COMPUTATIONAL RESEARCH in BOSTON and BEYOND (CRIBB)

Recently there has been an explosion of new computer architectures that combine multiple CPUs with Special Purpose Processors (SPPs). Examples of SPPs include Field Programmable Gate Arrays (FPGAs), Graphics Processing Units (GPUs) and the Cell Broadband Engine. Powerful multicomputer platforms that combine SPPs and CPUs in a single hardware architecture promise tremendous performance benefits. Applications that can run on any of these platforms, deliver performance and be easily ported to other platforms are highly desirable. Traditional programming practices, however, intertwine application code with hardware specific code such that porting entails a significant rewrite of the application and reuse of code is difficult.

VSIPL++ for Reconfigurable Computing (Vforce) is a middleware framework that extends VSIPL++ (a C++ extension of the Vector, Signal, and Image Processing Library) to include support for special purpose processors (SPPs). Vforce is an extensible framework that allows the same application code to run on different heterogeneous computing platforms. Vforce offers application-level portability, framework-level extensibility to new hardware, and system-level run time resource management. In particular, Vforce supports very late binding of the application to a specific hardware platform such that binding does not occur until run time.

In this talk, I will give a brief introduction to VSIPL++ and to some commercially available heterogeneous multicomputers. Then I will present the Vforce framework and explain how it supports the three goals of performance, productivity and portability. I will present our experience using Vforce on different hardware platforms (including those with FPGAs and those with GPUs) as well as different applications.