R-Car/Boards/Yocto-Gen3/OpenCL

From eLinux.org
Jump to: navigation, search

OpenCL support

The meta-opencl layers provide the following packages:

  • meta-ocl-rcar-gen3 layer:
    • cl-gles-user-module - OpenCL/OpenGL user module for Renesas R-Car R8A7795 SoC boards
  • meta-ocl-common layer:
    • libgfortran - GNU Fortran library (fixes)
    • opencv - Open Computer Vision library (fixes)
    • gflags - command line flags processing library
    • snappy - fast compressor/decompressor
    • lmdb - Lightning Memory-Mapped Database
    • openblas - optimized linear algebra library
    • viennacl - linear algebra library for computations on many-core architectures
    • clblas - library containing BLAS functions written in OpenCL
    • clblast - OpenCL BLAS library written in C++11
    • caffe - deep learning framework
    • renesas-opencl-sdk - separate set of Caffe/clBlas/ATLAS libraries and samples provided by Luxoft

Notes:

  • OpenCL is currently supported on the following boards:
    • Renesas R-Car H3/M3 SoC boards (Yocto 3.7.0 and above).
    • Renesas R-Car H3 SoC boards (prior to Yocto 3.7.0).
  • Luxoft renesas-opencl-sdk provides separate set of Caffe and clBlas libraries, and does not depend on caffe or clblas packages.

Downloading the prerequisites

Download basic Yocto layers as described in Yocto Start-Up Guide, and use the following commands to download meta-opencl layers:

  1. Clone meta-opencl git repository along with the other Yocto layers:
    git clone https://github.com/CogentEmbedded/meta-opencl.git
    
  2. Switch to the proper branch:

    • Yocto 2.12.0:
      git -C meta-opencl checkout v2.12.0
      
    • Yocto 2.19.0:
      git -C meta-opencl checkout v2.19.0
      
    • Yocto 2.23.0:

      git -C meta-opencl checkout v2.23.0
      
    • Yocto 3.7.0:

      git -C meta-opencl checkout v3.7.0
      
    • Yocto 3.9.0:

      git -C meta-opencl checkout v3.9.0
      

Download the OpenGL/OpenCL combined GPU libraries using the following link: https://www.renesas.com/us/en/solutions/automotive/rcar-download/rcar-opencl.html The zip archive should contain linux_gsx_binaries_cl_gles tarball that is to be used with meta-opencl/meta-ocl-rcar-gen3 layer.
Note: Yocto 3.7.0 tarball can be used with Yocto 3.9.0 as well.

Enabling OpenCL support in Yocto for R-Car Gen3 boards

Set up Yocto as usual, and use the following steps to enable OpenCL support before starting the build.

  1. Copy OpenCL/OpenGL binaries to the meta-opencl/meta-ocl-rcar-gen3 layer:
    • Yocto 3.7.0 and above:
      DEST=meta-opencl/meta-ocl-rcar-gen3/recipes-graphics/cl-gles-module/cl-gles-user-module 
      for f in *_linux_gsx_binaries_cl_gles.tar.bz2; do                                       
          cp ${f} ${DEST}/${f#EVA_} && echo "Copied: ${f}"                                      
      done
      

      Note: Strip the EVA_ prefix while copying when using evaluation version.

    • Prior to Yocto 3.7.0:

      cp r8a7795_linux_gsx_binaries_cl_gles3.tar.bz2 meta-opencl/meta-ocl-rcar-gen3/recipes-graphics/cl-gles-module/cl-gles-user-module/
      
  2. Add meta-python layer to BBLAYERS in bblayers.conf file:

      ${TOPDIR}/../meta-openembedded/meta-python \
  3. Add meta-opencl layers to BBLAYERS in bblayers.conf file:

      ${TOPDIR}/../meta-opencl/meta-ocl-common \
      ${TOPDIR}/../meta-opencl/meta-ocl-rcar-gen3 \
    
  4. Replace all occurrences of gles-user-module with cl-gles-user-module, and add PREFERRED_PROVIDER_virtual/opencl variable equal to "cl-gles-user-module" in local.conf:

    • Yocto 2.23.0 and above:
      PREFERRED_PROVIDER_virtual/libgles2 = "cl-gles-user-module"
      PREFERRED_PROVIDER_virtual/opencl = "cl-gles-user-module"
      
    • Prior to Yocto 2.23.0:

      PREFERRED_PROVIDER_virtual/libgles2 = "cl-gles-user-module"
      PREFERRED_PROVIDER_virtual/egl = "cl-gles-user-module"
      PREFERRED_PROVIDER_virtual/opencl = "cl-gles-user-module"
      
  5. Enable cl-gles-user-module instead of gles-user-module in local.conf:

    PREFERRED_PROVIDER_gles-user-module = "cl-gles-user-module"
    

Sample configuration files are available at meta-opencl/meta-ocl-rcar-gen3/docs/sample/conf.

Enabling Caffe support in Yocto

Once OpenCL is enabled, enable Caffe support by adding the following line to local.conf:

IMAGE_INSTALL_append = " caffe"

Caffe can be configured with PACKAGECONFIG variable. The following configuration features are supported:

  • opencv - enable OpenCV support (ON)
  • clblas - build using clBLAS instead of ViennaCL (OFF)
  • clblast - build using CLBlast instead of ViennaCL (OFF)
  • lmdb - enable LMDB support (OFF)
  • leveldb - enable LevelDB support (OFF)

Only OpenCV support is enabled by default. In case you need to use clBLAS instead of ViennaCL, add the following configuration variable to local.conf file:

PACKAGECONFIG_append_pn-caffe = " clblas"

If CLBlast support is needed, the following variable should be used:

PACKAGECONFIG_append_pn-caffe = " clblast"

In case database support is need, use the following variable in local.conf:

PACKAGECONFIG_append_pn-caffe = " leveldb lmdb"

Any number of features can be combined using PACKAGECONFIG variable, however clBLAS and CLBlast features are mutually exclusive, for example:

PACKAGECONFIG_append_pn-caffe = " clblas leveldb lmdb"

Finally, edit local.conf to add packages necessary for downloading and unpacking Caffe sample models:

IMAGE_INSTALL_append = " git python-argparse python-pyyaml python-six"

Enabling OpenCL SDK by Luxoft

Edit local.conf to add Renesas OpenCL SDK by Luxoft:

IMAGE_INSTALL_append += " renesas-opencl-sdk"

Building Yocto images

Start the build as usual:

bitbake core-image-weston

If the target board is to be used for natively building OpenCL applications, it might be useful to follow additional configuration steps below.

