high performance computing on graphics processing units: hgpu.org

The High Performance Embedded Computing Software Initiative (HPEC-SI) is developing a unified software framework for computation and communication for high performance signal processing tasks on parallel computers. The goal of the program is to address the high cost of software in Department of Defense (DoD) systems by improving the portability and productivity of signal processing application development, while simultaneously improving performance compared to current practices. The Vector, Signal, and Image Processing Library (VSIPL) is a portable application programming interface (API) that is widely used for embedded DoD signal processing systems. One portion of the HPEC-SI effort includes the development of C++ extensions for the existing VSIPL standard, called VSIPL++. Commodity graphics processing units (GPUs) are application-specific processors that implement a standardized three-dimensional graphics-rendering pipeline, and provide significant floating-point processing capacity at much lower cost, power consumption, and physical space compared to general-purpose processors. Recent changes in GPUs have increased programmability and flexibility in portions of the rendering pipeline, allowing non-graphics applications to exploit their computational capacity. Restrictions on the programming model, lack of appropriate tools, unusual performance behavior, and other factors make exploiting GPUs a costly, difficult, and time-consuming process for application developers. The embedded systems that VSIPL and VSIPL++ are commonly used on share several important characteristics with GPUs, making VSIPL++ well suited to abstract and exploit GPUs. This paper describes GPUVSIPL++, an implementation of portions of the VSIPL++ standard that exploits a GPU to accelerate computation beyond what is possible on a development workstation