Time q/p correlation blows up propagation performance.
As discussed on Mattermost, this change
D(3, 7) =
h * state.options.mass * state.options.mass *
stepper.charge(state.stepping) /
(stepper.momentum(state.stepping) *
std::hypot(1., state.options.mass / stepper.momentum(state.stepping)));
from !661 (merged) seems to obliterate propagation performance, at least the runtime of the fitting example increases by about a factor of 5.
It seems like this is an effect of the assignment of a non-zero value to D(3, 7)
rather than the calculation happening here.
@fklimpel can you follow up on this?