Non-linear solve crashes with PETSc error 63, but linear-solver works

[Stephan Schmidt]

Dear all,

There seems to be an odd behavioral difference between the linear and the non-linear solver.

If I manually (or using the derivative-command) linearize the form and feed it to the linear solver, then everything works fine.

If I pass the very same non-linear problem into the solve call to be automatically linearized and solved, then the solver call crashes with PETSc Error 63...

[Prof Garth Wells]

Looks to me like you're assembling only over part of the space for the nonlinear problem. I think you'll have zero blocks in your matrix.

[Stephan Schmidt]

Sure, that's possible. But even if you change the form a to

a = inner(du, u)*dx + inner(dp,p)*dx

in the above example, the behavior is exactly the same

[Prof Garth Wells]

You've hit the dreaded ufl.split vs dolfin.split (see Issue #194). Change

(u,p) = q.split()

to

(u,p) = split(q)

[Stephan Schmidt]

Thanks a lot! That has resolved the issue... Very interesting to see the two different effects of this wrt. the two solver calls...