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.

aicpu_ext_info.h 3.7 kB

3 years ago
4 years ago
4 years ago
3 years ago
4 years ago
3 years ago
4 years ago
4 years ago
3 years ago
4 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109
  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_AICPU_EXT_INFO_H_
  17. #define GE_HYBRID_AICPU_EXT_INFO_H_
  18. #include "external/ge/ge_api_error_codes.h"
  19. #include "cce/fwk_adpt_struct.h"
  20. #include "cce/aicpu_engine_struct.h"
  21. #include "graph/op_desc.h"
  22. #include "graph/ge_tensor.h"
  23. namespace ge {
  24. namespace hybrid {
  25. using AicpuShapeAndType = aicpu::FWKAdapter::ShapeAndType;
  26. using AicpuExtInfo = aicpu::FWKAdapter::ExtInfo;
  27. using AsyncWaitInfo = aicpu::FWKAdapter::AsyncWait;
  28. using AicpuSessionInfo = SessionInfo;
  29. class AicpuExtInfoHandler {
  30. public:
  31. AicpuExtInfoHandler(std::string node_name, uint32_t input_num, uint32_t output_num, UnknowShapeOpType unknown_type)
  32. : node_name_(std::move(node_name)),
  33. input_num_(input_num),
  34. output_num_(output_num),
  35. unknown_type_(unknown_type) {
  36. }
  37. ~AicpuExtInfoHandler() = default;
  38. uint8_t *GetExtInfo() const {
  39. return ext_info_.get();
  40. }
  41. size_t GetExtInfoLen() const {
  42. return ext_info_len_;
  43. }
  44. Status Parse(const std::string &ext_info);
  45. Status UpdateInputShapeAndType(uint32_t input_index, const GeTensorDesc &input_desc);
  46. Status UpdateOutputShapeAndType(uint32_t output_index, const GeTensorDesc &output_desc);
  47. Status UpdateSessionInfo(uint64_t session_id, uint64_t kernel_id, bool sess_flag);
  48. Status UpdateSessionInfoSessionId(uint64_t session_id);
  49. Status UpdateExecuteMode(bool flag);
  50. Status UpdateEventId(uint32_t event_id);
  51. Status GetOutputShapeAndType(uint32_t output_index, GeShape &shape, DataType &data_type);
  52. bool IsNeedRefreshIOAddr();
  53. int32_t GetTopicTypeFlag() const { return topic_type_flag_; }
  54. private:
  55. Status ParseExtShapeType(AicpuExtInfo *aicpu_ext_info);
  56. Status ParseExtInputShape(AicpuExtInfo *aicpu_ext_info);
  57. Status ParseExtOutputShape(AicpuExtInfo *aicpu_ext_info);
  58. Status ParseExtSessionInfo(AicpuExtInfo *aicpu_ext_info);
  59. Status ParseExtBitMap(AicpuExtInfo *aicpu_ext_info);
  60. Status ParseExtUpdateAddr(AicpuExtInfo *aicpu_ext_info);
  61. Status ParseExtTopicType(AicpuExtInfo *aicpu_ext_info);
  62. Status ParseExtAsyncWait(AicpuExtInfo *aicpu_ext_info);
  63. static Status UpdateShapeAndType(const GeShape &shape,
  64. DataType data_type,
  65. AicpuShapeAndType *shape_and_type);
  66. static void GetShapeAndType(const AicpuShapeAndType *shape_and_type,
  67. GeShape &shape,
  68. DataType &data_type);
  69. private:
  70. int32_t TopicTypeToRtsFlag(int32_t topic_type);
  71. const std::string node_name_;
  72. const uint32_t input_num_;
  73. const uint32_t output_num_;
  74. UnknowShapeOpType unknown_type_;
  75. AicpuSessionInfo *session_info_ = nullptr;
  76. AsyncWaitInfo *async_wait_ = nullptr;
  77. uint64_t *bit_map_ = nullptr;
  78. uint32_t *update_addr_ = nullptr;
  79. int32_t topic_type_flag_ = -1;
  80. std::unique_ptr<uint8_t[]> ext_info_;
  81. size_t ext_info_len_ = 0;
  82. std::vector<AicpuShapeAndType *> input_shape_and_type_;
  83. std::vector<AicpuShapeAndType *> output_shape_and_type_;
  84. };
  85. } // namespace hybrid
  86. } // namespace ge
  87. #endif // GE_HYBRID_AICPU_EXT_INFO_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