Any workarounds for TAO in a complex scalar type code?
Hi, I have a complex valued optimization code that requires non-linear least squares curve fitting at each iteration. I wanted to use Tao to do the non-linear least squares problem, but just discovered that Tao can only be used with real scalar types. Is there any workarounds? I was wondering if maybe I could somehow use two different Petsc installations at once, or is there any other ways to hack tao into working? Thanks, Lucas
Hi Lucas, In theory, you can attempt to solve complex-valued problems by concatenating the real and imaginary parts into a single real-valued solution vector and solving it as if it’s a real-valued problem. There’s a more practical issue of the PETSc build excluding the TAO source code entirely for complex builds. I think this is done because we cannot guarantee that TAO code has been completely rigorous in its distinction of PetscReal from PetscScalar. There may be places where we used Scalar and assumed it’s Real. The methods developed in the last few years have paid more attention to this but there’s a lot of older code floating around that might break even if you hacked the build into compiling TAO for complex configurations. I’m not entirely sure what’s necessary to enable TAO builds for complex PETSc. Perhaps Todd, Satish or Barry will have a better idea. If you want to give it a try though I’d be happy to help you resolve the errors as they come up so that it can be compiled for complex builds. Can’t provide convergence guarantees but it would at least let you attempt a solution. Alp On Jun 24, 2021, at 2:44 PM, Lucas Banting <[email protected]<mailto:[email protected]>> wrote: Hi, I have a complex valued optimization code that requires non-linear least squares curve fitting at each iteration. I wanted to use Tao to do the non-linear least squares problem, but just discovered that Tao can only be used with real scalar types. Is there any workarounds? I was wondering if maybe I could somehow use two different Petsc installations at once, or is there any other ways to hack tao into working? Thanks, Lucas
On Thu, Jun 24, 2021 at 3:25 PM Dener, Alp via petsc-users < [email protected]> wrote:
Hi Lucas,
In theory, you can attempt to solve complex-valued problems by concatenating the real and imaginary parts into a single real-valued solution vector and solving it as if it’s a real-valued problem.
There’s a more practical issue of the PETSc build excluding the TAO source code entirely for complex builds. I think this is done because we cannot guarantee that TAO code has been completely rigorous in its distinction of PetscReal from PetscScalar. There may be places where we used Scalar and assumed it’s Real. The methods developed in the last few years have paid more attention to this but there’s a lot of older code floating around that might break even if you hacked the build into compiling TAO for complex configurations.
If an expression does not support PetscComplex, shouldn't it report a compilation error?
I’m not entirely sure what’s necessary to enable TAO builds for complex PETSc. Perhaps Todd, Satish or Barry will have a better idea. If you want to give it a try though I’d be happy to help you resolve the errors as they come up so that it can be compiled for complex builds. Can’t provide convergence guarantees but it would at least let you attempt a solution.
Alp
On Jun 24, 2021, at 2:44 PM, Lucas Banting <[email protected]> wrote:
Hi,
I have a complex valued optimization code that requires non-linear least squares curve fitting at each iteration. I wanted to use Tao to do the non-linear least squares problem, but just discovered that Tao can only be used with real scalar types. Is there any workarounds?
I was wondering if maybe I could somehow use two different Petsc installations at once, or is there any other ways to hack tao into working?
Thanks, Lucas
$ git grep requiresscalar bound/makefile:#requiresscalar real complementarity/makefile:#requiresscalar real constrained/makefile:#requiresscalar real leastsquares/makefile:#requiresscalar real linesearch/impls/makefile:#requiresscalar real pde_constrained/makefile:#requiresscalar real quadratic/makefile:#requiresscalar real unconstrained/makefile:#requiresscalar real We added these requirements when Tao was brought back into the PETSc fold as Alp mentioned. It would be good to start eliminating the requirements when possible. I have removed the leastsquares/makefile:#requiresscalar real and have started a build to see how it goes. If everything goes well maybe we can get it working for complex without too much difficulty. Barry
On Jun 24, 2021, at 4:02 PM, Junchao Zhang <[email protected]> wrote:
On Thu, Jun 24, 2021 at 3:25 PM Dener, Alp via petsc-users <[email protected] <mailto:[email protected]>> wrote: Hi Lucas,
In theory, you can attempt to solve complex-valued problems by concatenating the real and imaginary parts into a single real-valued solution vector and solving it as if it’s a real-valued problem.
