Let the gridder choose appropriate nu and nv on its own
The wgridder
interface currently contains the dimensions nu
and nv
of the Fourier space grid, although they are just an implementation detail and are not used in any data products passed to or returned from the gridder. This has the potential for user errors:
- the user specifies a combination of
epsilon
,nu
andnv
, for which no suitable kernel can be found. Increasingnu
andnv
would make the problem solvable. - the user specifies large
epsilon
and highnu/nv
; this usually leads to a gridding task which spends more time than necessary on FFTs.
In principle the gridder has all necessary information to determine a good oversampling factor from the user-provided data. It won't be easy to compute the absolute best possible choice, but we can probably get close (and may be better than user intuition).
@parras, @g-landmanbester, @g-simonperkins what do you think?