Xabclib : 汎用的自動チューニングインターフェースOpenATLibを利用した反復解法ライブラリの開発(<特集>数値計算のための自動チューニング(続))

概要

論文の詳細を見る
Conventional numerical libraries with auto-tuning (AT) facility have no re-usability for their AT functions. To solve the re-usability problem of AT libraries, we have established Application Programming Interfaces (APIs) for AT. The library of the APIs is named OpenATLib. In addition, we have developed sparse iterative solvers using OpenATLib. The package is named Xabclib. We have also developed a new AT function, which is called "Numerical Policy" to specify optimization policy for AT from end-users. In the sparse matrix-vector multiplication part, we propose and implement a new parallelization method on thread processing level, named normalized non-zero elements method. The method balances the number of non-zero elements on rows, while the conventional method simply divides rows of the matrix. Due to the simple row division, the conventional method has load imbalance problem in some cases. The performance evaluation using one node (16 cores) of the T2K Open Supercomputer (U. Tokyo), we have established keeping good accuracy from end-user requirement for eigensolver and linear equations solver, respectively. In addition, we have obtained the factor of 2.8x speedup in a matrix compared to the execution of conventional method for sparse matrix-vector multiplication.
2010-12-24