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.

compile_nodes_pass.h 1.6 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
4 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748
  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_COMPILE_NODES_PASS_H_
  17. #define GE_GRAPH_PASSES_COMPILE_NODES_PASS_H_
  18. #include <memory>
  19. #include <string>
  20. #include <vector>
  21. #include <unordered_map>
  22. #include "inc/graph_pass.h"
  23. #include "init/gelib.h"
  24. namespace ge {
  25. ///
  26. /// compile nodes
  27. ///
  28. class CompileNodesPass : public GraphPass {
  29. public:
  30. CompileNodesPass() {}
  31. virtual ~CompileNodesPass() {}
  32. graphStatus Run(ComputeGraphPtr graph) override;
  33. private:
  34. graphStatus GetSupportedKernel(const NodePtr &node, const std::shared_ptr<GELib> instance, string &kernel_lib_name);
  35. bool CheckAccuracySupport(const OpsKernelInfoStorePtr &kernel_info, const std::shared_ptr<GELib> instance,
  36. const NodePtr &node, string& unsupported_reason);
  37. graphStatus CompileNodes(const std::shared_ptr<GELib> instance,
  38. std::unordered_map<string, vector<NodePtr>> &kernel_to_compile_nodes);
  39. };
  40. } // namespace ge
  41. #endif // GE_GRAPH_PASSES_COMPILE_NODES_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