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_model.h 4.6 kB

4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138
  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 GE_HYBRID_HYBRID_GRAPH_H_
  17. #define GE_HYBRID_HYBRID_GRAPH_H_
  18. #include <vector>
  19. #include <queue>
  20. #include <memory>
  21. #include "framework/common/ge_inner_error_codes.h"
  22. #include "graph/load/new_model_manager/data_inputer.h"
  23. #include "graph/load/new_model_manager/task_info/task_info.h"
  24. #include "graph/node.h"
  25. #include "hybrid/common/tensor_value.h"
  26. #include "hybrid/model/node_item.h"
  27. #include "hybrid/model/graph_item.h"
  28. #include "model/ge_root_model.h"
  29. namespace ge {
  30. namespace hybrid {
  31. class HybridModel {
  32. public:
  33. explicit HybridModel(GeRootModelPtr ge_model);
  34. ~HybridModel();
  35. Status Init();
  36. const NodeItem *GetNodeItem(const NodePtr &node) const;
  37. uint64_t GetSessionId() const {
  38. return root_runtime_param_.session_id;
  39. }
  40. GeModelPtr GetGeModel(const NodePtr &node) const;
  41. NodeItem *MutableNodeItem(const NodePtr &node);
  42. size_t TotalVarMemSize() const {
  43. return root_runtime_param_.var_size;
  44. }
  45. const uint8_t* GetVarMemBase() const {
  46. return var_mem_base_;
  47. }
  48. void SetDeviceId(uint32_t device_id) {
  49. device_id_ = device_id;
  50. }
  51. void SetModelId(uint32_t model_id) {
  52. model_id_ = model_id;
  53. }
  54. uint32_t GetModelId() const {
  55. return model_id_;
  56. }
  57. TensorValue* GetVariable(const string &name) const;
  58. NodePtr GetVariableNode(const string &name) const;
  59. const std::vector<domi::TaskDef>* GetTaskDefs(const NodePtr &node) const;
  60. const GraphItem *GetRootGraphItem() const;
  61. const GraphItem *GetSubgraphItem(const std::string &graph_name) const;
  62. const GraphItem *GetSubgraphItem(const ComputeGraphPtr &subgraph) const;
  63. const string &GetModelName() const;
  64. Status GetDynamicBatchInfo(std::vector<std::vector<int64_t>> &batch_info, int32_t &dynamic_type);
  65. void GetUserDesignateShapeOrder(std::vector<std::string> &user_input_shape_order);
  66. void GetModelAttr(std::vector<std::string> &dynamic_output_shape_info);
  67. Status GetInputOutputDescInfo(vector<InputOutputDescInfo> &input_desc,
  68. vector<InputOutputDescInfo> &output_desc,
  69. std::vector<uint32_t> &input_formats,
  70. std::vector<uint32_t> &outputFormats);
  71. Status GetInputDescInfo(vector<InputOutputDescInfo> &input_desc, std::vector<uint32_t> &formats);
  72. void CreateOutput(ConstGeTensorDescPtr &output_desc, InputOutputDescInfo &output, uint32_t &format_result);
  73. Status GetOutputDescInfo(vector<InputOutputDescInfo> &output_desc, std::vector<uint32_t> &formats);
  74. void CreateInputDimsInfo(const OpDescPtr &op_desc, Format format, InputOutputDescInfo &input);
  75. void SetModelDescVersion(bool is_new_model_desc) { is_new_model_desc_ = is_new_model_desc; }
  76. void SetInputDimsAndShapeRangesInfo(const vector<int64_t> &model_input_dims, std::vector<std::pair<int64_t, int64_t>> &shape_ranges,
  77. Format &format, InputOutputDescInfo &input);
  78. private:
  79. friend class HybridModelBuilder;
  80. friend class HybridModelAsyncExecutor;
  81. std::string model_name_;
  82. GeRootModelPtr ge_root_model_;
  83. std::map<uint32_t, NodeItem *> input_nodes_;
  84. std::map<std::string, NodePtr> constant_op_nodes_;
  85. std::map<std::string, NodePtr> device_variable_nodes_; //lint !e148
  86. std::map<std::string, NodePtr> host_variable_nodes_; //lint !e148
  87. std::map<std::string, std::unique_ptr<TensorValue>> variable_tensors_;
  88. std::map<NodePtr, std::vector<domi::TaskDef>> task_defs_;
  89. std::map<NodePtr, GeModelPtr> known_shape_sub_models_;
  90. std::unique_ptr<GraphItem> root_graph_item_;
  91. std::map<std::string, std::unique_ptr<GraphItem>> subgraph_items_;
  92. std::map<NodePtr, std::unique_ptr<NodeItem>> node_items_;
  93. bool is_new_model_desc_ = false; // support aipp
  94. // runtime fields
  95. uint32_t device_id_ = 0;
  96. uint32_t model_id_ = 0;
  97. uint8_t *var_mem_base_ = nullptr;
  98. RuntimeParam root_runtime_param_;
  99. };
  100. } // namespace hybrid
  101. } // namespace ge
  102. #endif // GE_HYBRID_HYBRID_GRAPH_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