Automatic Performance Tuning of Matrix Computations Based on Hierarchical Performance Models(<Special Topics>Auto-Tuning for Numerical Computations)

DOI Web Site 36 References
  • Yamamoto Yusaku
    神戸大学大学院システム情報学研究科計算科学専攻
  • Fukaya Takeshi
    名古屋大学大学院工学研究科計算理工学専攻博士後期課程

Bibliographic Information

Other Title
  • 階層的な性能モデルに基づく行列計算の自動チューニング(<特集>数値計算のための自動チューニング)
  • 階層的な性能モデルに基づく行列計算の自動チューニング
  • カイソウテキ ナ セイノウ モデル ニ モトズク ギョウレツ ケイサン ノ ジドウ チューニング

Search this article

Abstract

Linear algebra programs such as linear equation solvers and eigenvalue solvers have an inherent hierarchical structure; they are composed of calls to lower level computational routines such as inner product or matrixvector multiplication. Hence, one can predict their execution time by modeling the execution times of the lower level routines and accumulating them. Such hierarchical performance models can be used effectively for automatic performance tuning of linear algebra programs. In this paper, we survey recent studies along this line. In particular, we review several approaches for modeling the execution time of lower level routines, examine the conditions for the hierarchical tuning technique to be effective, and give an instance of linear algebra computation for which the technique is successful. We also discuss directions for future research, such as extension of the technique to deal with heterogeneous computing platforms, combination with algorithm generation techniques and incorporation into auto-tuning languages.

Journal

References(36)*help

See more

Details 詳細情報について

Report a problem

Back to top