[petsc-users] mixed precision methods with petsc

Barry Smith bsmith at mcs.anl.gov
Fri Jul 31 11:34:04 CDT 2015



   We tried adding some mixed precision support in PETSc a few years ago (what you found was a fossil of that testing). Unfortunately we have not come up with a good way to manage mixing precision in the same executable with the current PETSc design and implementation in C. So there is no support now and it is unlikely (without some brilliant idea from someone) that this will change.

   Barry

> On Jul 31, 2015, at 9:30 AM, LiRunhua <runhuali at hotmail.com> wrote:
> 
> I am curious about the mixed-precision capabilities in KSP. 
> What’s the effects with the option --with-mix-precision=1 when running ./configure?
> 
> Does it like this? The mix-precision CG solve, as an example, is that using single precision for some earlier iterations for accelerating, and using double precision for the rest iterations to keep high precision results.
> 
> Is there any other information about mix-precision methods, or for re-development if needed?
> 
> Thanks,
> Li Runhua



More information about the petsc-users mailing list