Small questions on ducc0 benchmarks #133
Description
Hi,
just stumbled across this project by accident ... great to see you find ducc0 useful!
And I must say this is the first time that someone got significant better performance from ducc0 SHTs in comparison with shtns! In my own experiments, they are roughly equal with shtns having a slight edge in some situations, especially in smaller transforms. But I definitely won't complain :-)
The accuracy benchmarks are the most interesting for me, however. Do you have a hypothesis about the reason why ducc0 is sometimes doing worse than the other implementations? Could it be a matter of quadrature weights, or is the geometry particularly weird?
If it happens on grids with decreasing numbers of points towards the poles, then I may have an idea where the discrepancy comes from.
In any case I would be very happy if you could point me to a concrete test script!
Best regards,
Martin