high performance computing on graphics processing units: hgpu.org

Many reconfigurable processors allow their instruction sets to be tailored according to the performance requirements of target applications. They have gained immense popularity in recent years because of this flexibility of adding custom instructions. However, most design automation algorithms for instruction set customization (like enumerating and selecting the optimal set of custom instructions) are computationally intractable. As such, existing tools to customize instruction sets of extensible processors rely on approximation methods or heuristics. In contrast to such traditional approaches, we propose to use GPUs (Graphics Processing Units) to efficiently solve computationally expensive algorithms in the design automation tools for extensible processors. To demonstrate our idea, we choose a custom instruction selection problem and accelerate it using CUDA (CUDA is a GPU computing engine). Our CUDA implementation is devised to maximize the achievable speedups by various optimizations like exploiting on-chip shared memory and register usage. Experiments conducted on well known benchmarks show significant speedups over sequential CPU implementations as well as over multi-core implementations.