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_profiler.h 1.4 kB

4 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
  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_PROFILER_H_
  17. #define GE_HYBRID_EXECUTOR_HYBRID_PROFILER_H_
  18. #include <atomic>
  19. #include <chrono>
  20. #include <mutex>
  21. #include <ostream>
  22. #include <string>
  23. #include <vector>
  24. namespace ge {
  25. namespace hybrid {
  26. class HybridProfiler {
  27. public:
  28. enum EventType {
  29. GENERAL,
  30. SHAPE_INFERENCE,
  31. COMPILE,
  32. EXECUTION,
  33. CALLBACKS
  34. };
  35. struct Event {
  36. std::chrono::system_clock::time_point timestamp;
  37. EventType event_type;
  38. std::string desc;
  39. };
  40. HybridProfiler();
  41. ~HybridProfiler() = default;
  42. void RecordEvent(EventType event_type, const char *fmt, ...);
  43. void Reset();
  44. void Dump(std::ostream &os);
  45. private:
  46. std::vector<Event> events_;
  47. std::atomic_int counter_;
  48. };
  49. } // namespace hybrid
  50. } // namespace ge
  51. #endif // GE_HYBRID_EXECUTOR_HYBRID_PROFILER_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