Rank k Cholesky Up/Down-dating on the GPU: gpucholmodV0.2

Abstract

In this note we briefly describe our Cholesky modification algorithm for streaming multiprocessor architectures. Our implementation is available in C++ with Matlab binding, using CUDA to utilise the graphics processing unit (GPU). Limited speed ups are possible due to the bandwidth bound nature of the problem. Furthermore, a complex dependency pattern must be obeyed, requiring multiple kernels to be launched. Nonetheless, this makes for an interesting problem, and our approach can reduce the computation time by a factor of around 7 for matrices of size 5000 by 5000 and k=16, in comparison with the LINPACK suite running on a CPU of comparable vintage. Much larger problems can be handled however due to the O(n) scaling in required GPU memory of our method.