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 4.3 kB

4 years ago
4 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110
  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 "graph/ge_local_context.h"
  24. #include "hybrid/common/npu_memory_allocator.h"
  25. #include "hybrid/common/tensor_value.h"
  26. #include "hybrid/executor/hybrid_profiler.h"
  27. #include "hybrid/executor/node_done_manager.h"
  28. #include "hybrid/executor/node_state.h"
  29. #include "hybrid/executor/rt_callback_manager.h"
  30. #include "hybrid/model/hybrid_model.h"
  31. // If expr is not SUCCESS, print the log and return the same value
  32. #define HYBRID_CHK_STATUS_RET(expr, ...) \
  33. do { \
  34. const ge::Status _status = (expr); \
  35. if (_status != ge::SUCCESS) { \
  36. if (_status == ge::END_OF_SEQUENCE) { \
  37. GELOGD("Got end of sequence"); \
  38. } else { \
  39. GELOGE(_status, __VA_ARGS__); \
  40. } \
  41. return _status; \
  42. } \
  43. } while (0)
  44. namespace ge {
  45. namespace hybrid {
  46. struct GraphExecutionContext {
  47. GraphExecutionContext();
  48. ~GraphExecutionContext() = default;
  49. void SetErrorCode(Status error_code);
  50. Status GetStatus() const;
  51. Status Synchronize(rtStream_t rt_stream);
  52. uint64_t session_id = 0;
  53. uint64_t context_id = 0;
  54. const HybridModel *model = nullptr;
  55. const GEThreadLocalContext *ge_context = nullptr;
  56. rtStream_t stream = nullptr;
  57. rtContext_t rt_context = nullptr;
  58. rtContext_t rt_gen_context = nullptr;
  59. std::unique_ptr<CallbackManager> callback_manager = nullptr;
  60. NpuMemoryAllocator *allocator = nullptr;
  61. mutable std::unique_ptr<HybridProfiler> profiler = nullptr;
  62. DumpProperties dump_properties;
  63. bool trace_enabled = false;
  64. bool dump_enabled = false;
  65. std::atomic_bool is_eos_;
  66. long profiling_level = 0;
  67. long iteration = 0;
  68. void *global_step = nullptr;
  69. private:
  70. Status status = SUCCESS;
  71. mutable std::mutex mu;
  72. };
  73. #define RECORD_PROFILING_EVENT(context, evt_type, fmt, category, node_name, ...) \
  74. do { \
  75. if ((context != nullptr) && (context)->profiler != nullptr) { \
  76. if (node_name != nullptr) { \
  77. context->profiler->RecordEvent(evt_type, "tid:%lu [%s@%ld] [%s] " fmt, \
  78. GeLog::GetTid(), node_name, context->iteration, category, \
  79. ##__VA_ARGS__); \
  80. } else { \
  81. context->profiler->RecordEvent(evt_type, "tid:%lu [%s] " fmt, GeLog::GetTid(), category, ##__VA_ARGS__); \
  82. }\
  83. } \
  84. } while (0)
  85. #define RECORD_MODEL_EXECUTION_EVENT(context, fmt, ...) \
  86. RECORD_PROFILING_EVENT((context), HybridProfiler::GENERAL, fmt, "ModelExecutor", nullptr, ##__VA_ARGS__)
  87. #define RECORD_SHAPE_INFERENCE_EVENT(context, name, fmt, ...) \
  88. RECORD_PROFILING_EVENT((context), HybridProfiler::SHAPE_INFERENCE, fmt, "ShapeInference", name, ##__VA_ARGS__)
  89. #define RECORD_COMPILE_EVENT(context, name, fmt, ...) \
  90. RECORD_PROFILING_EVENT((context), HybridProfiler::COMPILE, fmt, "Compilation", name, ##__VA_ARGS__)
  91. #define RECORD_EXECUTION_EVENT(context, name, fmt, ...) \
  92. RECORD_PROFILING_EVENT((context), HybridProfiler::EXECUTION, fmt, "Execution", name, ##__VA_ARGS__)
  93. #define RECORD_CALLBACK_EVENT(context, name, fmt, ...) \
  94. RECORD_PROFILING_EVENT((context), HybridProfiler::CALLBACKS, fmt, "Callback", name, ##__VA_ARGS__)
  95. } // namespace hybrid
  96. } // namespace ge
  97. #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