Section: New Results

CPU+GPU adaptive computation

We aim to automatically use CPU and GPU to jointly execute a parallel code. To ensure load balance between different PUs, thus to preserve performance, it is necessary to consider the underlying hardware and the program parameters. Compiler optimizations, execution context, hardware availability and specification make it difficult to determine execution times statically. To overcome this hurdle we rely on a portable and automatic method for predicting execution times of statically generated codes on multicore CPUs and on CUDA GPUs. This approach relies on three stages: automatic code generation, offline profiling of the target code and online prediction.

This is mainly the work of PhD student Jean-François Dollinger, advised by Vincent Loechner since 2011. Preliminary results, a "fastest-wins" algorithm between a multicore CPU and the best predicted GPU code version, was published in 2013 in ICPP. Our latest advances, load balancing code between multiple cores CPUs and multiple GPUs will be presented at the IMPACT 2015 workshop [25] in conjunction with the HiPEAC conference. We are currently preparing an extended journal paper to present this work, and Jean-François Dollinger will defend his PhD in 2015.