On Fri, Nov 22, 2013 at 12:24 PM, Jed Brown <[email protected]> wrote:
> Michael Povolotskyi <[email protected]> writes: > > > This sounds very reasonable for me and my colleagues to have > > MatSetTransposeType(A,MAT_TRANSPOSE_HERMITIAN). > > Our problems always require Hermitian transpose but I assume there are > > problems where it is not the case. > > Complex symmetric shows up in frequency-domain wave problems. Methods > like Sou-Cheng's CS-MINRES take advantage of that structure. > > > Actually my goal is to solve the transposed problem with PARDISO > > package. It looks like it is the fastest option for the problems we > > usually solve. > > > > By the way, it there a chance to have interface to Pardiso in the next > > PETSc release? My colleague Jose David Bermeo sent the modules for this > > some time ago for your approval. > > I was buried in a proposal deadline and then SC13, but I will go through > the pull requests that nobody has gotten to. I think the Pardiso > contribution is more-or-less ready to merge. For MatMatSolveTranspose, > I would recommend doing it in a new branch starting from 'master'. > Please include a reference implementation so that we can test without > Pardiso (which is not free software). The code should be similar to > MatMatSolve_SeqAIJ. Once that works (and has a test), you'll merge the > Pardiso branch and add MatMatSolveTranspose_Pardiso. > By analogy with MatTransposeMatMult() etc., should this be MatTransposeMatSolve()? Since 'Transpose' applies to the first Mat. Dmitry.
