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.

next_iteration_pass.h 3.2 kB

5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
4 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109
  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_GRAPH_PASSES_NEXT_ITERATION_PASS_H_
  17. #define GE_GRAPH_PASSES_NEXT_ITERATION_PASS_H_
  18. #include "inc/graph_pass.h"
  19. struct LoopCondGroup {
  20. LoopCondGroup() : loop_cond(nullptr) {}
  21. ge::NodePtr loop_cond; // LoopCond node
  22. std::vector<ge::NodePtr> enter_nodes; // Enter nodes
  23. std::vector<std::pair<ge::NodePtr, ge::NodePtr>> merge_next_pairs; // <Merge, NextIteration>
  24. };
  25. using LoopCondGroupPtr = std::shared_ptr<LoopCondGroup>;
  26. namespace ge {
  27. class NextIterationPass : public GraphPass {
  28. public:
  29. Status Run(ComputeGraphPtr graph);
  30. ///
  31. /// @brief Clear Status, used for subgraph pass
  32. /// @return SUCCESS
  33. ///
  34. Status ClearStatus() override;
  35. private:
  36. ///
  37. /// @brief Group Enter node
  38. /// @param [in] enter_node
  39. /// @return Status
  40. ///
  41. Status GroupEnterNode(const NodePtr &enter_node);
  42. ///
  43. /// @brief Group Enter nodes without batch_label attr
  44. /// @param [in] compute_graph
  45. /// @return Status
  46. ///
  47. Status GroupWithNoBatch(const ComputeGraphPtr &graph);
  48. ///
  49. /// @brief Find while groups
  50. /// @return Status
  51. ///
  52. Status FindWhileGroups();
  53. ///
  54. /// @brief Verify if valid
  55. /// @return bool
  56. ///
  57. bool VerifyWhileGroup();
  58. ///
  59. /// @brief Handle while group
  60. /// @param [in] graph
  61. /// @return Status
  62. ///
  63. Status HandleWhileGroup(ComputeGraphPtr &graph);
  64. ///
  65. /// @brief Create Active Node
  66. /// @param [in] graph
  67. /// @param [in] name
  68. /// @return ge::NodePtr
  69. ///
  70. NodePtr CreateActiveNode(ComputeGraphPtr &graph, const std::string &name);
  71. ///
  72. /// @brief Break NextIteration Link & add name to merge attr
  73. /// @param [in] next_node
  74. /// @param [in] merge_node
  75. /// @return Status
  76. ///
  77. Status BreakNextIteration(const NodePtr &next_node, NodePtr &merge_node);
  78. ///
  79. /// @brief find target node
  80. /// @param [in] node
  81. /// @param [in] target_type
  82. /// @param [in] is_input
  83. /// @param [in] batch_label
  84. /// @param [out] target_node
  85. /// @return Status
  86. ///
  87. Status FindTargetNode(const NodePtr &node, const std::string &target_type, bool is_input,
  88. const std::string &batch_label, NodePtr &target_node);
  89. // map<frame_name, vector<enter_node>>
  90. std::unordered_map<std::string, std::vector<NodePtr>> frame_enter_map_;
  91. // map<frame_name, map<batch_label, LoopCondGroup>>
  92. std::unordered_map<std::string, std::unordered_map<std::string, LoopCondGroupPtr>> loop_group_map_;
  93. };
  94. } // namespace ge
  95. #endif // GE_GRAPH_PASSES_NEXT_ITERATION_PASS_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