high performance computing on graphics processing units: hgpu.org

Present need to move towards heterogeneous architectures has been well established. This has increased the importance of parallelization of software to achieve good performance. Use of mixed architectures exponentially increases the need of the programmer to understand the intricacies of the underlying hardware to achieve optimal speedup. Obtaining optimal performance on one such architecture is difficult, but this process needs to be repeated for even small changes in the architecture or the source code. In these situations use of autotuning [1], [2] can not only reduce the amount of effort in trying to manually figure out the best optimization, but could also find configurations that are better than what was achieved by an expert on that architecture. We apply this machine learning technique on two Medical imagining applications. Like many other computationally intensive imaging applications these applications are best suited to running on inherently parallel architectures. We compare our results with a past study of the same applications where we modified the applications to run best on the given architectures. We achieved speedups of up to 159.8% improvement on an individual kernel basis, and 39.7% for overall application performance.