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_prof.h 3.2 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102
  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 INC_EXTERNAL_GE_GE_PROF_H_
  17. #define INC_EXTERNAL_GE_GE_PROF_H_
  18. #include <map>
  19. #include <string>
  20. #include <vector>
  21. #include "ge/ge_api_error_codes.h"
  22. namespace ge {
  23. enum ProfDataTypeConfig {
  24. kProfTaskTime = 0x0002,
  25. kProfAiCoreMetrics = 0x0004,
  26. kProfAicpuTrace = 0x0008,
  27. kProfTrainingTrace = 0x0800,
  28. kProfHcclTrace = 0x1000
  29. };
  30. enum ProfilingAicoreMetrics {
  31. kAicoreArithmaticThroughput = 0,
  32. kAicorePipeline = 1,
  33. kAicoreSynchronization = 2,
  34. kAicoreMemory = 3,
  35. kAicoreInternalMemory = 4,
  36. kAicoreStall = 5
  37. };
  38. typedef struct ProfAicoreEvents ProfAicoreEvents;
  39. typedef struct aclgrphProfConfig aclgrphProfConfig;
  40. ///
  41. /// @ingroup AscendCL
  42. /// @brief Initialize the profiling and set profiling configuration path
  43. /// @param [in] profiler_path: configuration path of profiling
  44. /// @param [in] length: length of configuration path
  45. /// @return Status result of function
  46. ///
  47. Status aclgrphProfInit(const char *profiler_path, uint32_t length);
  48. ///
  49. /// @ingroup AscendCL
  50. /// @brief Finalize profiling
  51. /// @return Status result of function
  52. ///
  53. Status aclgrphProfFinalize();
  54. ///
  55. /// @ingroup AscendCL
  56. /// @brief Create data of type aclgrphProfConfig
  57. /// @param [in] deviceid_list: device id list
  58. /// @param [in] device_nums: device numbers
  59. /// @param [in] aicore_metrics: type of aicore metrics
  60. /// @param [in] aicore_events: pointer to aicore events be reserved, only support NULL now
  61. /// @param [in] data_type_config: modules need profiling
  62. /// @return Status result of function
  63. ///
  64. aclgrphProfConfig *aclgrphProfCreateConfig(uint32_t *deviceid_list, uint32_t device_nums,
  65. ProfilingAicoreMetrics aicore_metrics, ProfAicoreEvents *aicore_events,
  66. uint64_t data_type_config);
  67. ///
  68. /// @ingroup AscendCL
  69. /// @brief Destroy data of type aclgrphProfConfig
  70. /// @param [in] profiler_config: config of profiling
  71. /// @return Status result of function
  72. ///
  73. Status aclgrphProfDestroyConfig(aclgrphProfConfig *profiler_config);
  74. ///
  75. /// @ingroup AscendCL
  76. /// @brief Start profiling of modules which is configured by profiler config
  77. /// @param [in] profiler_config: config of profiling
  78. /// @return Status result of function
  79. ///
  80. Status aclgrphProfStart(aclgrphProfConfig *profiler_config);
  81. ///
  82. /// @ingroup AscendCL
  83. /// @brief Stop profiling of modules which is configured by profiler config
  84. /// @param [in] profiler_config: config of profiling
  85. /// @return Status result of function
  86. ///
  87. Status aclgrphProfStop(aclgrphProfConfig *profiler_config);
  88. } // namespace ge
  89. #endif // INC_EXTERNAL_GE_GE_PROF_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