Generalized full sparse tiling of loop chains
Date
2013
Authors
Krieger, Christopher D., author
Strout, Michelle Mills, advisor
Böhm, Wim, committee member
Rajopadhye, Sanjay, committee member
Mueller, Jennifer, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
Computer and computational scientists are tackling increasingly large and complex problems and are seeking ways of improving the performance of their codes. The key issue faced is how to reach an effective balance between parallelism and locality. In trying to reach this balance, a problem commonly encountered is that of ascertaining the data dependences. Approaches that rely on automatic extraction of data dependences are frequently stymied by complications such as interprocedural and alias analysis. Placing the dependence analysis burden upon the programmer creates a significant barrier to adoption. In this work, we present a new programming abstraction, the loop chain, that specifies a series of loops and the data they access. Given this abstraction, a compiler, inspector, or runtime optimizer can avoid the computationally expensive process of formally determining data dependences, yet still determine beneficial and legal data and iteration reorderings. One optimization method that has been previously applied to irregular scientific codes is full sparse tiling. Full sparse tiling has been used to improve the performance of a handful of scientific codes, but in each case the technique had to be applied from scratch by an expert after careful manual analysis of the possible data dependence patterns. The full sparse tiling approach was extended and generalized as part of this work to apply to any code represented by the loop chain abstraction. Using only the abstraction, the generalized algorithm can produce a new data and iteration ordering as well as a parallel execution schedule. Insight into tuning a generalized full sparse tiled application was gained through a study of the different factors influencing tile count. This work lays the foundation for an efficient autotuning approach to optimizing tile count.
Description
Rights Access
Subject
compilers
run-time optimization
parallelization