328: Optimizing BigUint and Bigint multiplication with the Toom-3 algorithm r=cuviper a=kompass

Hi !

I finally implemented the Toom-3 algorithm ! I first tried to minimize the memory allocations by allocating the `Vec<BigDigit>` myself, as was done for Toom-2, but Toom-3 needs more complex calculations, with negative numbers. So I gave up this method, to use `BigInt` directly, and it's already faster ! I also chose a better threshold for the Toom-2 algorithm.

Before any modification :
```
running 4 tests
test multiply_0        ... bench:         257 ns/iter (+/- 25)
test multiply_1        ... bench:      30,240 ns/iter (+/- 1,651)
test multiply_2        ... bench:   2,752,360 ns/iter (+/- 52,102)
test multiply_3        ... bench:  11,618,575 ns/iter (+/- 266,286)
```

With a better Toom-2 threshold (16 instead of 4) :
```
running 4 tests
test multiply_0        ... bench:         130 ns/iter (+/- 8)
test multiply_1        ... bench:      19,772 ns/iter (+/- 1,083)
test multiply_2        ... bench:   1,340,644 ns/iter (+/- 17,987)
test multiply_3        ... bench:   7,302,854 ns/iter (+/- 82,060)
```

With the Toom-3 algorithm (with a threshold of 300):
```
running 4 tests
test multiply_0        ... bench:         123 ns/iter (+/- 3)
test multiply_1        ... bench:      19,689 ns/iter (+/- 837)
test multiply_2        ... bench:   1,189,589 ns/iter (+/- 29,101)
test multiply_3        ... bench:   3,014,225 ns/iter (+/- 61,222)
```

I think this could be optimized, but it's a first step !
This commit is contained in:
bors[bot] 2017-09-20 20:53:40 +00:00
commit 4896746fec
2 changed files with 74 additions and 13 deletions

View File

@ -78,6 +78,11 @@ fn multiply_2(b: &mut Bencher) {
multiply_bench(b, 1 << 16, 1 << 16);
}
#[bench]
fn multiply_3(b: &mut Bencher) {
multiply_bench(b, 1 << 16, 1 << 17);
}
#[bench]
fn divide_0(b: &mut Bencher) {
divide_bench(b, 1 << 8, 1 << 6);

View File

@ -8,6 +8,7 @@ use traits::{Zero, One};
use biguint::BigUint;
use bigint::BigInt;
use bigint::Sign;
use bigint::Sign::{Minus, NoSign, Plus};
@ -225,20 +226,18 @@ fn mac_digit(acc: &mut [BigDigit], b: &[BigDigit], c: BigDigit) {
return;
}
let mut b_iter = b.iter();
let mut carry = 0;
let (a_lo, a_hi) = acc.split_at_mut(b.len());
for ai in acc.iter_mut() {
if let Some(bi) = b_iter.next() {
*ai = mac_with_carry(*ai, *bi, c, &mut carry);
} else if carry != 0 {
*ai = mac_with_carry(*ai, 0, c, &mut carry);
} else {
break;
}
for (a, &b) in a_lo.iter_mut().zip(b) {
*a = mac_with_carry(*a, b, c, &mut carry);
}
assert!(carry == 0);
let mut a = a_hi.iter_mut();
while carry != 0 {
let a = a.next().expect("carry overflow during multiplication!");
*a = adc(*a, 0, &mut carry);
}
}
/// Three argument multiply accumulate:
@ -250,13 +249,23 @@ fn mac3(acc: &mut [BigDigit], b: &[BigDigit], c: &[BigDigit]) {
(c, b)
};
// Karatsuba multiplication is slower than long multiplication for small x and y:
// We use three algorithms for different input sizes.
//
if x.len() <= 4 {
// - For small inputs, long multiplication is fastest.
// - Next we use Karatsuba multiplication (Toom-2), which we have optimized
// to avoid unnecessary allocations for intermediate values.
// - For the largest inputs we use Toom-3, which better optimizes the
// number of operations, but uses more temporary allocations.
//
// The thresholds are somewhat arbitrary, chosen by evaluating the results
// of `cargo bench --bench bigint multiply`.
if x.len() <= 32 {
// Long multiplication:
for (i, xi) in x.iter().enumerate() {
mac_digit(&mut acc[i..], y, *xi);
}
} else {
} else if x.len() <= 256 {
/*
* Karatsuba multiplication:
*
@ -375,6 +384,53 @@ fn mac3(acc: &mut [BigDigit], b: &[BigDigit], c: &[BigDigit]) {
},
NoSign => (),
}
} else {
// Toom-3 multiplication:
//
// Toom-3 is like Karatsuba above, but dividing the inputs into three parts.
// Both are instances of Toom-Cook, using `k=3` and `k=2` respectively.
//
// FIXME: It would be nice to have comments breaking down the operations below.
let i = y.len()/3 + 1;
let x0_len = cmp::min(x.len(), i);
let x1_len = cmp::min(x.len() - x0_len, i);
let y0_len = i;
let y1_len = cmp::min(y.len() - y0_len, i);
let x0 = BigInt::from_slice(Plus, &x[..x0_len]);
let x1 = BigInt::from_slice(Plus, &x[x0_len..x0_len + x1_len]);
let x2 = BigInt::from_slice(Plus, &x[x0_len + x1_len..]);
let y0 = BigInt::from_slice(Plus, &y[..y0_len]);
let y1 = BigInt::from_slice(Plus, &y[y0_len..y0_len + y1_len]);
let y2 = BigInt::from_slice(Plus, &y[y0_len + y1_len..]);
let p = &x0 + &x2;
let q = &y0 + &y2;
let p2 = &p - &x1;
let q2 = &q - &y1;
let r0 = &x0 * &y0;
let r4 = &x2 * &y2;
let r1 = (p + x1) * (q + y1);
let r2 = &p2 * &q2;
let r3 = ((p2 + x2)*2 - x0) * ((q2 + y2)*2 - y0);
let mut comp3: BigInt = (r3 - &r1) / 3;
let mut comp1: BigInt = (r1 - &r2) / 2;
let mut comp2: BigInt = r2 - &r0;
comp3 = (&comp2 - comp3)/2 + &r4*2;
comp2 = comp2 + &comp1 - &r4;
comp1 = comp1 - &comp3;
let result = r0 + (comp1 << 32*i) + (comp2 << 2*32*i) + (comp3 << 3*32*i) + (r4 << 4*32*i);
let result_pos = result.to_biguint().unwrap();
add2(&mut acc[..], &result_pos.data);
}
}