Hisui: an Image and Video Processing Framework with Auto-optimizer

In order to develop a high-performance image and video processing program, it is necessary to tune the program to a target platform. However, this task not only requires a high level of knowledge of computer architecture and parallel processing, but also requires an enormous amount of development time. It makes the cost of program development a significant problem. To solve this problem, we are developing Hisui that aims to achieve both high productivity and high performance in the development of image and video processing programs. Hisui consists of a highly abstract Hisui language and Hisui compiler that compiles programs written using Hisui language. In this study, we design an automatic optimization function for Hisui and implement it on Hisui compiler. This function can automatically apply multi-threading, vectorization, and cache blocking. In order to design this function, we first investigated these optimization methods and identified the appropriate optimization application criteria and methods. Then, based on the results of the survey, we design an automatic optimization function and implement it on Hisui compiler. As a result of evaluating the usefulness of the improved image and video processing framework, we confirmed that the programs optimized by Hisui compiler are up to 3.0 times faster than those optimized by a combination of Halide and its Auto-scheduler.