Dev Ojha
Dev Ojha
I'm not really sure its useful to know if you're doing a coset FFT vs regular FFT. Are you worried about knowledge of the extra ~5% of runtime? If so,...
Had a discussion with @Pratyush regarding an API for this. 1) Make Domain have two constructors, `new_subgroup(subgroup_size)` and `new_coset(coset_size, offset)`. 2) Remove cosetFFT / cosetIFFT methods 3) Add a convenience...
The statistical significance measurements are very important for micro-benches though. Agreed that they don't give that much at the larger benchmark sizes, and just cause delays
I have run into that problem as well, I'm not sure I'd classify the statefulness as a feature, since it does become a pain to track which change was benched...
A relevant data point is how FFTW does this. They have the caller create a "plan" for FFTs based on the domain size, which then AFAIU does precompilation and preprocessing...
Wow blstrs' fp12 multiplications are significantly faster! (The fp multiplication times are the same in both, but the fp12 multiplication in blstrs is half the time) One of the things...
I think Alex implemented these in C++ as part of his libff update, that could perhaps be a useful reference
Ah just checked, it was implementing Tonelli Shanks (I mistook it with a conversation we had about this exact paper though lol) Sorry for the confusion
These are known at compile time, so theres no run time test needed with an API change. (The coefficient is a constant)
Agreed that its probably not that notable on a second pass through, its called like 4 times for bls12