Hello everyone…I am starting to implement the cyclic reduction scheme to solve tri-diagonal systems…has anyone done something similar?Any suggestion perhaps?
I did it, about year and half ago - I’d say it is the right approach for CUDA-based tri-diagonal solver, and if implemented with care you could reach great performance boost.
There was a good paper given a few weeks ago at the ACM Principles and Practice of Parallel Programming conference held in Bangalore on parallel tridiagonal solvers using CUDA, including parallel cyclic reduction methods IIRC. You might want to track down a copy.
Interesting read (the paper is here)… The approach and results match my experience.
Thanks for the link :rolleyes: actually it seems to be a very good paper…
I found that paper a while ago, but I’ve been scratching my head trying to write decent code for it, to no avail :(
Turned to an explicit solver I coded which works great, though has severe mathematical stability issues.