high performance computing on graphics processing units: hgpu.org

In this report we’ve created a linear algebra API using OpenCL, for use with MATLAB. We’ve demonstrated that the individual linear algebra components can be faster when using the GPU as compared to the CPU. We found that the API is heavily memory bound, but still faster than MATLAB in our testcase. The API components were autotuned to obtain higher performance, though the components were still bound by memory transfer rates. MEX was used for the bindings from the API to MATLAB. The API was since used for modelling the poisson problem, and was able to solve the problem. We saw in this problem that we could create more specialized components for solving the problem, and obtain faster solving times. The linear algebra components excelled at rapid prototyping, being almost as easy to write as MATLAB code, and MATLAB code could be mostly replaced line by line with code from the API. The experiences from the poisson problem was taken on to the wave equation in 2d, and we observed the same trends.