KLU sparse solver and SPGMR preconditioning for IDA solver #86
Description
I am trying to pass Jacobian as a sparse matrix through a jac function for the IDA solver. However, the code only take two types of linear solvers i.e. dense or SPGMR.
I understand it requires the interface for sparse KLU solver library in IDA which is not yet implemented in Assimulo. So, I proceeded with the SPGMR solver to pass a preconditioned matrix as per SUNDIALS documentation.
In the Assimulo documentation examples, I could find an example (ida_with_jac_spgmr.py) where jacv (Jacobian times vector) function is implemented. In this example, I figured there is no preconditioning used. Then looked at another example in CVODE (Explicit_Problem) where preconditioning is used (cvode_with_preconditioning.py).
However, trying out the same structure for IDA (Implicit_Problem) didn't work. The error was, the options for preconditioning were not available for IDA solver. I came to a conclusion about the behavior of this by looking into the source code of sundials.pyx.
---------------------------------Snippet of source code-------------------------------------------------------------------------
# From [sundials.pyx file](https://github.com/modelon-community/Assimulo/blob/master/src/solvers/sundials.pyx)---provides only two solvers-DENSE/SPGMR
if self.options["linear_solver"] == 'DENSE':
elif self.options["linear_solver"] == 'SPGMR':
# PREC_NONE implies no preconditioner option is supplied to SUNDIALS
IF SUNDIALS_VERSION >= (3,0,0):
#Create the linear solver
IF SUNDIALS_VERSION >= (4,0,0):
IF SUNDIALS_VERSION >= (6,0,0):
self.sun_linearsolver = SUNDIALS.SUNLinSol_SPGMR(self.yTemp, **PREC_NONE**, 0, ctx)
ELSE:
self.sun_linearsolver = SUNDIALS.SUNLinSol_SPGMR(self.yTemp, **PREC_NONE**, 0)
ELSE:
self.sun_linearsolver = SUNDIALS.SUNSPGMR(self.yTemp, PREC_NONE, 0)
#Attach it to IDAS
IF SUNDIALS_VERSION >= (4,0,0):
flag = SUNDIALS.IDASetLinearSolver(self.ida_mem, self.sun_linearsolver, NULL)
ELSE:
flag = SUNDIALS.IDASpilsSetLinearSolver(self.ida_mem, self.sun_linearsolver)
From the comments I added in the code snippet from the source code, I understood that the preconditioning functionality for the SPGMR solver is not exposed in sundials.pyx file.
Let me know if I my understanding is correct. If yes, is there a chance for this functionality to be exposed?
It would be really helpful if you could provide some information on this.