On Friday, 26 February 2016 at 13:21:12 UTC, Guillaume Piolat
wrote:
We are not talking of the same thing. I was thinking about the
table frequency cutoff which is 2x lower every level of mipmap
Ok. I think is most common to use high levels of oversampling for
tables so one can get better SNR using cheap interpolation. But
I've lately thought a bit about encoding the complexity of a
segment of the function in the table so that the renderer can
choose an interpolation technique that matches the location in
the table (e.g. straight line, use lerp; high frequency wobble,
use sinc). To make it work for realtime one would have to track
the cost and revert to lower quality when the budget has been
spent.