The goal of this thesis is to provide Erlang with an easy to use, high performance, data parallel skeleton library exploiting GPUs’ processing power. We choose OpenCL to be vendor neutral. The skeletons must be easy to use, possibly be a nearly perfect drop-in replacement to standard Erlang functions with similar semantic (minimal disruption principle).
We developed a prototype Erlang library, using OpenCL, implementing the data parallel skeletons to accelerate computations on lists of Erlang floats. The user code for such computations is specified writing OpenCL kernels using the OpenCL C programming language. The library has a very user-friendly API since it is modeled after Erlang’s list module, with which every Erlang programmer is familiar. The outcome of the tests shows, on one hand, a significant performance improvement over a pure Erlang implementation and, on the other hand, an acceptable performance loss over a C++/OpenCL reference implementation.