| 9d4fc724 | 28-Aug-2019 |
Lawrence Mitchell <lawrence@wence.uk> |
Patch: Use work vectors for patch solves
Rather than making vectors for each patch, just make one big enough
for the largest patch and reuse. Significantly increases performance
for small patch size
Patch: Use work vectors for patch solves
Rather than making vectors for each patch, just make one big enough
for the largest patch and reuse. Significantly increases performance
for small patch sizes.
In addition, remove log events from inner loop (which also costs a not
insignificant amount of time).
show more ...
|
| bc960bbf | 14-Jan-2019 |
Jed Brown <jed@jedbrown.org> |
PCBDDC: add lumped variant, aka M_1
This variant, also known as M_1 in the literature, skips extension of
balanced interface values to the interior by solving a Dirichlet
problem. It gives up a fac
PCBDDC: add lumped variant, aka M_1
This variant, also known as M_1 in the literature, skips extension of
balanced interface values to the interior by solving a Dirichlet
problem. It gives up a factor of H/h in condition number relative to
Dirichlet (standard) BDDC, but may be appropriate for small
subdomains, inexact subdomain solvers, and/or as an ingredient in
composite preconditioners.
Based-on-patch-by: Stefano Zampini <stefano.zampini@gmail.com>
show more ...
|
| daa275bd | 25-Sep-2019 |
Satish Balay <balay@mcs.anl.gov> |
Merge branch 'mark/gamg-optimizations-cpupinning-logicfixes-reducedgridlayout' into 'master'
high level GAMG optimizations, pin coarse grids to CPU, add new reduced...
See merge request petsc/petsc
Merge branch 'mark/gamg-optimizations-cpupinning-logicfixes-reducedgridlayout' into 'master'
high level GAMG optimizations, pin coarse grids to CPU, add new reduced...
See merge request petsc/petsc!2061
show more ...
|