Mixed Tensor/NonTensor for Gen

[jeremylt]

Last step in the *gen refactor. This will allow us to run operators that have a mix of tensor product and non-tensor bases. 2D is easier, 3D will take more thought.

[jeremylt]

Sample "flat" version

template <int P_1D, int Q_1D>
inline __device__ void WeightTensor2dFlattened(SharedData_Cuda &data, const CeedScalar *__restrict__ q_weight_1d, CeedScalar *w) {
  const int max = P_1D < Q_1D ? P_1D : Q_1D;

  WeightTensor2d_Core<Q_1D>(data, data.t_id_x % max, data.t_id_x / max, q_weight_1d, w);
}

data.t_id_x goes up to max(P_1D * P_1D, Q_1D * Q_1D) here, so to map back to the old version we need to pass t_id_x % max to use as the tensor t_id_x and t_id_y / max to use as the tensor t_id_y.

[jeremylt]

Ok, I need to check on the memory in the data.slice for this approach

Also.... I'm not sure how 3D will be tackled

[jeremylt]

Ok, there's going to need to be more disentangling, as the operator I'm trying to target has different dim for different nodal spaces

[jeremylt]

Ok, separate dims for each field now, but theres some bug that's giving wrong results

[jeremylt]

ugh, T_1D is wrong for this strategy. Pondering

Idea - move slice or stand up fully separate versions? Probably the fully separate is the way to go at this point

[jeremylt]

Getting closer. 2D Tensor + 3D NonTensor is working now. Need 3D Tensor + 3D NonTensor next.

[jeremylt]

I'm starting to think this was not worth any of the time spent here - I think the restrictions also need to be hacked around with. I'm going to abandon this effort for now and switch to AtPoints assembly

[jeremylt]

Holy crap, we've got non-tensor with 2D tensor working perfectly now. Just need non-tensor with 3D tensor.

[jeremylt]

Ok, ready for review. Plan is squash + merge