Benchmark software for ArrayFire

The benchmarking program requires the following system-level libraries:

• ncurses
• python
• bokeh

On Ubuntu these dependencies are most easily installed via. the package manager an Anaconda Python. First install ncurses via. the package manager:

sudo apt-get install libncurses5-dev

Next, download and install Anaconda from Continuum Analytics. Once this is complete, run

conda install bokeh

which will automatically download and install all required packages.

Build and install the ArrayFire library following instructions here:

• ArrayFire

Note, you may install ArrayFire to a non-system path if needed.

Basic building instructions:

git clone --recursive https://github.com/bkloppenborg/arrayfire_benchmark.git
cd arrayfire_benchmark
cd build
cmake ..
make

If you have ArrayFire installed in a non-standard location, specify the directory which contains the ArrayFireConfig* files. These files may be found in the share/ArrayFire subdirectory of the installation folder. For example, if ArrayFire were installed locally to /opt/ArrayFire then we would modify the cmake command above to be:

cmake -DArrayFire_DIR=/opt/ArrayFire/share/ArrayFire ..

Install ArrayFire using the installer. Advanced users can opt to use custom builds, but this document will not detail steps for that.

You will also need Boost. You can install it using the Boost binary installers for Windows (VS2013 builds).

Open the CMake GUI. Source directory is arrayfire-benchmark and build directory is arrayfire-benchmark/build. Hit configure.

You may need to add/change the following:

• Add BOOST_ROOT and point it to the boost install directory.
• Add BOOST_LIBRARYDIR and point it to BOOST_ROOT/lib64-msvc-12.0

Run configure again.

Once the Boost libraries are found, you will need to add a prefix of "lib" to all the boost libraries. Example:

BOOST_SYSTEM_LIBRARY_RELEASE will change from C:/boost/lib64-msvc-12.0/boost_system-vc120-mt-1_58.lib to C:/boost/lib64-msvc-12.0/libboost_system-vc120-mt-1_58.lib.

Run configure once again. Then generate.

Now open the build/af_benchmark solution and build it.

First generate a series of benchmark results by running one of benchmark_cpu, benchmark_opencl, or benchmark_cuda programs with the -r output_file.csv option. These three programs have the same set of options which may be seen using the -h option.

After this, use the scripts/standalone-plot.py to visualize individual results from the benchmark suite (specify the -h option to see possibilities).

If you need to plot a lot of results, modify the scripts/make-plot.py script accordingly.