Cuda + cuDNN Tips & Tricks

cuda_tips_and_tricks.md

Cuda + cuDNN Tips & Tricks

Installation

CUDA + cuDNN Installation:

• Install Cuda (2 options):

  • Download from NVIDIA website. Will install to /Developer/NVIDIA/CUDA=#-# (macOS / OS X), with symlinks in /usr/local/cuda/.
  • Alternatively, use brew cask install cuda.

• Download CuDNN from NVIDIA website into local_cudnn_path, then:

  • sudo mkdir /usr/local/cudnn
  • sudo cp -r local_cudnn_path/* /usr/local/cudnn/

• Setup paths in ~/.bash_profile:

  • export CUDA_HOME=/usr/local/cuda
  • export CUDNN_HOME=/usr/local/cudnn
  • macOS / OS X:
    • export DYLD_LIBRARY_PATH=$DYLD_LIBRARY_PATH:$CUDA_HOME/lib:$CUDNN_HOME/lib
  • Linux:
    • sudo apt-get install libcupti-dev
    • export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CUDA_HOME/lib64:$CUDNN_HOME/lib64:$CUDA_HOME/extras/CUPTI/lib64

• Test:

  cp -r /usr/local/cuda/samples ~/cuda-samples
  pushd ~/cuda-samples
  make
  popd
  ~/cuda-samples/bin/x86_64/darwin/release/deviceQuery
  

• Troubleshoot Failed to initialize NVML: Driver/library version mismatch error (via https://stackoverflow.com/a/45319156):

  sudo rmmod nvidia_drm
  sudo rmmod nvidia_modeset
  sudo rmmod nvidia_uvm
  sudo rmmod nvidia
  nvidia-smi