  1. Enable Fortran support in local.conf:
    FORTRAN_forcevariable = ",fortran"
    RUNTIMETARGET_append_pn-gcc-runtime = " libquadmath"
    IMAGE_INSTALL_append = " gfortran gfortran-symlinks libgfortran libgfortran-dev"
    
  2. Enable cmake in local.conf:

    IMAGE_INSTALL_append = " cmake"
    
  3. Enable boost development libraries in local.conf:

    IMAGE_INSTALL_append = " boost-dev boost-staticdev"
    
  4. Enable other development libraries in local.conf:

    IMAGE_INSTALL_append = " gflags-dev snappy-dev protobuf-dev leveldb-dev lmdb-dev opencv-dev opencv-apps"
    
  5. Enable BLAS development libraries in local.conf:

    IMAGE_INSTALL_append = " caffe-dev clblas-dev clblast-dev viennacl-dev openblas-dev"
    
  6. Prior to Yocto 3.7.0: disable cpio.gz image generation in local.conf to avoid 2GB cpio image size limitation issues:

    IMAGE_FSTYPES_remove = "cpio.gz"
    
  7. Remove Linux kernel sources and LTP to make SDK image smaller if necessary:

    IMAGE_INSTALL_remove = "kernel-devsrc ltp"
    
  8. Build SDK image:

    bitbake core-image-weston-sdk
    

Testing OpenCL on R-Car Gen3 boards

Use the following command to start OpenCL unit tests on the board:

ocl_unit_test

The application should perform the unit tests, and print results to the console. The last message should indicate the number of passed tests:

Finished 31 tests in 171.5 seconds: 31 passed, 0 failed (100.00%)

Note: OpenCL unit test is included in the cl-gles-module-dev package which is added to core-image-weston-sdk by default.

Testing ViennaCL

Use the following command to display ViennaCL information:

viennacl-info

Make sure that GPU device type is used:

Type:                          GPU

Testing Caffe

Before Caffe classification test can be started, Caffe model has to be downloaded and unpacked. The next two steps can be done on the target board directly, provided the board is connected to the internet, and DNS is properly setup. Otherwise, the model can be downloaded to the host system and copied to the target root filesystem later.

  1. Download Caffe sources:
    git clone -b opencl -n git://github.com/BVLC/caffe.git
    cd caffe
    git checkout -b tmp e0f77c3b5f4837615f05b097ba3f2a05d7413e58
    
  2. Download and unpack Caffe model:

    ./scripts/download_model_binary.py models/bvlc_reference_caffenet
    ./data/ilsvrc12/get_ilsvrc_aux.sh
    
  3. Run Caffe classification test on the target board from Caffe source directory:

    classification \
    models/bvlc_reference_caffenet/deploy.prototxt \
    models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel \
    data/ilsvrc12/imagenet_mean.binaryproto \
    data/ilsvrc12/synset_words.txt \
    examples/images/cat.jpg
    
  4. The test should detect the cat in the image, and produce the following output to the console:

    Use GPU with device ID 0
    ---------- Prediction for examples/images/cat.jpg ----------
    0.3134 - "n02123045 tabby, tabby cat"
    0.2380 - "n02123159 tiger cat"
    0.1235 - "n02124075 Egyptian cat"
    0.1004 - "n02119022 red fox, Vulpes vulpes"
    0.0715 - "n02127052 lynx, catamount"
    

Testing Renesas OpenCL SDK by Luxoft

  1. Export RENESAS_OPENCL_SDK_ROOT directory, for example:
    export RENESAS_OPENCL_SDK_ROOT=/opt/renesas/renesas-opencl-sdk-1.0.5
    
  2. Create SDK working directory:

    mkdir -p /home/root/RenoclSDK
    
  3. Copy binaries and libs from RENESAS_OPENCL_SDK_ROOT to SDK working directory:

    cp $RENESAS_OPENCL_SDK_ROOT/__output/* /home/root/RenoclSDK
    
  4. Copy resources from RENESAS_OPENCL_SDK_ROOT to SDK working directory:

    cp -r $RENESAS_OPENCL_SDK_ROOT/Resources /home/root/RenoclSDK
    
  5. Export LD_LIBRARY_PATH:

    export LD_LIBRARY_PATH=/home/root/RenoclSDK:$LD_LIBRARY_PATH
    
  6. Change current working directory to the SDK working directory:

    cd /home/root/RenoclSDK
    
  7. Run Face Detection Samples, for example:

    ./BasicFaceDetHaar -image Resources/images/singleFace.jpg
    ./OpenCLFaceDetHaar -image Resources/images/singleFace.jpg
    
  8. Run Deep Learning Samples, for example:

    ./BasicCaffeSegNet -image Resources/Segnet/0001TP_008550.png
    ./OpenCLCaffeSegNet -image Resources/Segnet/0001TP_008550.png
    

For further details, please refer to the documentation available at $RENESAS_OPENCL_SDK_ROOT/documents.