Abstracts of the 2015 east coast computer algebra day

Abstract

In the past decade, the introduction of low-level heterogeneous programming models, in particular CUDA, has brought supercomputing to the level of the desktop computer. However, these models bring notable challenges, even to expert programmers. Indeed, fully exploiting the power of hardware accelerators with CUDA-like code often requires significant code optimization e.ort. While this development can certainly yield high performance, it is desirable for some programmers to avoid the explicit management of device initialization and data transfer between memory levels. To this end, high-level models for accelerator programming, like OpenMP and OpenACC, have become an important research direction. With these models, programmers only need to annotate their C/C++ code to indicate which code portion is to be executed on the device and how data maps between host and device.