This saves 4 point doublings by unwrapping the first loop iteration,
as well as 63 field multiplications (one per iteration) by managing
curve model choice explicitly.
Building the docs currently doesn't work, because rustdoc enables parts of the
code (to document them) which then don't check (because there are missing
exports or dependencies). This **should** fix the issue, but there's no way to
test without publishing a new version.
This doesn't (yet) give any speedup over the non-precomputed multiscalar
multiplication, and it's not clear that it's a good idea to commit to
supporting it in the future. Removing it means that it's not committed-to as
part of the public API, but the source is still there in the tree if we want to
revisit it later.
Since computation of the 0 term in reduction requires a multiplication with a
4-cycle latency, this ensures that the rest of the computation can start before
the 0 term is finished
This splits the `FieldElement51x4` type into two types:
- `F51x4Reduced` (with reduced limbs)
- `F51x4Unreduced` (with unreduced limbs)
The reduction is implemented as a `From` impl to convert one type to the other.
The output of a multiplication is now a `F51x4Unreduced`. The reason is that
the inputs to IFMA operations must be at most 52 bits, so it's not possible to
perform an addition of (51+epsilon)-bit values and still be small enough to be
used as an input to multiplication. So, it doesn't make sense to perform a
reduction at the end of a multiplication, because the reduced values will be
fed into an addition or subtraction, which then needs to be re-reduced.
This begins to attempt to restructure the source tree so that the common parts
are common and the different parts are different.
The backend is now split into two parts:
- serial (containing the implementation using serial formulas and mixed-model arithmetic).
- vector (containing the implementation using parallel formulas and single-model arithmetic).
The serial scalar_mul tree is now under backend::serial::scalar_mul.
The avx2 scalar_mul tree is now under backend::avx2::scalar_mul.
`FieldElement32` -> `FieldElement2625`
`FieldElement64` -> `FieldElement51`
`Scalar32` -> `Scalar29`
`Scalar64` -> `Scalar52`
This naming is more accurate and would let us add an ADX backend later.
This change provides a common convention for using allocator-dependent
features with:
#![cfg(feature = "alloc")]
When available, `Vec` is imported consistently as `prelude::Vec`, which
means modules that need access to `Vec` can simply do:
use prelude::*;
and if an allocator is available, `Vec` will be in the crate prelude.
This allows all `alloc` vs `std` gating to be handled in `lib.rs`,
`build.rs`, and `prelude.rs` so the rest of the codebase doesn't have to
do any gating whatsoever.
Unfortunately, Rust selects `i32` as the type for an integer literal
when the literal has no other type constraints. This means that someone
cannot write `Scalar::from(1)`, as Rust will choose `i32` as the type for
`1`, and we don't `impl From<i32> for Scalar`.
We could implement `From` conversions for signed integers, but since
`Scalar` operations should be constant-time by default, this would
require us to extract the sign bit of the integer and use it to
conditionally select between the positive and negative of Scalar
constructed from the value bits. This is more expensive than the
unsigned operation, and I don't think it's what anyone really wants.
Making API consumers specify that their literals are unsigned is
slightly annoying, but better than the above alternative.
It would also be nice to change `Scalar::from_hash` to be
`impl<D: Digest<OutputSize = U64>> From<D> for Scalar`,
but this isn't currently allowed by Rust (since that `impl` "could"
conflict with the `impl From<u8>` if someone decided that `u8` should
`impl Digest`).
Each backend can now be selected by an individual feature:
- `u32_backend` for `backend::u32`;
- `u64_backend` for `backend::u64`;
- `avx2_backend` for `backend::avx2`;
The `u64_backend` is selected by default, since most people use X64 and we have
no way to select based on target (see discussion in #126). However, these
changes mean that it is possible to select the backend explicitly, and if we
had the ability to select target-default features, we could do so easily.
Change Scalar::non_adjacent_form() to take a width parameter.
This rewrite also makes it faster, although it's probably a ways off
from optimal. I don't know how much it matters.
TODO: write up description of why this computes the same thing.
Thanks to @oleganza for pointing out an error reading bits across words
in an earlier version of this code.
This change required some work, because the to-be-stabilized SIMD functions
don't allow non-constant `imm8`s. Previously, the `stdsimd` functions had a
constifying macro that ensured that the immediates were known. The dalek code
used this to build helper functions which would be inlined into different
places where the immediates were known. Unfortunately, since constexprs aren't
fully supported in Rust yet, this is done by a hidden compiler attribute, and
there's no way to propagate "constness".
To deal with this, some of the functions are specialized (e.g.,
`square_and_negate_D` instead of taking a mask), and others use an enum.
The serial (`u32`/`u64`) implementations use a multiple curve models, passing
between extended and projective coordinates when performing addition and
doubling (respectively). But the AVX2 backend doesn't, so in order to write a
single scalar mult implementation, we have to either abstract over the curve
models or have two implementations.
A generic solution is possible but extremely unreadable: the scalar mul
implementation would be parameterized over the point types used by the serial
implementations, with many where clauses describing how the types relate. The
AVX2 types could then be substituted in the appropriate places.
Instead we just duplicate the code into the `avx2` backend.
This should contain generic implementations of scalar multiplication algorithms
that can be used with multiple backends. The goal is to move the existing
scalar multiplication code into this submodule, then call it from the
user-facing API. This can also contain code for things we can't do now, like
multiscalar multiplication with precomputation.
Since Criterion can only benchmark public API, these changes just drop
all internal benchmarks (e.g., benchmarks for field operations). But
those are usually microbenchmarks whose meaning is kind of questionable
anyways, so I don't think this is a big loss.
The `bench` feature disappears, since Criterion works on stable Rust.
