[petsc-users] oddity in MatTranspose() under petsc4py

[Ethan Coon]

Nevermind... the issue is in numpy being smarter about limiting copies
than it used to be, not petsc4py.  Forcing numpy to copy the array after
the transpose fixes the problem.

Which brings up the question as to whether petsc4py should be smart
enough to check vals.flags['OWNDATA'] and make a copy if not true before
it grabs the underlying c-array?

Ethan

On Fri, 2013-01-18 at 10:55 -0700, Ethan Coon wrote:
> First, apologies for not looking into this more closely, but I wanted to
> make sure I wasn't missing something stupid before I rebuilt petsc-dev
> and petsc4py-dev and started writing PETSc-only versions to see what is
> going on.
> 
> See the following odd result:
> 
> 
> In [1]: from petsc4py import PETSc
> 
> In [2]: import numpy as np
> 
> In [3]: n = [0,1]
> 
> In [4]: vals = np.array([[0,1],[0,0]], 'd')
> 
> In [5]: A = PETSc.Mat().createAIJ(2,2)
> 
> In [6]: A.assemble()
> 
> In [7]: B = A.duplicate()
> 
> In [8]: B.setValues(n,n,vals, PETSc.InsertMode.ADD_VALUES)
> 
> In [9]: B.assemble()
> 
> In [10]: B.transpose(out=A)
> Out[10]: <petsc4py.PETSc.Mat at 0x2afb7e0>
> 
> In [11]: A.view()
> Matrix Object: 1 MPI processes
>   type: seqbaij
> row 0: (0, 0)  (1, 0) 
> row 1: (0, 1)  (1, 0) 
> 
> In [12]: vals
> Out[12]: 
> array([[ 0.,  1.],
>        [ 0.,  0.]])
> 
> In [13]: A.setValues(n,n,-vals.transpose(), PETSc.InsertMode.ADD_VALUES)
> 
> In [14]: A.assemble()
> 
> In [15]: A.view()
> Matrix Object: 1 MPI processes
>   type: seqbaij
> row 0: (0, 0)  (1, -1) 
> row 1: (0, 1)  (1, 0) 
> 
> where I would have expected A to be all zeros...
> 
> This is on petsc4py-3.3 and petsc-3.3-p5, and I know this code worked on
> older versions of petsc4py/petsc.  If I'm not missing something dumb
> I'll look a little closer into whether it is PETSc or petsc4py...
> 
> Thanks,
> 
> Ethan
> 
> 
> 

-- 
------------------------------------
Ethan Coon
Post-Doctoral Researcher
Applied Mathematics - T-5
Los Alamos National Laboratory
505-665-8289

http://www.ldeo.columbia.edu/~ecoon/
------------------------------------