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_execution_context.h 3.2 kB

4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283
  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_HYBRID_EXECUTION_CONTEXT_H_
  17. #define GE_HYBRID_EXECUTOR_HYBRID_EXECUTION_CONTEXT_H_
  18. #include <atomic>
  19. #include <unordered_map>
  20. #include "common/blocking_queue.h"
  21. #include "common/properties_manager.h"
  22. #include "framework/common/debug/ge_log.h"
  23. #include "hybrid/common/npu_memory_allocator.h"
  24. #include "hybrid/common/tensor_value.h"
  25. #include "hybrid/executor/hybrid_profiler.h"
  26. #include "hybrid/executor/node_done_manager.h"
  27. #include "hybrid/executor/node_state.h"
  28. #include "hybrid/executor/rt_callback_manager.h"
  29. #include "hybrid/model/hybrid_model.h"
  30. namespace ge {
  31. namespace hybrid {
  32. struct GraphExecutionContext {
  33. void SetErrorCode(Status error_code);
  34. Status GetStatus() const;
  35. uint64_t session_id = 0;
  36. const HybridModel *model = nullptr;
  37. rtStream_t stream = nullptr;
  38. rtContext_t rt_context = nullptr;
  39. rtContext_t rt_gen_context = nullptr;
  40. std::unique_ptr<CallbackManager> callback_manager;
  41. NpuMemoryAllocator *allocator = nullptr;
  42. mutable std::unique_ptr<HybridProfiler> profiler;
  43. DumpProperties dump_properties;
  44. bool trace_enabled = false;
  45. bool dump_enabled = false;
  46. long profiling_level = 0;
  47. long iteration = 0;
  48. Status status = SUCCESS;
  49. mutable std::mutex mu;
  50. };
  51. #define RECORD_PROFILING_EVENT(context, evt_type, fmt, category, node_name, ...) \
  52. do { \
  53. if ((context != nullptr) && (context)->profiler != nullptr) { \
  54. if (node_name != nullptr) { \
  55. context->profiler->RecordEvent(evt_type, "tid:%lu [%s] [%s] " fmt, GeLog::GetTid(), node_name, category, ##__VA_ARGS__);\
  56. } else { \
  57. context->profiler->RecordEvent(evt_type, "tid:%lu [%s] " fmt, GeLog::GetTid(), category, ##__VA_ARGS__); \
  58. }\
  59. } \
  60. } while (0)
  61. #define RECORD_MODEL_EXECUTION_EVENT(context, fmt, ...) \
  62. RECORD_PROFILING_EVENT((context), HybridProfiler::GENERAL, fmt, "ModelExecutor", nullptr, ##__VA_ARGS__)
  63. #define RECORD_SHAPE_INFERENCE_EVENT(context, name, fmt, ...) \
  64. RECORD_PROFILING_EVENT((context), HybridProfiler::SHAPE_INFERENCE, fmt, "ShapeInference", name, ##__VA_ARGS__)
  65. #define RECORD_COMPILE_EVENT(context, name, fmt, ...) \
  66. RECORD_PROFILING_EVENT((context), HybridProfiler::COMPILE, fmt, "Compilation", name, ##__VA_ARGS__)
  67. #define RECORD_EXECUTION_EVENT(context, name, fmt, ...) \
  68. RECORD_PROFILING_EVENT((context), HybridProfiler::EXECUTION, fmt, "Execution", name, ##__VA_ARGS__)
  69. #define RECORD_CALLBACK_EVENT(context, name, fmt, ...) \
  70. RECORD_PROFILING_EVENT((context), HybridProfiler::CALLBACK, fmt, "Callback", name, ##__VA_ARGS__)
  71. } // namespace hybrid
  72. } // namespace ge
  73. #endif // GE_HYBRID_EXECUTOR_HYBRID_EXECUTION_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