState { kInit, kLoaded };
  30. enum SubExeGraphType { kInitExeGraph, kMainExeGraph, kDeInitExeGraph, kSubExeGraphTypeEnd };
  31. inline const char *GetSubExeGraphTypeStr(SubExeGraphType type) {
  32. constexpr const char *kSubExeGraphTypeStrs[kSubExeGraphTypeEnd] = {"Init", "Main", "DeInit"};
  33. return kSubExeGraphTypeStrs[type];
  34. }
  35. struct ModelExecuteArg {
  36. rtStream_t stream;
  37. };
  38. static_assert(std::is_standard_layout<ModelExecuteArg>::value, "The class ModelExecuteArg must be a POD");
  39. class VISIBILITY_EXPORT ModelV2Executor {
  40. public:
  41. static std::unique_ptr<ModelV2Executor> Create(const ge::ComputeGraphPtr &root_graph);
  42. ge::graphStatus Load();
  43. ge::graphStatus Execute(const ModelExecuteArg &arg, Tensor **inputs, size_t input_num, Tensor **outputs,
  44. size_t output_num);
  45. ge::graphStatus ExecuteSync(Tensor **inputs, size_t input_num, Tensor **outputs, size_t output_num);
  46. ge::graphStatus UnLoad();
  47. const ModelDesc &GetModelDesc() const;
  48. void SetModelDesc(ModelDesc *model_desc);
  49. ExeGraphExecutor &GetMainExeGraphExecutor() {
  50. return graphs_[kMainExeGraph];
  51. }
  52. ExecutorSubscribersScheduler &GetSubscribers();
  53. const ExecutorSubscribersScheduler &GetSubscribers() const;
  54. ModelV2Executor(const ModelV2Executor &) = delete;
  55. ModelV2Executor(ModelV2Executor &&) = delete;
  56. ModelV2Executor &operator=(const ModelV2Executor &) = delete;
  57. ModelV2Executor &operator=(ModelV2Executor &&) = delete;
  58. private:
  59. friend class ModelV2ExecutorBuilder;
  60. friend class ModelV2ExecutorTestHelper;
  61. ModelV2Executor();
  62. private:
  63. std::array<ExeGraphExecutor, kSubExeGraphTypeEnd> graphs_;
  64. ResourceGuard resource_guard_;
  65. ModelDesc *model_desc_ = nullptr;
  66. rtStream_t default_stream_ = nullptr;
  67. ExecutorSubscribersScheduler subscribers_;
  68. ExecutorState state_ = ExecutorState::kInit;
  69. };
  70. } // namespace gert
  71. #endif // AIR_CXX_RUNTIME_V2_CORE_MODEL_V_2_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