Closed dfarns closed 3 years ago
It is just an experiment with estimated costs but applied to a larger space of plans. It probably produces slower plans in practice — there's certainly no guarantee that it is better, since it is still limited by the fact that it is not performing any timing measurements.
Thank you for the explanation.
I see this in the source as one of the "undocumented guru" flags, but can you please provide a description? Would this somehow guarantee a recurring plan for the same problem, as with FFTW_ESTIMATE, but choose a "better" plan and improve execute performance? Thanks!