Hi Gaetan, Thanks for pointing me to the example. I think I understand more now. It seems that setPythonContext should point to an object that performs the multi(x, y) operation, which takes x and modifies y inside the multi() function. What does setUp() do?
Also, can I assume that the implementation is Pythonic, such that the shell can operate on my own custom vector objects, rather than PETSc Vecs? Defining the multi function to do this would be easy, but I wanted to make sure that the KSP solver wouldn't complain for some reason. Thanks, Jonathan Tu On Jul 30, 2014, at 4:48 PM, Gaetan Kenway <[email protected]> wrote: > Hi Jonathan > > I've successfully used petsc shell matrices from Python. There is an example > of how to do it in demo/poisson2d/poisson2d.py. The example is a little > sparse, but it has all the important information. The most important is to > define the > > def mult(self, mat, x, y): > > routine inside of a Python class that is given in mat.setPythonContext() > command. > > Hope that helps. > > Gaetan > > > On Wed, Jul 30, 2014 at 7:29 PM, Jonathan Tu <[email protected]> wrote: > Hi, > > I would like to solve a large linear system. I have a linear operator that > operates on a "state" vector. The linear operator does things like take the > sum of the state elements, take the cross product of the state with a known > vector, and then multiply the state by a large (distributed) matrix. You > could imagine it like this: > > [sum of elements, cross product with other vector, multiplication with large > matrix] = A * x > > Now, I realize it is possible to implement all of this using a large matrix. > After all, these are all linear operations. However, I thought it might be > faster and less error-prone to use some of numpy's built-in operations, for > the sums and cross products, and only use a distributed matrix for the last > part of the output. > > From the PETSc documentation, it seems like a matrix shell would be perfect > for this. I could define my own mult operator that does all of these things > in steps, using built-in numpy functions when appropriate and direct matrix > multiplication when appropriate. However, I cannot figure out how to do this > in petsc4py. I found somewhere online that perhaps createPython is the way I > need to go about this, but I couldn't figure out the details. > > Any ideas? > > > > Jonathan Tu >
