Introduce non-prime finite fields in fp

[JoeyBF]

This adds support for prime-power fields of order <2^16. We construct the multiplication table for the field F_(p^d) by identifying it with F_p[a]/P, where P is the appropriate Conway polynomial. Since there aren't that many fields of that order that aren't prime, I just hardcoded the Conway polynomials in a separate file.

The FqVectorP struct (previously FpVectorP) takes in a generic field instead of a prime, and all the managing of the internals of a limb is delegated to the underlying field.

Note that right now we only have a monomorphized FpVector, where the field is prime. This allows us to use u32s in our API. I suspect that the better move would be to have a dedicated FpElement struct that would be a wrapper around a u32, but would ensure that it belongs in the right range at compile time. I have started experimenting with this, but it gets really unwieldy to constantly create field elements instead of just passing regular numbers (in the algebra crate for example).

Implementing larger fields would not be impossible, but it would require more mucking around with Conway polynomials (that are only known for finitely many qs), and switching to a different representation for storing field elements and multiplication tables.

[JoeyBF]

@hoodmane do you want more time to review this or are we good to merge?

[hoodmane]

I'm going to try to do a more detailed review.

[JoeyBF]

No worries, I know it's a substantial change

[JoeyBF]

One question I'm still pondering: it is better to implement arithmetic operations on the fields or the elements?

On one hand, implementing them on fields means that the elements themselves can be very lightweight, since the field will be responsible for knowing how to handle them. However, we're now required to keep a copy of the field around, and always write, say, fq.add(a,b) instead of just a + b. On the other hand, implementing the operations on elements requires them to keep track of their field, but is much better for ergonomics.

I'm thinking the second approach is better. It solves the problem of ergonomics, which will make it simpler to just treat elements as things that we can manipulate algebraically. It will require us to use something like FpElement instead of u32 for Fp::Element, which might require some significant changes to the code, that will also allow us to ensure that an element is reduced mod p at the type level.

It might seem drastic to keep a copy of the field for every single element, but P2, P3, P5, and P7 are all ZSTs, and Fp<P> is just a thin wrapper, so there's no memory overhead for those primes. If odd-primes is disabled, even ValidPrime is a ZST. At larger primes, I think it's reasonable to take the hit on memory usage, since we're not as worried about performance as for the smaller primes.

[hoodmane]

I agree it's nice to be able to write a + b. If I understand correctly you're thinking of something like:

FieldElement<FieldType> {
   field: FieldType
   data: FieldType::ElementData
}

And then your point is for most of the fields we are calculating, field is zero sized so there's no overhead just a compile-time tag. For the fields that aren't zero-sized it doubles the amount of stuff we carry around, but they are already slow anyways?

[JoeyBF]

Yeah that's the idea. Basically tagging every element with its field. So SmallFq::Element would still be SmallFqElement, but the latter would go from just being a bare Option<u32> to

struct SmallFqElement<P> {
    field: SmallFq<P>,
    data: Option<u32>,
}

which would double the memory usage. That would also mean that we would change Fp::Element from u32 to some struct FpElement with the shape

struct FpElement<P> {
    field: Fp<P>,
    data: u32
}

but as you say the extra field attribute would be zero-sized in the cases where we care most about performance.

I'm not sure if we want to change our FpVector API to take in FpElements and have the caller handle the scalars, or just convert from u32s internally. I'll have to start implementing the changes and see what works best.

[JoeyBF]

Closing in favor of #159