On Fri, Mar 1, 2024 at 10:28 AM Fabian Wermelinger <fab4...@posteo.ch> wrote:
> Dear All, I am implementing a linear solver interface in a flow solver > with support for PETSc. My application uses a parallel CSR representation > and it manages the memory for it. I would like to wrap PETSc matrices (and > vectors) around it such > ZjQcmQRYFpfptBannerStart > This Message Is From an External Sender > This message came from outside your organization. > > ZjQcmQRYFpfptBannerEnd > > Dear All, > > I am implementing a linear solver interface in a flow solver with support for > PETSc. My application uses a parallel CSR representation and it manages the > memory for it. I would like to wrap PETSc matrices (and vectors) around it > such > that I can use the PETSc solvers as well. I plan to use > MatMPIBAIJSetPreallocationCSR and VecCreateMPIWithArray for lightweight > wrapping. The matrix structure is static over the course of iterations. I am > using a derived context class to host the PETSc related context. This context > holds references to the PETSc matrix and vectors and KSP/PC required to call > the > solver API later in the iteration loop. I would like to create as much as > possible during creation of the context at the beginning of iterations (the > context will live through iterations). > > My understanding is that MatMPIBAIJSetPreallocationCSR and > VecCreateMPIWithArray > DO NOT copy such that I can wrap the PETSc types around the memory managed by > the hosting linear solver framework in the application. The system matrix and > RHS (the pointers to these arrays are passed to MatMPIBAIJSetPreallocationCSR > and VecCreateMPIWithArray, respectively) is assembled by the application > before > any call to a linear solver. > > Given this setting: for every iteration, my plan is the PETSc information from > the context (Mat, Vec, KSP) and simply call KSPSolve without any other PETSc > calls (still assuming the matrix structure is static during iteration). > > What is not clear to me: > > Are there any MatSetValues/VecSetValues calls followed by > MatAssembly/VecAssembly(Begin/End) calls required for this setting? > > I don't believe so. You are doing the assembly. If you want to tell PETSc (KSP) that you changed the matrix so that it will re-setup the solvers (eg, refactor) then you can call KSPSetOperators. (You might want to increment the state of the Mat so that other things like TS know the state has changed, maybe that is all that you need to do) > The data in > the arrays for which pointers have been passed to > MatMPIBAIJSetPreallocationCSR > and VecCreateMPIWithArray is computed prior to any solver call in an > iteration, > such that I am assuming no additional "set value" calls through PETSc are > required -> am I missing something important by assuming this? > > That sounds fine, but keep in mind that you need to use our data layout for the blocks if you use our MatVec, etc. Thanks, Mark > Thank you for taking the time! > > -- > fabs > >
