AMDGPU OpenCL Caffe

Enablement for Caffe (OpenCL) on AMD GPUs

System Requirement

• Ubuntu 16.04 or equivalent
  • Note that a supported (valid) opencl enabled driver stack is needed from AMD
• AMD Radeon GPU (tested on):
  • Polaris (RX460)
  • Fiji (R9 Fury, Fury X)
  • Tonga (FirePro W7100)

1. Update to latest Canonical release

• Update the apt metadata: sudo apt-get update
• Upgrade the packages: sudo apt-get upgrade

2. Install general dependencies

• sudo apt-get install libprotobuf-dev libleveldb-dev libsnappy-dev libopencv-dev libhdf5-serial-dev protobuf-compiler libatlas-base-dev libblas-dev libgflags-dev libgoogle-glog-dev liblmdb-dev libboost-all-dev cmake git python-numpy python-pip cmake

3. Install AMDGPU-PRO Driver for Linux

• **Optional - You can install driver by executing the helper script "driver_install.sh" script that is supplied in this repository.
• Download AMDGPU-PRO Driver
• Follow instruction to install AMDGPU-PRO Driver for Ubuntu
• Reboot to the kernel that loads the installed drivers

4. Download Caffe

• Download Caffe (on this example, download repository to home folder)
  • > cd ~ ; git clone https://github.com/BVLC/caffe
  • > cd caffe ; git checkout opencl
  • > export OCLCAFFE=$PWD
• NOTE - tested commit c61d48746b2df1d237c64abc7416342ce98f3251

5. Copy ViennaCL library

• Download ViennaCL for Linux
• Extract downloaded folder > tar -xvf ViennalCL-<version>.tar.gz
• Copy extracted folder to Caffe > cp -r ViennalCL-<version> ~/caffe

6. Install Caffe

• Create directory for cmake: > mkdir build
• Navigate to build: > cd build
• Configure and cmake Caffe
  • > cmake -DViennaCL_INCLUDE_DIR=../ViennaCL-<version> -DOPENCL_INCLUDE_DIRS=../ViennaCL-<version>/CL/ -DOPENCL_LIBRARIES=/opt/amdgpu-pro/lib/x86_64-linux-gnu/libOpenCL.so.1 ..
  • Note: By default OpenCL libraries for AMDGPU is at /opt/amdgpu-pro/x86_64-linux-gnu/libOpenCL.so.1
• Compile Caffe: > make
  • Note: You can enable multithreaded compilation with-j<# threads_available> to speed up the process. (i.e. for an 8 core machine we can use : make -j8)
• Install Caffe: > make install
• Test Caffe: > make runtest
  • Note: For multithreaded build, add -j<# cores_available> to speed up process. (i.e. for an 8 core machine we can use : make -j8)

7. Try CIFAR (CANADIAN INSTIT. FOR ADVANCED RESEARCH DATA TRAINING)

• Go to your caffe path : cd $OCLCAFFE

• Invoke the commands from the caffe top module folder (i.e ~/caffe):

  • ./data/cifar10/get_cifar10.sh
  • ./examples/cifar10/create_cifar10.sh
  • ./examples/cifar10/train_quick.sh

• Note: You can substitute train_quick with train_full as needed. This will increase the time needed to fully train the model, however.

• Enabling Caffe's pythonic interface (PyCaffe):

  • Navigate to python in /caffe > cd python
  • Install all required packages with pip > for req in $(cat requirements.txt); do sudo -H pip install $req; done

• Optional: Add PyCaffe module to $PYTHONPATH > export PYTHONPATH=$OCLCAFFE/python:$PYTHONPATH