June 22–26, 2014
Leipzig, Germany

Presentation Details

Name: (10a) Cache-Aware Sparse Matrix Format for GPU
Time: Thursday, June 26, 2014
10:30 am - 11:00 am
Room:   Hall 4
CCL - Congress Center Leipzig
Breaks:10:30 am - 11:00 am Coffee Break
07:30 am - 10:30 am Welcome Coffee
Presenter:   Yusuke Nagasaka, Tokyo Institute of Technology
Abstract:   We propose two novel sparse matrix formats, SJDS and NUSJDS, which are designed to improve GPU cache hit rate in SPMV by uniformly dividing matrix into sub-matrices by columns. SJDS achieves up to 10x performance gain compared to other existing formats. NUSJDS enables further speed-up by adaptively adjusting the sizes of sub-matrices.

Yusuke Nagasaka, Akira Nukada & Satoshi Matsuoka, Tokyo Institute of Technology