Development of Innovative Library for Singular Value Decomposition Suited to Multi-Core Processors
Project/Area Number |
20246027
|
Research Category |
Grant-in-Aid for Scientific Research (A)
|
Allocation Type | Single-year Grants |
Section | 一般 |
Research Field |
Engineering fundamentals
|
Research Institution | Kyoto University |
Principal Investigator |
|
Co-Investigator(Kenkyū-buntansha) |
TSUJIMOTO Satoshi 京都大学, 大学院・情報学研究科, 准教授 (60287977)
KIMURA Kinji 京都大学, 大学院・情報学研究科, 特定准教授 (10447899)
YAMAMOTO Yusaku 神戸大学, 大学院・システム情報学研究科, 教授 (20362288)
IWASAKI Masahi 京都府立大学, 大学院・生命環境科学研究科, 准教授 (30397575)
TAKATA Masami 奈良女子大学, 大学院・人間文化研究科, 助教 (20397574)
|
Project Period (FY) |
2008 – 2011
|
Project Status |
Completed (Fiscal Year 2011)
|
Budget Amount *help |
¥27,690,000 (Direct Cost: ¥21,300,000、Indirect Cost: ¥6,390,000)
Fiscal Year 2011: ¥7,410,000 (Direct Cost: ¥5,700,000、Indirect Cost: ¥1,710,000)
Fiscal Year 2010: ¥8,970,000 (Direct Cost: ¥6,900,000、Indirect Cost: ¥2,070,000)
Fiscal Year 2009: ¥11,310,000 (Direct Cost: ¥8,700,000、Indirect Cost: ¥2,610,000)
|
Keywords | 数理工学(数理的解析・計画・設計・最適化) / 特異値分解 / マルチコアプロセッサ / 特異値計算 / 原点シフト戦略 / GPGPU / 並列計算 / ライブラリ / クラウドコンピューティング / ライブラリー / 可積分系 / 計算アルゴリズム / ダルブ変換 / グラフィックスプロセッサ / 単精度GPU / 倍精度GPU / GPU / 原点シフト |
Research Abstract |
This project obtains the following fruits in parallel computations with multi-core CPU environments. 1) A shift strategy based on the generalized Newton shift is completed for computing bidiagonal singular values by the mdLVs. 2) A speed up is established in reorthogonalization of singular vectors by using the Householder transformation with the compact WY representation. 3) A speed up of preconditioning process from given dense matrix to bidiagonal is performed. 4) Evaluation both of performance and accuracy of some singular value decomposition algorithms including I-SVD and D & C in double precision computation in GPGPU are discussed. 5) New integrable algorithms for computing eigenvalues and singular values of some class of band matrices are designed. The resulting code DBDSLV of the fast I-SVD algorithm in terms of 1)-4) can be downloaded in the web page as well as many presentations have been given in international workshops and annual meetings and many research papers have been published.
|
Report
(6 results)
Research Products
(62 results)