Hello,
this is a feature request.
Adding Mixed Precision Iterative Refinement (single or even better half
precision based) to the cusp krylov solver in order to leverage the single or
half precision hardware performance for double precision simulations should
result in a speedup of 3 to 4.5 for larger sparse matrix problems. This would
be of great benefit to my CFD research as well as that of others working e.g.
on energy efficiency projects (wind power...).
Mixed precision has been implemented e.g. in the QUDA library for QCD on GPUs
(see attached files), the magma library or a mixed precision GPU Krylov solver
for lattice QFD (Boston University).
I attached a couple of related papers that might be of interest.
I am happy to support related testing, documentation... activities as required.
Kind regards,
Klaus
Original issue reported on code.google.com by k_burk...@yahoo.com on 16 Jun 2012 at 7:12
Original issue reported on code.google.com by
k_burk...@yahoo.comon 16 Jun 2012 at 7:12Attachments: