SIParCS 2017- Cuong Manh Nguyen
Portable Performance of Wavelet Compression Using Lifting Scheme
High performance computing requirements have brought many heterogeneous computational resources, such as multiple- and many-core architectures, into a single machine. However, the architecture, programming and memory models of these devices differ dramatically. Utilizing heterogeneous resources with the diversity of devices becomes a major challenge. For many applications, a new programming model which can adapt to heterogeneous computational resources is essential to achieving portability and high performance. In this study, we implement the lifting scheme wavelet transform using both data parallel primitives from VTK-m, an architecture-agnostic toolkit, and native CUDA to investigate the potential for portable performance over multiple architectures. Our results show that the performance of VTK-m is comparable with nVidia's hardware-specific CUDA programming language.
Mentors: John Clyne, Samuel Li