Auto-vectorization: recent development and prospect

Abstract

The technology of SIMD is developing rapidly, and quite a few auto-vectorization methods have been proposed.Auto-vectorization can automatically translate scalar programs into vector programs based on SIMD extension, decrease workload of the programmers in coding vector programs, and effectively improve performance of programs.Based on that, the research achievements in the field of automatic vectorization in recent 10 years were analyzed and summarized.The key problems and major breakthroughs in automatic vectorization were classified from four aspects:semantic-maintaining analysis and transformation, vectorization grouping analysis and transformation, processor-oriented analysis and transformation, and performance evaluation analysis.Furtherly, the development trends and research directions of the four aspects were prospected.

Keywords

• auto-vectorization
• SIMD extension
• compiling technology
• data level parallelism
• performance optimization