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.

single_op.h 3.3 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495
  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_SINGLE_OP_SINGLE_OP_H_
  17. #define GE_SINGLE_OP_SINGLE_OP_H_
  18. #include <cstdint>
  19. #include <memory>
  20. #include <mutex>
  21. #include <string>
  22. #include <vector>
  23. #include "common/ge_inner_error_codes.h"
  24. #include "framework/executor/ge_executor.h"
  25. #include "runtime/stream.h"
  26. #include "task/op_task.h"
  27. #include "cce/aicpu_engine_struct.h"
  28. namespace ge {
  29. class SingleOp {
  30. public:
  31. SingleOp(std::mutex *stream_mutex, rtStream_t stream);
  32. ~SingleOp();
  33. Status ExecuteAsync(const std::vector<DataBuffer> &inputs, const std::vector<DataBuffer> &outputs);
  34. void SetStream(rtStream_t stream);
  35. void SetSessionID(uint64_t session_id);
  36. private:
  37. Status ValidateArgs(const std::vector<DataBuffer> &inputs, const std::vector<DataBuffer> &outputs);
  38. Status UpdateArgs(const std::vector<DataBuffer> &inputs, const std::vector<DataBuffer> &outputs);
  39. Status GetArgs(const std::vector<DataBuffer> &inputs, const std::vector<DataBuffer> &outputs);
  40. friend class SingleOpModel;
  41. std::mutex *stream_mutex_;
  42. rtStream_t stream_ = nullptr;
  43. std::vector<void *> input_addr_list_;
  44. std::vector<size_t> input_sizes_;
  45. std::vector<void *> output_addr_list_;
  46. std::vector<size_t> output_sizes_;
  47. std::vector<uintptr_t> args_;
  48. uint64_t aicpu_session_id_ = 0;
  49. std::vector<OpTask *> tasks_;
  50. std::vector<std::vector<uintptr_t *>> arg_table_;
  51. };
  52. class DynamicSingleOp {
  53. public:
  54. DynamicSingleOp(uintptr_t resource_id, std::mutex *stream_mutex_, rtStream_t stream);
  55. ~DynamicSingleOp();
  56. Status ExecuteAsync(const vector<GeTensorDesc> &input_desc,
  57. const std::vector<DataBuffer> &inputs,
  58. std::vector<GeTensorDesc> &output_desc,
  59. std::vector<DataBuffer> &outputs);
  60. void SetSessionID(uint64_t session_id);
  61. private:
  62. friend class SingleOpModel;
  63. Status ValidateParams(const vector<GeTensorDesc> &input_desc,
  64. const std::vector<DataBuffer> &inputs,
  65. std::vector<GeTensorDesc> &output_desc,
  66. std::vector<DataBuffer> &outputs) const;
  67. Status AllocateWorkspaces(const std::vector<int64_t> &workspace_sizes,
  68. std::vector<void *> &workspaces);
  69. Status ExecuteTbeTask(const vector<GeTensorDesc> &input_desc,
  70. const vector<void *> &inputs,
  71. vector<GeTensorDesc> &output_desc,
  72. vector<void *> &outputs);
  73. std::unique_ptr<OpTask> op_task_;
  74. uintptr_t resource_id_ = 0;
  75. std::mutex *stream_mutex_;
  76. rtStream_t stream_ = nullptr;
  77. size_t num_inputs_ = 0;
  78. size_t num_outputs_ = 0;
  79. uint64_t aicpu_session_id_ = 0;
  80. };
  81. } // namespace ge
  82. #endif // GE_SINGLE_OP_SINGLE_OP_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