We can make use of the libCEED partial assembly decomposition to accelerate error estimation quite a bit, at least to perform the interpolation of the smooth and coarse flux fields at quadrature points. This currently involves recomputing a lot of quantities we already have access to, including the element Jacobians and other geometry data.
This would also enable error estimation to take place fully on the GPU.
We can make use of the libCEED partial assembly decomposition to accelerate error estimation quite a bit, at least to perform the interpolation of the smooth and coarse flux fields at quadrature points. This currently involves recomputing a lot of quantities we already have access to, including the element Jacobians and other geometry data.
This would also enable error estimation to take place fully on the GPU.