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 3.1 kB

4 years ago
4 years ago
4 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113
  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 <vector>
  19. #include "external/ge/ge_api_error_codes.h"
  20. #include "graph/node.h"
  21. #include "graph/op_desc.h"
  22. #include "framework/common/types.h"
  23. #include "hybrid/common/tensor_value.h"
  24. namespace ge {
  25. namespace hybrid {
  26. class NodeTask;
  27. class NodeExecutor;
  28. struct FusedSubgraph {
  29. std::map<uint32_t, std::vector<GeTensorDescPtr>> input_mapping;
  30. std::map<uint32_t, OpDescPtr> output_mapping;
  31. std::vector<NodePtr> nodes;
  32. ComputeGraphPtr graph;
  33. };
  34. bool IsControlOp(const std::string &op_type);
  35. // for caching static information across execution
  36. struct NodeItem {
  37. ~NodeItem() = default;
  38. static Status Create(const NodePtr &node, std::unique_ptr<NodeItem> &node_item);
  39. const std::string &NodeName() const {
  40. return node_name;
  41. }
  42. const std::string &NodeType() const {
  43. return node_type;
  44. }
  45. OpDescPtr GetOpDesc() const {
  46. return node->GetOpDesc();
  47. }
  48. bool IsInputShapeStatic(int index) const;
  49. GeTensorDescPtr MutableOutputDesc(int index) const {
  50. return op_desc->MutableOutputDesc(static_cast<uint32_t>(index));
  51. }
  52. GeTensorDescPtr MutableInputDesc(int index) const;
  53. Status GetCanonicalInputIndex(uint32_t index, int &canonical_index) const;
  54. bool IsControlOp() const;
  55. void SetToDynamic();
  56. std::string DebugString() const;
  57. NodePtr node;
  58. OpDesc *op_desc;
  59. int node_id = -1;
  60. int num_inputs = 0;
  61. int num_outputs = 0;
  62. int input_start = -1;
  63. int output_start = -1;
  64. bool is_dynamic = false;
  65. bool has_observer = false;
  66. bool has_optional_inputs = false;
  67. bool is_output_shape_static = true;
  68. UnknowShapeOpType shape_inference_type = DEPEND_IN_SHAPE;
  69. std::string node_name;
  70. std::string node_type;
  71. std::vector<ge::NodePtr> dependents_for_shape_inference;
  72. std::vector<ge::NodePtr> dependents_for_execution;
  73. std::set<int> to_const_output_id_list;
  74. // src_output_id, dst_anchor_id, dst_node
  75. vector<vector<pair<int, NodeItem *>>> outputs;
  76. std::shared_ptr<NodeTask> kernel_task;
  77. std::unique_ptr<FusedSubgraph> fused_subgraph;
  78. const NodeExecutor *node_executor = nullptr;
  79. std::map<int, ge::NodePtr> ref_outputs;
  80. std::map<int, int> reuse_inputs;
  81. std::map<int, int> reuse_outputs;
  82. int num_static_input_shapes = 0;
  83. private:
  84. explicit NodeItem(NodePtr node);
  85. Status Init();
  86. std::vector<bool> is_input_shape_static_;
  87. std::vector<uint32_t> input_desc_indices_;
  88. };
  89. } // namespace hybrid
  90. } // namespace ge
  91. #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