halo2_axiom/fft/baseline.rs
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128
//! This contains the baseline FFT implementation
use ff::Field;
use super::recursive::FFTData;
use crate::{
arithmetic::{self, log2_floor, FftGroup},
multicore,
};
/// Performs a radix-$2$ Fast-Fourier Transformation (FFT) on a vector of size
/// $n = 2^k$, when provided `log_n` = $k$ and an element of multiplicative
/// order $n$ called `omega` ($\omega$). The result is that the vector `a`, when
/// interpreted as the coefficients of a polynomial of degree $n - 1$, is
/// transformed into the evaluations of this polynomial at each of the $n$
/// distinct powers of $\omega$. This transformation is invertible by providing
/// $\omega^{-1}$ in place of $\omega$ and dividing each resulting field element
/// by $n$.
///
/// This will use multithreading if beneficial.
fn best_fft<Scalar: Field, G: FftGroup<Scalar>>(a: &mut [G], omega: Scalar, log_n: u32) {
let threads = multicore::current_num_threads();
let log_threads = log2_floor(threads);
let n = a.len();
assert_eq!(n, 1 << log_n);
for k in 0..n {
let rk = arithmetic::bitreverse(k, log_n as usize);
if k < rk {
a.swap(rk, k);
}
}
//let start = start_measure(format!("twiddles {} ({})", a.len(), threads), false);
// precompute twiddle factors
let twiddles: Vec<_> = (0..(n / 2))
.scan(Scalar::ONE, |w, _| {
let tw = *w;
*w *= ω
Some(tw)
})
.collect();
//stop_measure(start);
if log_n <= log_threads {
let mut chunk = 2_usize;
let mut twiddle_chunk = n / 2;
for _ in 0..log_n {
a.chunks_mut(chunk).for_each(|coeffs| {
let (left, right) = coeffs.split_at_mut(chunk / 2);
// case when twiddle factor is one
let (a, left) = left.split_at_mut(1);
let (b, right) = right.split_at_mut(1);
let t = b[0];
b[0] = a[0];
a[0] += &t;
b[0] -= &t;
left.iter_mut()
.zip(right.iter_mut())
.enumerate()
.for_each(|(i, (a, b))| {
let mut t = *b;
t *= &twiddles[(i + 1) * twiddle_chunk];
*b = *a;
*a += &t;
*b -= &t;
});
});
chunk *= 2;
twiddle_chunk /= 2;
}
} else {
recursive_butterfly_arithmetic(a, n, 1, &twiddles)
}
}
/// This perform recursive butterfly arithmetic
fn recursive_butterfly_arithmetic<Scalar: Field, G: FftGroup<Scalar>>(
a: &mut [G],
n: usize,
twiddle_chunk: usize,
twiddles: &[Scalar],
) {
if n == 2 {
let t = a[1];
a[1] = a[0];
a[0] += &t;
a[1] -= &t;
} else {
let (left, right) = a.split_at_mut(n / 2);
multicore::join(
|| recursive_butterfly_arithmetic(left, n / 2, twiddle_chunk * 2, twiddles),
|| recursive_butterfly_arithmetic(right, n / 2, twiddle_chunk * 2, twiddles),
);
// case when twiddle factor is one
let (a, left) = left.split_at_mut(1);
let (b, right) = right.split_at_mut(1);
let t = b[0];
b[0] = a[0];
a[0] += &t;
b[0] -= &t;
left.iter_mut()
.zip(right.iter_mut())
.enumerate()
.for_each(|(i, (a, b))| {
let mut t = *b;
t *= &twiddles[(i + 1) * twiddle_chunk];
*b = *a;
*a += &t;
*b -= &t;
});
}
}
/// Generic adaptor
pub fn fft<Scalar: Field, G: FftGroup<Scalar>>(
a: &mut [G],
omega: Scalar,
log_n: u32,
_data: &FFTData<Scalar>,
_inverse: bool,
) {
best_fft(a, omega, log_n)
}