Hi,
what is the reason behind having gradient fusion off when antialias is on? Performance reason? If so what in particular?
Best
Gradient fusion is where we untransform the gradients inside the same kernel that computes the flux.
This paradigm does not make sense when flux anti-aliasing is enabled, however. There, the flux is evaluated at the quadrature points whereas we obtain the transformed gradients at the solution points. Thus, fusion is not viable. Instead, we use a separate kernel to untransform the gradients at the solution points which are then interpolated to the quadrature points so we can evaluate the flux. Hence, at the time we evaluate the flux there is nothing to fuse.
Regards, Freddie.
2 Likes