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.

allreduce_fusion_pass.h 2.9 kB

4 years ago
4 years ago
1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556
  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_OPTIMIZE_OPTIMIZER_ALLREDUCE_FUSION_PASS_H_
  17. #define GE_GRAPH_OPTIMIZE_OPTIMIZER_ALLREDUCE_FUSION_PASS_H_
  18. #include <unordered_set>
  19. #include <utility>
  20. #include <vector>
  21. #include "inc/graph_pass.h"
  22. namespace ge {
  23. //
  24. class AllReducePass : public GraphPass {
  25. public:
  26. Status Run(ge::ComputeGraphPtr graph) override;
  27. private:
  28. Status GetPeerOutDataToInData(std::unordered_set<void *> &anchorSet,
  29. vector<ge::OutDataAnchorPtr> &peerOutDataAnchorVec, ge::NodePtr &srcNodePtr,
  30. ge::OpDescPtr &dstOpDescPtr);
  31. Status GetPeerOutDataToInControl(std::unordered_set<void *> &anchorSet,
  32. vector<ge::OutDataAnchorPtr> &peerOutDataToInControlVec, ge::NodePtr &srcNodePtr);
  33. Status GetPeerOutControlToInControl(std::unordered_set<void *> &anchorSet,
  34. vector<ge::OutControlAnchorPtr> &peerOutControlToInControlVec,
  35. ge::NodePtr &srcNodePtr);
  36. Status GetPeerAnchorFromOutData(std::unordered_set<void *> &anchorSet,
  37. vector<std::pair<int, ge::InDataAnchorPtr>> &peerInDataFromOutDataVec,
  38. vector<std::pair<int, ge::InControlAnchorPtr>> &peerInControlFromOutDataVec,
  39. ge::NodePtr &srcNodePtr, ge::OpDescPtr &dstOpDescPtr, int &index);
  40. Status GetPeerInControlFromOutControl(std::unordered_set<void *> &anchorSet,
  41. vector<ge::InControlAnchorPtr> &peerInControlFromOutControlVec,
  42. ge::NodePtr &srcNodePtr);
  43. Status GetPeerOutDataToInData(std::unordered_set<void *> &anchorSet,
  44. std::vector<ge::OutDataAnchorPtr> &peerOutDataAnchorVec,
  45. ge::NodePtr &srcNodePtr);
  46. Status GetPeerInAnchorToOutData(std::unordered_set<void *> &anchorSet,
  47. std::vector<std::pair<int, ge::InDataAnchorPtr>> &fusionOpPeerInDataAnchor,
  48. std::vector<std::pair<int, ge::InControlAnchorPtr>>&fusionOpPeerInControlFromOutData,
  49. ge::NodePtr &srcNodePtr);
  50. };
  51. } // namespace ge
  52. #endif // GE_GRAPH_OPTIMIZE_OPTIMIZER_ALLREDUCE_FUSION_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