CAPP

Abstract

Parallel programming can provide higher computational performance over a sequential implementation by making use of the many cores available in parallel systems. However, the parallel-capable devices introduce complexity into the programming model. Current parallel programming API's such as OpenCL and CUDA provide interfaces to the parallel devices, but are complex and result in code which includes cross-cutting components, such as setting up the parallel programming context, compiling the parallel kernel, and transferring data between the host and device memory spaces when the kernel is executed. A C++ Aspect-Oriented based Parallel Programming (CAPP) framework is developed, using AspectC++ and OpenCL, which defines aspects to remove the cross-cutting components from the C++ code. The aspects set up the OpenCL context, compile the OpenCL kernel, and manage the data transfer between the memory spaces each time a kernel is executed. The aspects are woven into the C++ code before compilation, rather than at run time, which improves performance. An interface is provided for executing OpenCL kernels from the C++ code, essentially providing parallel programming in C++. The framework was applied to the SAXPY and Black-Scholes option pricing problems. Computational performance was, on average, 1.4-7% slower than the OpenCL implementation and up to 9 times faster than the sequential C++ implementation for the Black-Scholes problem. The amount of code was greatly reduced from the OpenCL implementation, and the resulting CAPP framework C++ code was simple and modular, resembling the sequential C++ implementation code. The amount of CAPP code required for the examples was found to be comparable to that of C++ AMP, another parallel programming framework.