Programming Constructs to Enable Formation of Efficient Algorithm Mapping for ExaScale Processors

Frontier Technology, Inc. (FTI) and its research partner University of Florida (UF) propose to develop new software tools for addressing the requirements of semi-automatically programming emerging classes of computational hardware for Exa-scale computing so as to benefit the DoD through reduced software production and overall life cycle costs with greater portability, fewer defects, and energy reduction. Phase I will: Co-design higher level algorithmic expressions with specific optimizations for parallelism and locality; Perform proof of concept measurements of improvements beyond what is achieved with common high level languages; Develop a mathematical basis of the underlying optimization model to illustrate the exploitation of parallelism and locality across stages of sensor signal processing; Benchmark model kernels relevant to Air Force sensor missions. Phase I will design, analyze, prototype, and demonstrate implementations of the above-listed computational strategies on a multicore CPU, graphics processing unit (GPU), or CPU-GPU cluster, with analysis of cluster size and proof of scalability. Phase II will implement new higher-level algorithm-specific language and optimizations; optimize these on mission relevant sensor data processing kernels; establish porting benefits; demonstrate parallelism, locality across stages of sensor computational chains, and methodology for application to Air Force missions and commercialization.