Multicore Optimizations for Sparse Matrix Kernels
Richard Vuduc
Computational Science and Engineering Division, Georgia Tech
April 14, 2008
10:00 AM
5700, D307
Host: Jeff Vetter (vetter@ornl.gov)
ABSTRACT:
I will present recent work, joint with the Berkeley Benchmarking and OPtimization (BeBOP) group, on optimizing the performance of sparse matrix-vector multiplication for current multicore systems. I will also give a broad overview of the design and implementation of OSKI, an autotuned sparse BLAS-like library, emphasizing how its internal data structure representation supports optimizing transformations relevant to multicore.
This work is joint with Sam Williams, Lenny Oliker, John Shalf, James Demmel, and Katherine Yelick.
# # #