MaTAXPY with same nonzero pattern is much faster.
> On Jan 30, 2018, at 10:10 PM, Marius Buerkle <mbuer...@web.de> wrote: > > Thanks. While B'*C*B may not have the same nonzero pattern as A it does not > matter as this is done by MaTAXPY with different nonzero pattern it should > not matter when using MatPtAPNumeric_MPIAIJ_MPIAIJ_scalable() with a > restricted reduction. I will give it a try. > > > > > To do this most efficiently you would compute exactly the part of F = B'*C*B > you need (that matches the nonzero pattern of A) do a MatAXPY() with same > numerical pattern and then just stick the values of D into the original > matrix (without worrying about excluding the new nonzeros since there won't > be any). > > MatPtAPNumeric_MPIAIJ_MPIAIJ_scalable() does the numerical triple product so > the only special code you need is to rewrite > MatPtAPSymbolic_MPIAIJ_MPIAIJ_scalable() so that that it takes another matrix > (your A) that determines the nonzero pattern you want and from this > constructs the needed symbolic information needed to do the (special, > reduced) matrix triple product. > > If you can this great, if not then I suspect there is not much you can do to > improve the efficiency of what you have already done. > > Barry > > > > On Jan 30, 2018, at 9:15 PM, Marius Buerkle <mbuer...@web.de> wrote: > > > > The matrix A is the extracted submatrix and I am basicallly doing something > > like this D=A+B'*C*B. Where A, B, and C are sparse but B'*C*C will clearly > > generate some addional nonzero entries in D. And D should go back into the > > bigger matrix but only the entries which are already nonzero in A. > > > > > > Could you tell us exactly what matrix matrix products you are doing? > > > > Barry > > > > > > > On Jan 30, 2018, at 6:53 PM, Marius Buerkle <mbuer...@web.de> wrote: > > > > > > Barry, > > > > > > Thanks for you reply. The pulled matrix is symmetric but that's it. At > > > the moment I am doing a copy of the matrix right after MatCreateSubMatrix > > > to keep it's nonzero structure. To insert the matrix which I obtained > > > after the multiplications back into the bigger matrix I use the initially > > > copied Matrix to get the nonzero elements with MatGetRow and as I know > > > the offset of the SubMatrix relative to the elements of the big-one I can > > > use MatSetValues to insert the elements row by row. That kinda works but > > > it is rather inefficient. Is there anyway to avoid copying the whole > > > submatrix in the beginning, as I don't need the actual values but only > > > the positions of the nonzero elements in one way or the other. > > > > > > Marius > > > > > > > > > > > >> Marius, > > >> > > >> We don't provide a way to insert a "generic" sparse matrix into a bigger > > >> matrices (dense matrices coming from element stiffness matrices yes) so > > >> I don't see any simple solution. >Does the submatrix you pull out have > > >> any particular structure, what does it represent? > > >> > > >> Barry > > > > > > > > >>> > > >>> Hi ! > > >>> > > >>> I have the followng problem. I create a Submatrix containing a subset > > >>> of row/columns of the original matrix. After some matrix > > >>> multiplications the non-zero strucutre of the resulting matrix changed. > > >>> Now I want to insert this submatix back into the original one keeping > > >>> only the non-zero entries which are present in the original matrix > > >>> discarding the others which accumulated due to the matrix > > >>> multiplications. Is there an easy or also not so easy way to do this ? > > >>> > > >>> Best, > > >>> Marius > > > > > >
