| 9d01615f | 01-Apr-2015 |
Fande Kong <fdkong.jd@gmail.com> |
Rewrite an implementation of matincreaseoverlap
The old version depends a gloabl variable C->rmap->N. When the
problem size is small, it is pretty well. But when the problem size is
large, for exampl
Rewrite an implementation of matincreaseoverlap
The old version depends a gloabl variable C->rmap->N. When the
problem size is small, it is pretty well. But when the problem size is
large, for example, 1 billion, the old version does not work since it
requires too much memory. In this new implementation, we remove this dependent,
memory allocation only depends on the local size of the problem. And we also
use PetscSF to exchange data instead of the use of the MPI since if Jed does any
improvements on SF we will gain without changing code.
When overlap=1, we possibly do not involve any communication.
We use ex40.c as a test code. The new implementation is off by default. Use -mat_overalp_k 1
to turn on.
make ex40
make runex40_2
show more ...
|