Introduction
The OpenCL platform model has the concept of local memory, which is mapped to scrach-pad memory for a given OpenCL implementation. The question is whether using local memory is always benefitial? To answer this question, we have developed a code generator and a composer (Aristotle) for different memory access patterns. With the help of Aristotle, we can generate benchmarks with and without local memory use. In this way, we can quantify and predict the benefits of using local memory on diverse processors.
Our Approach
Reference
J. Fang, H. Sips, and A. L. Varbanescu, "Quantifying the performance impacts of using local memory for many-core processors," in Multi-/Many-core Computing Systems (MuCoCoS), 2013 IEEE 6th International Workshop on. IEEE, 2013, pp. 1-10. [Online]. Available: http://dx.doi.org/10.1109/mucocos.2013.6633601
J. Fang, H. Sips, and A. L. Varbanescu, "Aristotle: A performance impact indicator for the OpenCL kernels using local memory," Scientific Programming, vol. (Accepted).
Support or Contact
Having trouble with the code? Let us know via j.fang@tudelft.nl.