There’s a more practical issue of the PETSc build excluding the TAO source code entirely for complex builds. I think this is done because we cannot guarantee that TAO code has been completely rigorous in its distinction of PetscReal from PetscScalar. There may be places where we used Scalar and assumed it’s Real. The methods developed in the last few years have paid more attention to this but there’s a lot of older code floating around that might break even if you hacked the build into compiling TAO for complex configurations. If an expression does not support PetscComplex, shouldn't it report a compilation error?
I’m not entirely sure what’s necessary to enable TAO builds for complex PETSc. Perhaps Todd, Satish or Barry will have a better idea. If you want to give it a try though I’d be happy to help you resolve the errors as they come up so that it can be compiled for complex builds. Can’t provide convergence guarantees but it would at least let you attempt a solution.
Alp
On Jun 24, 2021, at 2:44 PM, Lucas Banting <[email protected] <mailto:[email protected]>> wrote:
Hi,
I have a complex valued optimization code that requires non-linear least squares curve fitting at each iteration. I wanted to use Tao to do the non-linear least squares problem, but just discovered that Tao can only be used with real scalar types. Is there any workarounds?
I was wondering if maybe I could somehow use two different Petsc installations at once, or is there any other ways to hack tao into working?
Thanks, Lucas
Least squares has two supported types: brgn and pounders. They will not compile with complex. I suspect the problem, at least with brgn, is its handling of bound constraints. Perhaps for complex numbers the handling of constraints can simply be #ifdef out? If one does not need constraints. Barry You can start by removing the leastsquares/makefile:#requiresscalar real adding it into the pounders directory and then compiling and "fixing" the code in brng.c to work with complex. But BRNG is a wrapper around TAOBNTL so you need to get that to also build correctly with complex.
On Jun 24, 2021, at 10:11 PM, Barry Smith <[email protected]> wrote:
$ git grep requiresscalar bound/makefile:#requiresscalar real complementarity/makefile:#requiresscalar real constrained/makefile:#requiresscalar real leastsquares/makefile:#requiresscalar real linesearch/impls/makefile:#requiresscalar real pde_constrained/makefile:#requiresscalar real quadratic/makefile:#requiresscalar real unconstrained/makefile:#requiresscalar real
We added these requirements when Tao was brought back into the PETSc fold as Alp mentioned. It would be good to start eliminating the requirements when possible.
I have removed the leastsquares/makefile:#requiresscalar real and have started a build to see how it goes. If everything goes well maybe we can get it working for complex without too much difficulty.
Barry
On Jun 24, 2021, at 4:02 PM, Junchao Zhang <[email protected] <mailto:[email protected]>> wrote:
On Thu, Jun 24, 2021 at 3:25 PM Dener, Alp via petsc-users <[email protected] <mailto:[email protected]>> wrote: Hi Lucas,
In theory, you can attempt to solve complex-valued problems by concatenating the real and imaginary parts into a single real-valued solution vector and solving it as if it’s a real-valued problem.
There’s a more practical issue of the PETSc build excluding the TAO source code entirely for complex builds. I think this is done because we cannot guarantee that TAO code has been completely rigorous in its distinction of PetscReal from PetscScalar. There may be places where we used Scalar and assumed it’s Real. The methods developed in the last few years have paid more attention to this but there’s a lot of older code floating around that might break even if you hacked the build into compiling TAO for complex configurations. If an expression does not support PetscComplex, shouldn't it report a compilation error?
I’m not entirely sure what’s necessary to enable TAO builds for complex PETSc. Perhaps Todd, Satish or Barry will have a better idea. If you want to give it a try though I’d be happy to help you resolve the errors as they come up so that it can be compiled for complex builds. Can’t provide convergence guarantees but it would at least let you attempt a solution.
Alp
On Jun 24, 2021, at 2:44 PM, Lucas Banting <[email protected] <mailto:[email protected]>> wrote:
Hi,
I have a complex valued optimization code that requires non-linear least squares curve fitting at each iteration. I wanted to use Tao to do the non-linear least squares problem, but just discovered that Tao can only be used with real scalar types. Is there any workarounds?
I was wondering if maybe I could somehow use two different Petsc installations at once, or is there any other ways to hack tao into working?
Thanks, Lucas
participants (4)
-
Barry Smith -
Dener, Alp -
Junchao Zhang -
Lucas Banting