Python CuPy on MacBookPro

目次

FrontPage
・Random
・機械学習
・さくらVPS
・Fedora13
・SuSe10
・Docker
・Ansible
・Java
・Scala
・Python
・Ruby
・Lisp
・Computer
・GIS
・HTML
・Culture
・Link

訪問者

total: 2099
today: 4
yesterday: 6
now: 3

更新

最新の10件

人気の10件

MenuBar

CUDAのインストール †

→Python CUDA on MacBookPro

CUDA関連の環境変数の設定 (plist) †

• Shellから動かすぶんには、~/.bashrc とかに定義すればいいけど、PyCharm? に環境変数を渡すためには launchd に設定する必要がある この設定は macOS 10.12 Sierra 以降は無効 → どうすればいいかは後述
  • launchd は macOS の systemd/initd
• ~/Library/LaunchAgents?/ に plist 形式でログイン時に実行するコマンドを定義する
• -~/Library/LaunchAgents?/PATH.plist

  <?xml version="1.0" encoding="UTF-8"?>
  <!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
  <plist version="1.0">
  <dict>
  <key>Label</key>
  <string>PATH</string>
  <key>ProgramArguments</key>
  <array>
    <string>/bin/launchctl</string>
    <string>setenv</string>
    <string>PATH</string>
    <string>/Developer/NVIDIA/CUDA-9.1/bin:$PATH</string>
  </array>
  <key>RunAtLoad</key>
  <true/>
  </dict>
  </plist>
  

  • ~/Library/LaunchAgents?/DYLD_LIBRARY_PATH.plist

    <?xml version="1.0" encoding="UTF-8"?>
    <!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
    <plist version="1.0">
    <dict>
    <key>Label</key>
    <string>DYLD_LIBRARY_PATH</string>
    <key>ProgramArguments</key>
    <array>
      <string>/bin/launchctl</string>
      <string>setenv</string>
      <string>DYLD_LIBRARY_PATH</string>
      <string>/usr/local/cuda/lib:/Developer/NVIDIA/CUDA-9.1/lib:$DYLD_LIBRARY_PATH</string>
    </array>
    <key>RunAtLoad</key>
    <true/>
    </dict>
    </plist>
    

• この設定は macOS 10.12 Sierra 以降はこうする
  • 上記の設定だと、CuPy? が Cuda のライブラリを見つけられずに、実行時の (PythonからGPUコードへの) コンパイルでエラーが起きて動かない

    cupy.cuda.nvrtc.NVRTCError: NVRTC_ERROR unknown (7)
    cupy.cuda.compiler.CompileException: nvrtc: error: failed to load builtins

• Cuda関係のライブラリは、$(HOME)/lib:/usr/local/lib/:/lib/:/usr/lib のどこかに Symbolic link を作る

  #!/bin/sh
  for file in $1/*
  do
    b=`basename $file`
    ln -vs $file ~/lib/$b
  done
  

  みたいなシェルスクリプトを作って

  [~]$ ./mklink.sh /Developer/NVIDIA/CUDA-9.1/lib
  /Users/atsushi/lib/libcublas.9.1.dylib -> /Developer/NVIDIA/CUDA-9.1/lib/libcublas.9.1.dylib
  /Users/atsushi/lib/libcublas.dylib -> /Developer/NVIDIA/CUDA-9.1/lib/libcublas.dylib
  /Users/atsushi/lib/libcublas_device.a -> /Developer/NVIDIA/CUDA-9.1/lib/libcublas_device.a
  /Users/atsushi/lib/libcublas_static.a -> /Developer/NVIDIA/CUDA-9.1/lib/libcublas_static.a
  /Users/atsushi/lib/libcudadevrt.a -> /Developer/NVIDIA/CUDA-9.1/lib/libcudadevrt.a
  /Users/atsushi/lib/libcudart.9.1.dylib -> /Developer/NVIDIA/CUDA-9.1/lib/libcudart.9.1.dylib
  ...

CuPy?のインストール †

$ git clone https://github.com/cupy/cupy.git
$ cd cupy
$ git checkout v4.2.0
$ python setup.py clean
$ MACOSX_DEPLOYMENT_TARGET=10.13 CC=clang CXX=clang++ python setup.py install
$ pip install .
Processing /Users/atsushi/PycharmProjects/cupy
Requirement already satisfied: numpy>=1.9.0 in /Users/atsushi/.pyenv/versions/anaconda3-5.1.0/envs/jupyter-env/lib/python3.6/site-packages (from cupy==4.2.0) (1.14.3)
Requirement already satisfied: six>=1.9.0 in /Users/atsushi/.pyenv/versions/anaconda3-5.1.0/envs/jupyter-env/lib/python3.6/site-packages (from cupy==4.2.0) (1.11.0)
Requirement already satisfied: fastrlock>=0.3 in /Users/atsushi/.pyenv/versions/anaconda3-5.1.0/envs/jupyter-env/lib/python3.6/site-packages (from cupy==4.2.0) (0.3)
Building wheels for collected packages: cupy
  Running setup.py bdist_wheel for cupy ... done
  Stored in directory: /private/var/folders/cm/p00lclws0jxdmjlb89qyqq580000gn/T/pip-ephem-wheel-cache-47vpnvhd/wheels/f3/4b/b6/cf3092440fb9ebf35bfb25a4de43ee2a9609fae446258288e9
Successfully built cupy
Installing collected packages: cupy
  Found existing installation: cupy 4.2.0
    Uninstalling cupy-4.2.0:
      Successfully uninstalled cupy-4.2.0
Successfully installed cupy-4.2.0

なんか動いた †

import cupy as cp
a = cp.array([1,2,3,4,5])
b = cp.array([6,7,8,9,10])
c = a + b
print (c)

[ 7  9 11 13 15]

Python