8 Feb
2015
8 Feb
'15
4:07 p.m.
On 07/02/2015 07:00, Mudit Sharma wrote:
After creating QR,LU and Cholesky for Boost..can we develop algorithms for determining which of QR,LU,Cholesky..etc. should be used in solving linear equations based on analyzing the running time and numerical stability specific to the problem. Is it feasible? Thanks!
I know of several teams in academia where this is a research topic. One of Joel Falcou's PhD students, Ian Masliah, has worked on this, you can try discussing it with him. One of his goals is also to choose the best pivoting method or detect whether it's a good idea to use mixed-precision methods or to offload on the GPU.