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.

ge_profiling.h 1.5 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445
  1. /**
  2. * Copyright 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 INC_FRAMEWORK_COMMON_GE_PROFILING_H_
  17. #define INC_FRAMEWORK_COMMON_GE_PROFILING_H_
  18. #include "ge/ge_api_error_codes.h"
  19. #include "toolchain/prof_callback.h"
  20. const int MAX_DEV_NUM = 64;
  21. enum ProfCommandHandleType {
  22. kProfCommandhandleInit = 0,
  23. kProfCommandhandleStart,
  24. kProfCommandhandleStop,
  25. kProfCommandhandleFinalize,
  26. kProfCommandhandleModelSubscribe,
  27. kProfCommandhandleModelUnsubscribe
  28. };
  29. struct ProfCommandHandleData {
  30. uint64_t profSwitch;
  31. uint32_t devNums; // length of device id list
  32. uint32_t devIdList[MAX_DEV_NUM];
  33. uint32_t modelId;
  34. };
  35. ge::Status RegProfCtrlCallback(MsprofCtrlCallback func);
  36. ge::Status RegProfSetDeviceCallback(MsprofSetDeviceCallback func);
  37. ge::Status RegProfReporterCallback(MsprofReporterCallback func);
  38. ge::Status ProfCommandHandle(ProfCommandHandleType type, void *data, uint32_t len);
  39. #endif // INC_FRAMEWORK_COMMON_GE_PROFILING_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