Add FEEvaluationImplBasisChange::do_backward_hessians #10869
Conversation
peterrum
force-pushed
the
do_backward_hessians
branch
from
b216381
to
dc4e443
Compare
| do_backward_hessians( | ||
| const unsigned int n_components, | ||
| const dealii::AlignedVector<Number2> &transformation_matrix, | ||
| const bool add_into_result, | ||
| Number * values_in, | ||
| Number * values_out, | ||
| const unsigned int basis_size_1_variable = | ||
| dealii::numbers::invalid_unsigned_int, | ||
| const unsigned int basis_size_2_variable = | ||
| dealii::numbers::invalid_unsigned_int) |
There was a problem hiding this comment.
I must admit that I do not like to duplicate the do_backward function for the sake of bypassing the assumption of a zero somewhere in the values call in that function. We should make the interfaces uniform by just providing an apply function in the tensor product evaluators with an enum informing us about special cases, see also #7039.
There was a problem hiding this comment.
I agree that we should try not to introduce an additional specialization for hessian. But let's not remove the values/gradients/hessians-function calls now from all places (I can put it on my TODO list). I have introduced as suggested an enum (name is up to discussion) and use this now in the FEEvaluationImplBasisChange class. What do you think, @kronbichler?
There was a problem hiding this comment.
Absolutely, there is no need to clean things up now, neither need we remove the values/gradients/hessian calls now. But I was thinking of providing all evaluators with a static void apply function and uniform template parameters along the variants. Or, even better, there is a single code to run through the tensor products and just different 1D matrix-vector products.
peterrum
force-pushed
the
do_backward_hessians
branch
from
85cba06
to
a9a0e44
Compare
Extracted from #10830. Needed for applying the inverse matrix in the ECL context.