Advanced Code Generation in Digital Signal Processing

Numerical software of the future will be different from what it looks nowadays. Nevertheless, reliability, portability across the rapidly evolving platforms, and satisfactory run time efficiency will remain to be standard requirements requested from high quality software. However, continually increasing problem sizes, computing environments which are exceedingly difficult to utilize, and new ways of software usage via the Internet and computational grids make it harder and harder to achieve these traditional goals. Special purpose numerical software in conjunction with special purpose compilers will gain in importance. This evolution of a radically new way of software development will be pursued in the proposed research project. The ultimate goal is to automate the process of architecture dependent performance tuning. The proposed two-year research project will deal with the machine generation of high-performance implementations of discrete linear transform algorithms that are automatically adapted to a given computing platform. Code generated by applying the newly developed methods will take full advantage of modern architectural features and instruction set extensions, deep memory hierarchies, software initiated prefetching, as well as multiple CPUs in both shared memory and distributed memory multiprocessors, up to IBM`s top performance computers of BG/L type. Research will concentrate on discrete linear transforms like Fourier transforms, sine and cosine transforms, as well as wavelet transforms.