> On Sep 28, 2020, at 5:32 PM, Sam Guo <[email protected]> wrote: > > Another idea I am pursuing is to prefix real and complex using some compiler > flags. Didn’t figure out how to do this cross platform for shared libs yet.
This is very hairy because there are some symbols that can be common and some that are not common. I could never figure out a rational general purpose approach. This is definitely something we need to support in any future refactorization https://gitlab.com/petsc/petsc/-/issues/643 <https://gitlab.com/petsc/petsc/-/issues/643> > > On Monday, September 28, 2020, Matthew Knepley <[email protected] > <mailto:[email protected]>> wrote: > On Mon, Sep 28, 2020 at 5:01 PM Sam Guo <[email protected] > <mailto:[email protected]>> wrote: > Hi Matt, > Since I use MUMPS as preconditioner, complex uses too much memory if my > input matrix is real. Ideally if I can compile real and complex into > different symbols (like MUMPS) , I can load both version without conflict. > > What I mean to say is that it would be great if it were as simple as using > two different symbols, but unfortunately the problem is more difficult. I was > trying to use > the example of templates. This would be a very intrusive change no matter > what technology you are using. > > So your main memory usage is from the MUMPS factorization, and you cannot > afford to double that usage? > > You could consider writing a version of AIJ that stores real entries, but > allows complex vector values. It would promote to complex for the row dot > product. > However, you would also have to do the same work for all the solves you do > with MUMPS. > > I think it would be much easier to just decompose your complex work into real > and imaginary parts and use PETSc with real scalars to compute them > separately. > Since you know your matrices have 0 imaginary part, this becomes very > straightforward. > > Thanks, > > Matt > > Thanks, > Sam > > On Mon, Sep 28, 2020 at 12:52 PM Matthew Knepley <[email protected] > <mailto:[email protected]>> wrote: > On Mon, Sep 28, 2020 at 3:43 PM Sam Guo <[email protected] > <mailto:[email protected]>> wrote: > Hi Stefano and PETSc dev team, > I want to try your suggestion to always load complex version of PETSc but > if my input matrix A is real, I want to create shell matrix to matrix-vector > and factorization using real only. > > I do not think that will work as you expect. I will try to explain below. > > I still need to understand how MatRealPart works. Does it just zero out > the image numerical values or does it delete the image memory? > > When we have complex values, we use the "complex" type to allocate and store > them. Thus you cannot talk about just the memory to store imaginary parts. > MatRealPart sets the imaginary parts of all the matrix elements to zero. > > If my input matrix A is real, how do I create a shell matrix to matrix > -vector multiplication y=A*x where A is real, PestcScalar = complex, x and y > are Vec? I notice there is a VecRealPart but it seems it just zeros the image > numerical values. It seems I still have to create a PetscReal pointer to > copy the real part of PetacScalar pointers like following. Can you comment on > it? > > What you suggest would mean rewriting the matrix multiplication algorithm by > hand after extracting the values. I am not sure if this > is really what you want to do. Is the matrix memory really your limiting > factor? Even if you tried to do this with templates, the memory > from temporaries would be very hard to control. > > Thanks, > > Matt > > Thanks, > Sam > > PetscScalar *px = nullptr; > VecGetArrayRead(x, &px); > PetscScalar *py = nullptr; > VecGetArray(y, &py); > int localSize = 0; > VecGetLocalSize(x, &localSize); > std::vector<PetasReal> realX(localSize); // I am using c++ to call PETSc > > //retrieve real part > for(int i = 0; i < localSize; i++) realX[i] = PetscRealPart(px[i]); > > // do real matrix-vector multiplication > // realY=A*realX > // here where realY is std::vector<PetscReal> > > //put real part back to py > for(int i = 0; i < localSize; i++) pv[i] = realY[i]; > VecRestoreArray(y,&py); > > On Tue, May 26, 2020 at 1:49 PM Sam Guo <[email protected] > <mailto:[email protected]>> wrote: > Thanks > > On Tuesday, May 26, 2020, Stefano Zampini <[email protected] > <mailto:[email protected]>> wrote: > All the solvers/matrices/vectors works for PetscScalar types (i.e. in your > case complex) > If you need to solve for the real part only, you can duplicate the matrix and > call MatRealPart to zero out the imaginary part. But the solve will always > run in the complex space > You should not be worried about doubling the memory for a matrix (i.e. real > and imaginary part) > > >> On May 26, 2020, at 11:28 PM, Sam Guo <[email protected] >> <mailto:[email protected]>> wrote: >> >> complex version is needed since matrix sometimes is real and sometimes is >> complex. I want to solve real matrix without allocating memory for imaginary >> part((except eigen pairs). >> >> On Tuesday, May 26, 2020, Zhang, Hong <[email protected] >> <mailto:[email protected]>> wrote: >> You can build PETSc with complex version, and declare some variables as >> 'PETSC_REAL'. >> Hong >> >> From: petsc-users <[email protected] >> <mailto:[email protected]>> on behalf of Sam Guo >> <[email protected] <mailto:[email protected]>> >> Sent: Tuesday, May 26, 2020 1:00 PM >> To: PETSc <[email protected] <mailto:[email protected]>> >> Subject: [petsc-users] using real and complex together >> >> Dear PETSc dev team, >> Can I use both real and complex versions together? >> >> Thanks, >> Sam > > > > -- > What most experimenters take for granted before they begin their experiments > is infinitely more interesting than any results to which their experiments > lead. > -- Norbert Wiener > > https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/> > > > -- > What most experimenters take for granted before they begin their experiments > is infinitely more interesting than any results to which their experiments > lead. > -- Norbert Wiener > > https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/>
