Fix #351

[prj-]

No description provided.

[cmd8]

Thanks for the quick turnaround. I am having problems with this patch. Looks like they can be reproduced with newton-2d-PETSc.edp.

[prj-]

I knew this wasn't that easy to fix, otherwise I would have already fixed it...

[prj-]

I see what is happening, your varf was "easy" because there was no integration. Here, there is some integration, so we need to cache the values from before the assembly, do the assembly, and override the assembly by the cached values.

[cmd8]

If that all seems like too much to bury beneath a simple += operator, maybe that operator should be deprecated and eventually removed?

[prj-]

Yes, because it’s indeed really not clear what on combined with varf should do (should it apply BC after the summation or before the summation but after the assembly?)
I’ll see how easy it is to remove it, because at the same time, it’s functioning alright with no on.

[cmd8]

To me, it feels like on() should be interpreted as a part of the variational form. That is, it should apply the BC after assembly, but before summation. Otherwise, any natural BCs will be "enforced" before summation while any essential BCs will be enforced afterwards.

[prj-]

Not my finest work, but it should work now.

[cmd8]

Thanks, almost there. One last consideration for the complex case. The discrepancy there is smaller, but not irrelevant.

mesh Th = square(10,10);
fespace Vh(Th, P1);
Vh<complex> u;

varf varform(du, v) = int2d(Th)(x*y*(x-1)*(y-1)*(dx(v) + dy(v) + 1i*v)) + on(1, du = 0);

u[] = 1;
complex[int] temp = varform(0, Vh, tgv = -1);
u[] += temp;
cout << u[].l2 << endl;

u[] = 1;
u[] += varform(0, Vh, tgv = -1);
cout << u[].l2 << endl;
// These should be equal?

[prj-]

Of course... my bad, should be alright now.

[cmd8]

Looks good! So this operator will remain a part of FF and I can safely use it in my code (without fear of deprecation)?

[prj-]

It can stay, but it comes at a cost of three vectors instead of a single one (for now). So the user-facing code is maybe clearer, but slightly more memory-demanding under the hood (of course, these are merely vectors, so it should be in no way a bottleneck for most codes).

[cmd8]

Got it. Thanks for your quick work. 5-star service 😊