Expression and Loop Libraries for High-Performance Code Synthesis

Abstract

To simultaneously provide rapid application development and high performance, developers of scientific and multimedia applications often mix languages, using scripting languages to glue together high-performance components written in compiled languages. While this can be a useful development strategy, it distances application developers from the optimization process while adding complexity to the development tool chain. Recently, we introduced the Synthetic Programming Environment (SPE) for Python, a library for generating and executing machine instructions at run-time. The SPE is an effective platform for optimizing serial and parallel applications without relying on intermediate languages. In this paper, we use the SPE to develop two code generation libraries, one for scalar and vector (SIMD) expression evaluation and another for parallel and high-performance loop generation. Using these libraries, we implement two high performance kernels and demonstrate how to achieve high levels of performance by providing the appropriate abstractions to users. Our results show that scripting languages can be used to generate high-performance code and suggest that providing optimizations as user-level libraries is an effective strategy for managing the complexity of high-performance applications.