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.

node_item.h 3.6 kB

4 years ago
4 years ago
4 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130
  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_MODEL_NODE_ITEM_H_
  17. #define GE_HYBRID_MODEL_NODE_ITEM_H_
  18. #include <mutex>
  19. #include <vector>
  20. #include "external/ge/ge_api_error_codes.h"
  21. #include "graph/node.h"
  22. #include "graph/op_desc.h"
  23. #include "framework/common/types.h"
  24. #include "hybrid/common/tensor_value.h"
  25. namespace ge {
  26. namespace hybrid {
  27. class NodeTask;
  28. class NodeExecutor;
  29. struct FusedSubgraph {
  30. std::map<int, std::vector<GeTensorDescPtr>> input_mapping;
  31. std::map<int, OpDescPtr> output_mapping;
  32. std::vector<NodePtr> nodes;
  33. ComputeGraphPtr graph;
  34. };
  35. bool IsControlOp(const std::string &op_type);
  36. // for caching static information across execution
  37. struct NodeItem {
  38. ~NodeItem() = default;
  39. static Status Create(const NodePtr &node, std::unique_ptr<NodeItem> &node_item);
  40. const std::string &NodeName() const {
  41. return node_name;
  42. }
  43. const std::string &NodeType() const {
  44. return node_type;
  45. }
  46. OpDescPtr GetOpDesc() const {
  47. return node->GetOpDesc();
  48. }
  49. bool IsInputShapeStatic(int index) const;
  50. GeTensorDescPtr MutableOutputDesc(int index) const;
  51. Status UpdateInputDesc(int index, const GeTensorDesc &tensor_desc);
  52. GeTensorDescPtr MutableInputDesc(int index) const;
  53. Status GetInputDesc(int index, GeTensorDesc &tensor_desc) const;
  54. Status GetOutputDesc(int index, GeTensorDesc &tensor_desc) const;
  55. Status GetCanonicalInputIndex(uint32_t index, int &canonical_index) const;
  56. bool IsControlOp() const;
  57. bool IsHcclOp() const;
  58. void SetToDynamic();
  59. std::string DebugString() const;
  60. NodePtr node;
  61. OpDesc *op_desc;
  62. int node_id = -1;
  63. int group = -1;
  64. int num_inputs = 0;
  65. int num_outputs = 0;
  66. int input_start = -1;
  67. int output_start = -1;
  68. bool is_dynamic = false;
  69. bool has_observer = false;
  70. bool has_optional_inputs = false;
  71. bool is_output_shape_static = true;
  72. bool is_need_force_infershape = false;
  73. UnknowShapeOpType shape_inference_type = DEPEND_IN_SHAPE;
  74. std::string node_name;
  75. std::string node_type;
  76. std::vector<ge::NodePtr> dependents_for_shape_inference;
  77. std::vector<ge::NodePtr> dependents_for_execution;
  78. std::set<int> to_const_output_id_list;
  79. // src_output_id, dst_anchor_id, dst_node
  80. vector<vector<pair<int, NodeItem *>>> outputs;
  81. std::shared_ptr<NodeTask> kernel_task;
  82. std::unique_ptr<FusedSubgraph> fused_subgraph;
  83. const NodeExecutor *node_executor = nullptr;
  84. std::map<int, ge::NodePtr> ref_outputs;
  85. std::map<int, int> reuse_inputs;
  86. std::map<int, int> reuse_outputs;
  87. int num_static_input_shapes = 0;
  88. bool is_profiling_report = false;
  89. private:
  90. explicit NodeItem(NodePtr node);
  91. Status Init();
  92. Status InitInputsAndOutputs();
  93. void ResolveOptionalInputs();
  94. Status ResolveDynamicState();
  95. Status ResolveStaticInputsAndOutputs();
  96. void ResolveUnknownShapeType();
  97. GeTensorDescPtr DoGetInputDesc(int index) const;
  98. std::vector<bool> is_input_shape_static_;
  99. std::vector<uint32_t> input_desc_indices_;
  100. mutable std::mutex mu_;
  101. };
  102. } // namespace hybrid
  103. } // namespace ge
  104. #endif // GE_HYBRID_MODEL_NODE_ITEM_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