Support user-specified W_prototype <: AbstractSciMLOperator
#1996
Conversation
gaurav-arya
force-pushed
the
ag-wprot
branch
4 times, most recently
from
67ea0a4
to
36f4048
Compare
equivalent to before)
|
@ChrisRackauckas, @utkarsh530, @vpuri3 this is finally ready:) |
|
why do we need a type? why is it not just integrator.gamma = ScalarOp(0.0; update_func)
integrator.W = integrator.M - integrator.gamma * integrator.J |
|
You mean why not remove the |
|
LGTM. Yes, I believe |
|
@gaurav-arya can we resolve this here? |
test/interface/wprototype_tests.jl
Outdated
| # TODO: in future, ensure and test that polyalg chooses the best linear solve for all. | ||
| sol = solve(prob, Rosenbrock23(linsolve = KrylovJL_GMRES())) | ||
| sol_J = solve(prob_J, Rosenbrock23(linsolve = KrylovJL_GMRES())) # note: direct linsolve in this case is broken, see #1998 | ||
| sol_W = solve(prob_W, Rosenbrock23(linsolve = KrylovJL_GMRES())) |
There was a problem hiding this comment.
Test a newton solver as well, like FBDF. Also test Radau5
There was a problem hiding this comment.
Added an FBDF test. RadauIIA5 does not seem to support sparse Jacobians though:
OrdinaryDiffEq.jl/src/caches/firk_caches.jl
Line 238 in b807375
There was a problem hiding this comment.
Yeah Radau5 will need more work.
WOperatorto match that of a SciMLOperator.W_prototype <: AbstractSciMLOperator.Dependent on SciML/SciMLBase.jl#473 and SciML/StochasticDiffEq.jl#540. The intent is for this feature to be undocumented until it is more battle tested. Needs tests
(I have a much more involved set of changes that replaces the existing
WOperatorwith a lazy SciMLOperator. But I've come to the conclusion that this is not something to be done lightly, e.g. there aremuladdoptimizations that are not easy to replace with the lazy purely general approach. Instead, I'm just addingAbstractSciMLOperatorbranches to first allow the W-operator to be manually swapped out for anAbstractSciMLOperator, while keeping the originalWOperatoraround.)