Previous works in the literature have shown the feasibility of general purpose computations for non-visual applications on low-end mobile graphics processors using graphics APIs. These works focused only on the functional aspects of the software, ignoring the implementation details and therefore their performance implications due to their particular micro-architecture. Since various steps in such applications can be implemented in multiple ways, we identify optimisation opportunities, explore the different options and evaluate them. We show that the implementation details can significantly affect the obtained performance with discrepancies up to 3 orders of magnitude and we demonstrate the effectiveness of our proposal on two embedded platforms, obtaining more than 16× speedup over benchmarks designed following OpenGL ES 2 best practices.

This work has been partially supported by the Spanish Ministry of Science and Innovation under grant TIN2015-65316-P and the HiPEAC Network of Excellence.

Peer Reviewed