README

Image Skeletonisation
=====================

A performance comparison with Thrust


Description
-----------

This project was realised during the fall semester 2014 at EPFL for CS-425,
Program parallelisation on PC clusters.

It aims to compare a theoretical analysis of the parallelisation of the image
skeletonisation problem against an actual implementation in C++ using Thrust,
a high level algorithm library with CUDA, TBB and OpenMP backends. The goal is
to compare the speedup of the parallel implementation and highlight the effect
of some hardware properties on the scalability of the performances.

The full report is available in `Image Skeletonisation - report.pdf`.


Build Instructions
------------------

To run all implementations and measure running time, use `make run`. You will
need the software described in the `Hardware & Software Specifications` section
of the report.

Binaries of the required boost, SFML and libjpeg dependencies for Mac OS X are
provided for convenience. If you are not using this operating system or the same
compilers, you will need to build those yourself and potentially edit the
`Makefile`.

Adding `-DENABLE_SAVE_IMAGE` to the compilation flags will enable saving the
skeletonised image and include I/O times in the running time measurement.

Similarly, `-DENABLE_PARALLEL_SHORT_CIRCUIT_WHITE` enable some optimisations
in the parallel implementation. Note that on CUDA this might decrease the
speedup due to thread divergence.


License
-------

The source code included in this project is distributed under the zlib/png
license. Please refer to LICENSE for the full text.

The boost library is distributed under Boost Software License, which is
available in lib/BOOST_LICENSE. The SFML library is distributed under
zlib/png license, which is available in lib/SFML_LICENSE. License for libjpeg is
available in lib/LIBJPEG_LICENSE.