109
reads

Lung-Sheng Chien
2014-08-15  11:00 - 12:00
Room 430, Astronomy and Mathematics Building

GPU has shown its power on dense linear algebra, for example, HPL (high performance LINPACK) for TOP500. However GPU is still not widely used in sparse direct solver because of sequential nature of the algorithm. In this talk, I will introduce a technique, called domino scheme, which can avoid global synchronization. Such technique can be used in sparse triangular solve, incomplete Cholesky factorization or incomplete LU factorization. Also it can be an alternative of supernodal or multifrontal approach in sparse direct solver. During this talk, you will learn how to trace a data dependence graph without global synchronization.