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 5.4 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
4 years ago
4 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116
  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 "graph/load/model_manager/davinci_model.h"
  25. #include "hybrid/common/npu_memory_allocator.h"
  26. #include "hybrid/common/tensor_value.h"
  27. #include "hybrid/executor/hybrid_profiler.h"
  28. #include "hybrid/executor/node_done_manager.h"
  29. #include "hybrid/executor/node_state.h"
  30. #include "hybrid/executor/rt_callback_manager.h"
  31. #include "hybrid/model/hybrid_model.h"
  32. // If expr is not SUCCESS, print the log and return the same value
  33. #define HYBRID_CHK_STATUS_RET(expr, ...) \
  34. do { \
  35. const ge::Status _status = (expr); \
  36. if (_status != ge::SUCCESS) { \
  37. if (_status == ge::END_OF_SEQUENCE) { \
  38. GELOGD("Got end of sequence"); \
  39. } else { \
  40. GELOGE(_status, __VA_ARGS__); \
  41. } \
  42. return _status; \
  43. } \
  44. } while (0)
  45. namespace ge {
  46. namespace hybrid {
  47. struct GraphExecutionContext {
  48. GraphExecutionContext();
  49. ~GraphExecutionContext() = default;
  50. void SetErrorCode(Status error_code);
  51. Status GetStatus() const;
  52. Status Synchronize(rtStream_t rt_stream);
  53. Status DumpExceptionInfo(const std::vector<rtExceptionInfo> &exception_infos);
  54. uint64_t session_id = 0;
  55. uint64_t context_id = 0;
  56. const HybridModel *model = nullptr;
  57. const GEThreadLocalContext *ge_context = nullptr;
  58. rtStream_t stream = nullptr;
  59. rtStream_t hccl_stream = nullptr;
  60. rtContext_t rt_context = nullptr;
  61. rtContext_t rt_gen_context = nullptr;
  62. std::unique_ptr<CallbackManager> callback_manager = nullptr;
  63. NpuMemoryAllocator *allocator = nullptr;
  64. mutable std::unique_ptr<HybridProfiler> profiler = nullptr;
  65. DumpProperties dump_properties;
  66. bool trace_enabled = false;
  67. bool dump_enabled = false;
  68. ExceptionDumper exception_dumper;
  69. std::vector<std::shared_ptr<ge::DavinciModel>> davinci_model;
  70. std::atomic_bool is_eos_{false};
  71. static long profiling_level;
  72. long iteration = 0;
  73. void *global_step = nullptr;
  74. private:
  75. Status status = SUCCESS;
  76. mutable std::mutex mu;
  77. };
  78. #define RECORD_PROFILING_EVENT(context, evt_type, fmt, category, node_name, ...) \
  79. do { \
  80. if (ge::hybrid::GraphExecutionContext::profiling_level > 0) { \
  81. if ((context != nullptr) && (context)->profiler != nullptr) { \
  82. if (node_name != nullptr) { \
  83. context->profiler->RecordEvent(evt_type, "tid:%lu [%s@%ld] [%s] " fmt, \
  84. GeLog::GetTid(), node_name, context->iteration, category, ##__VA_ARGS__); \
  85. } else { \
  86. context->profiler->RecordEvent(evt_type, "tid:%lu [%s] " fmt, GeLog::GetTid(), category, ##__VA_ARGS__); \
  87. } \
  88. } \
  89. } \
  90. } while (0)
  91. #define RECORD_MODEL_EXECUTION_EVENT(context, fmt, ...) \
  92. RECORD_PROFILING_EVENT((context), HybridProfiler::GENERAL, fmt, "ModelExecutor", nullptr, ##__VA_ARGS__)
  93. #define RECORD_SHAPE_INFERENCE_EVENT(context, name, fmt, ...) \
  94. RECORD_PROFILING_EVENT((context), HybridProfiler::SHAPE_INFERENCE, fmt, "ShapeInference", name, ##__VA_ARGS__)
  95. #define RECORD_COMPILE_EVENT(context, name, fmt, ...) \
  96. RECORD_PROFILING_EVENT((context), HybridProfiler::COMPILE, fmt, "Compilation", name, ##__VA_ARGS__)
  97. #define RECORD_EXECUTION_EVENT(context, name, fmt, ...) \
  98. RECORD_PROFILING_EVENT((context), HybridProfiler::EXECUTION, fmt, "Execution", name, ##__VA_ARGS__)
  99. #define RECORD_CALLBACK_EVENT(context, name, fmt, ...) \
  100. RECORD_PROFILING_EVENT((context), HybridProfiler::CALLBACKS, fmt, "Callback", name, ##__VA_ARGS__)
  101. } // namespace hybrid
  102. } // namespace ge
  103. #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