Computer Generation of Multicore Software for Software-Defined Radio

Future software platforms starting with today’s multicores will be massively parallel, offer unprecedented performance, and will be exceedingly difficult to program. In this project, we will work towards automatic development of high-performance software for these platforms. Specifically, we will design a tool that can computer-generate the performance critical components of software-defined radio (SDR) given only a high-level mathematical specification. Our tool will automatically perform all difficult optimizations including parallelization, vectorization, and memory hierarchy tuning at a high –level of abstraction before mapping to actual source code. This way, inherent shortcomings of compilers are overcome. As SDR components we consider the discrete Fourier transform, correlation, convolution, and Viterbi decoder as required by SDR. As target platform, we will use Clearspeed with 192 cores) and Tilera (with 64 cores), but note that our approach can be easily ported to others. Our goal is to eventually achieve with our computer-generated code the same performance, or even better, of any competing hand-tuned implementations of the same functionality.