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.

atomic_addr_clean_pass.h 3.2 kB

5 years ago
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
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106
  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_ATOMIC_ADDR_CLEAN_PASS_H_
  17. #define GE_GRAPH_PASSES_ATOMIC_ADDR_CLEAN_PASS_H_
  18. #include <vector>
  19. #include "graph/graph.h"
  20. #include "inc/graph_pass.h"
  21. namespace ge {
  22. /*
  23. * Atomic addr clean task fusion
  24. * Find all atomic op in graph,and insert one AtomicAddrClean op.
  25. * To clean atomic output and workspace once for all.
  26. * before iteration starts, empty AtomicAdd output, workspace memory
  27. * op1 op1
  28. * | |
  29. * op2(atomic) ==> op2
  30. * | | \
  31. * op3(atomic) op3 -AtomicClean
  32. */
  33. class AtomicAddrCleanPass : public GraphPass {
  34. public:
  35. Status Run(ComputeGraphPtr graph);
  36. Status ClearStatus() override;
  37. private:
  38. /**
  39. * HandleLoopGraph
  40. * @param graph
  41. * @return
  42. */
  43. Status HandleLoopGraph(ComputeGraphPtr &graph, const vector<NodePtr> &atomic_node_vec);
  44. /**
  45. * HandleNormalGraph
  46. * @param graph
  47. * @return
  48. */
  49. Status HandleNormalGraph(ComputeGraphPtr &graph, const vector<NodePtr> &atomic_node_vec);
  50. /**
  51. * Insert atomic clean node to graph
  52. * @param graph
  53. * @return
  54. */
  55. NodePtr InsertAtomicAddrCleanNode(ComputeGraphPtr &graph);
  56. /**
  57. * Link control anchor from atomic clean node to atomic node
  58. * @param atomic_node
  59. * @param atomic_clean_node
  60. * @return
  61. */
  62. Status LinkToAtomicNode(const NodePtr &atomic_node, NodePtr &atomic_clean_node);
  63. /**
  64. * Link atomic clean node to all potential precedence nodes which may execute before atomic clean node
  65. * @param graph
  66. * @param atomic_clean_node
  67. * @return
  68. */
  69. Status LinkToPotentialPrecedenceNode(ComputeGraphPtr &graph, NodePtr &atomic_clean_node);
  70. /**
  71. * Check if this node is atomic op.
  72. * @param node
  73. * @return
  74. */
  75. bool IsAtomicOp(const NodePtr &node);
  76. /**
  77. * Handle atomic node in unknown graph
  78. * @param atomic_node_vec: atomic node vector in unknown graph
  79. * @return
  80. */
  81. Status CompileUnknownGraphOp(const vector<NodePtr> &atomic_node_vec);
  82. Status HandleDispersedAtomicNodes(ComputeGraphPtr &graph, const std::vector<NodePtr> &atomic_node_vec,
  83. std::vector<NodePtr> &common_atomic_nodes);
  84. bool CheckAtomicFromOpsKernel(const NodePtr &node);
  85. bool IsOutputIndexPeerInputAtomic(const NodePtr &node, int64_t output_index);
  86. bool CheckSkipInsertInLoopGraph(const NodePtr &node);
  87. vector<NodePtr> hcom_node_vec_;
  88. bool is_loop_graph_ = false;
  89. };
  90. } // namespace ge
  91. #endif // GE_GRAPH_PASSES_ATOMIC_ADDR_CLEAN_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