Auto-tuning of Numerical Programs by Block Multi-color Ordering Code Generation and Job-Level Parallel Execution

摘要

Multi-color ordering is a parallel ordering that allows programs to be parallelized by application to sequentially executed parts of the programs. While multi-color ordering parallelizes sequentially executed parts with data dependences and increases the number of parts executed in parallel, improved performance by multi-color ordering is sensitive to differences in the architectures and systems on which the programs are executed. This sensitivity requires us to tune the numbers of colors; i.e., modify programs for each architecture and system. In this work, we develop a code generator based on multi-color ordering and automatically tune the number of colors using a job-level parallel scripting language Xcrypt. Furthermore, we support block multi-color ordering that avoids the disadvantage of stride accesses in the original multi-color ordering, and evaluate and clarify the effectiveness of block multi-color ordering.