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.

Dockerfile 1.6 kB

123456789101112131415161718192021222324252627282930313233343536373839404142
  1. # this dockerfile used for graphengine build
  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. FROM ubuntu:18.04
  17. RUN apt-get update \
  18. && apt-get install -y git g++ wget unzip clang-format-9 build-essential lcov vim
  19. # install for doxygen
  20. RUN apt-get install -y graphviz doxygen
  21. # install for graph ensy engine
  22. RUN cpan install -y Graph::Easy
  23. RUN wget https://cmake.org/files/v3.16/cmake-3.16.7-Linux-x86_64.tar.gz
  24. RUN mkdir -p /opt/cmake-3.16.7 \
  25. && tar -xvf cmake-3.16.7-Linux-x86_64.tar.gz -C /opt/cmake-3.16.7 --strip-components=1 \
  26. && ln -sf /opt/cmake-3.16.7/bin/* /usr/bin/ \
  27. && mv /usr/bin/clang-format-9 /usr/bin/clang-format
  28. RUN wget https://github.com/ccup/lcov/archive/refs/tags/add_lcov.tar.gz -O add_lcov.tar.gz \
  29. && mkdir -p /opt/addlcov1.0.0 \
  30. && tar -xvf add_lcov.tar.gz -C /opt/addlcov1.0.0 \
  31. && mv /opt/addlcov1.0.0/lcov-add_lcov/bin/lcov /usr/bin/addlcov
  32. ENV PROJECT_HOME=/code/Turing/graphEngine
  33. RUN echo "alias ge=/code/Turing/graphEngine/scripts/ge.sh">>~/.bashrc

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