DocumentCode :
752241
Title :
Automatic partitioning of parallel loops with parallelepiped-shaped tiles
Author :
Rastello, F. ; Robert, Y.
Author_Institution :
ST Microelectron., Grenoble, France
Volume :
13
Issue :
5
fYear :
2002
fDate :
5/1/2002 12:00:00 AM
Firstpage :
460
Lastpage :
470
Abstract :
In this paper, an efficient algorithm to implement loop partitioning is introduced and evaluated. We start from results of Agarwal et al. (1995) whose aim is to minimize the number of accessed data throughout the computation of a tile; this number is called the cumulative footprint of the tile. We improve these results along several directions. First, we derive a new formulation of the cumulative footprint, allowing for an analytical solution of the optimization problem stated by Agarwal et al.. Second, we deal with arbitrary parallelepiped-shaped tiles, as opposed to rectangular tiles. We design an efficient heuristic to determine the optimal tile shape in this general setting and we show its usefulness using both examples of Agarwal et al. and a large collection of randomly generated data.
Keywords :
cache storage; program compilers; accessed data; automatic parallel loop partitioning; cumulative footprint; efficient algorithm; heuristic; optimization problem; parallelepiped-shaped tiles; randomly generated data; tile computation; Arithmetic; Computer science; Costs; Microelectronics; Optimization methods; Partitioning algorithms;
fLanguage :
English
Journal_Title :
Parallel and Distributed Systems, IEEE Transactions on
Publisher :
ieee
ISSN :
1045-9219
Type :
jour
DOI :
10.1109/TPDS.2002.1003856
Filename :
1003856
Link To Document :
بازگشت