Look at the source for DMDAVecGetArray() it calls either VecGetArray1d, 
VecGetArray2d, VecGetArray3d. You can call the VecGetArrayNd directly setting 
the local start and stop you want.


Thanks, I'll look into this.

      As a PETSc developer, of course, I would recommend keeping your 
local/temporary data also in Vecs and using the DMDAVecGetArray() for access to 
those also and having all code written in the “local patch style” with loops 
i=gxs; i<gxs+gxm  I think the code is clearer and easier to reason about than 
having each process from 0 to vxm etc.

Yeah, I definitely see the attraction and I may ultimately go that route. As a non-PETSc developer however, it seems preferable to absolutely minimize the reliance on PETSc data management for code re-use in a non-PETSc application going forward. I'd be eager to hear from other devs on how they approach this.




On 03/06/2014 09:23 AM, Barry Smith wrote:
On Mar 6, 2014, at 8:05 AM, Mark Lohry <[email protected]> wrote:

I'm using DMDAs for managing my DOF data on structured grids, so processes have 
access to local array chunks from i=gxs; i<gxs+gxm. I have a lot of 
local/temporary data that I don't want PETSc to manage and I'm trying to keep PETSc 
as segregated from the rest of code as possible, so I'd prefer for any functions to 
be ignorant of particular ordering, and just access from i=0; i<gxm so that I'm 
using 0-indexed arrays locally. What's the appropriate way to go about doing this / 
getting direct access to the beginning of global-indexed local array?
    Mark,

    Look at the source for DMDAVecGetArray() it calls either VecGetArray1d, 
VecGetArray2d, VecGetArray3d. You can call the VecGetArrayNd directly setting 
the local start and stop you want.

    Barry

   As a PETSc developer, of course, I would recommend keeping your local/temporary 
data also in Vecs and using the DMDAVecGetArray() for access to those also and 
having all code written in the “local patch style” with loops i=gxs; i<gxs+gxm  
I think the code is clearer and easier to reason about than having each process 
from 0 to vxm etc.


Reply via email to