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.

infer_base_pass.h 3.0 kB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465
  1. /**
  2. * Copyright 2021 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_INFER_BASE_PASS_H_
  17. #define GE_GRAPH_PASSES_INFER_BASE_PASS_H_
  18. #include "graph/passes/base_pass.h"
  19. namespace ge {
  20. class InferBasePass : public BaseNodePass {
  21. public:
  22. Status Run(NodePtr &node) override;
  23. graphStatus InferAndUpdate(NodePtr &node, bool before_subgraph, std::set<NodePtr> &changed_nodes);
  24. void PrintInOutTensors(const NodePtr &node, const std::string &phase);
  25. protected:
  26. virtual std::string SerialTensorInfo(const GeTensorDescPtr &tensor_desc) const = 0;
  27. virtual bool NeedInfer(const NodePtr &node) const;
  28. virtual graphStatus Infer(NodePtr &node) = 0;
  29. /**
  30. * Update the output TensorDesc by src TensorDesc. This will be called when updating peer node input desc.
  31. * @param src, input TensorDesc
  32. * @param dst, output TensorDesc to be updated
  33. * @return
  34. */
  35. virtual graphStatus UpdateTensorDesc(const GeTensorDescPtr &src, GeTensorDescPtr &dst, bool &changed) = 0;
  36. /**
  37. * Update the output TensorDesc for nodes which contain subgraphs.
  38. * In dynamic multi-dims/batch/images size scene, the update process maybe different,
  39. * in which case, the `InferBasePass` will call method `UpdateOutputFromSubgraphsForMultiDims` instead.
  40. * @param src, input TensorDesc from NetOutput nodes in all subgraphs
  41. * @param dst, output TensorDesc to be updated
  42. * @return
  43. */
  44. virtual graphStatus UpdateOutputFromSubgraphs(const std::vector<GeTensorDescPtr> &src,
  45. GeTensorDescPtr &dst) = 0;
  46. virtual graphStatus UpdateOutputFromSubgraphsForMultiDims(const std::vector<GeTensorDescPtr> &src,
  47. GeTensorDescPtr &dst) = 0;
  48. private:
  49. void AddChangedNodesImmediateRepass(const std::set<NodePtr> &changed_nodes);
  50. bool ContainsSubgraph(const NodePtr &node);
  51. std::vector<ComputeGraphPtr> GetCurNodeSubgraphs(const NodePtr &node);
  52. graphStatus UpdateTensorDescToSubgraphData(NodePtr &node);
  53. graphStatus UpdateTensorDescToParentNodeOutput(NodePtr &node);
  54. graphStatus UpdateParentNodeContainsSubgraphs(NodePtr &node,
  55. const std::vector<std::vector<GeTensorDescPtr>> &ref_out_tensors);
  56. graphStatus UpdateTensorDescToPeerInputs(NodePtr &node, std::set<NodePtr> &changed_nodes);
  57. };
  58. } // namespace ge
  59. #endif // GE_GRAPH_PASSES_INFER_BASE_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