Using Architecture and Code Optimization Techniques to Create Fast and Effective Data Compressors

Опубликовано 7 сентября 2016, 16:43

This talk illustrates how to exploit computer architecture and compiler optimization ideas to achieve breakthroughs in lossless program-trace and floating-point compression, which in turn benefit architecture and compiler researchers. We start with an overview of value prediction hardware. Then we show how these prediction concepts can be used to design trace compression algorithms that typically outperform other approaches in both speed and compression ratio. Next, we present the compiler optimization techniques we employed to automatically synthesize such compressors and tailor them to user-specified trace formats. Finally, we discuss the necessary modifications to convert our algorithm into a fast and effective lossless floating-point compression algorithm.