Automating GPU computing in MATLAB
Indiana University, Bloomington, IN, USA
Proceedings of the international conference on Supercomputing, ICS ’11, 2011
@inproceedings{shei2011automating,
title={Automating GPU computing in MATLAB},
author={Shei, C.Y. and Ratnalikar, P. and Chauhan, A.},
booktitle={Proceedings of the international conference on Supercomputing},
pages={245–254},
year={2011},
organization={ACM}
}
MATLAB is a popular software platform for scientific and engineering software writers. It offers a high level of abstraction for fundamental mathematical operations and extensive highly optimized domain-specific libraries for several scientific and engineering disciplines. With the recent availability of GPU libraries for MATLAB, it has become possible to easily exploit GPGPUs as coprocessors. However, this requires changing the code by carefully declaring variables that would live on the GPU, breaking the simplicity of the MATLAB programming model. We present a fully automatic source-level compilation technique to exploit a given GPU library for MATLAB, enabling coarse-grained heterogeneous parallelism across CPU and GPU. Our approach is based on empirically characterizing the library’s functions, in order to build a comparative model of their performance on the CPU and GPU, which is then used along with a data communication cost model to maximize parallelism by selectively offloading some computation on the GPU. We achieve this by phrasing the problem as a binary integer linear programming problem aimed at minimizing CPU-GPU data movement, and using a hierarchical approach to keep the computational complexity in check. We have implemented our approach in a source-level MATLAB compiler, and present experimental results on a set of MATLAB kernels and applications using the GPUmat library. We show speedups of up to 7 times when the GPU is harnessed, compared to a standalone 8-core CPU.
August 28, 2011 by hgpu
