为什么这个Rust程序这么慢？我错过了什么？

Question

我读了曼哈顿的最小距离度量,并重写了作者在Rust中的"天真"实现.C++变体是:

#include <utility>
#include <cstdio>
#include <cstdlib>

std::pair<int, int> pointsA[1000001];
std::pair<int, int> pointsB[1000001];

int main() {
    int n, t;
    unsigned long long dist;

    scanf("%d", &t);

    while(t-->0) {
        dist = 4000000000LL;
        scanf("%d", &n);

        for(int i = 0; i < n; i++) {
            scanf("%d%d", &pointsA[i].first, &pointsA[i].second);
        }

        for(int i = 0; i < n; i++) {
            scanf("%d%d", &pointsB[i].first, &pointsB[i].second);
        }

        for(int i = 0; i < n ;i++) {
            for(int j = 0; j < n ; j++) {
                if(abs(pointsA[i].first - pointsB[j].first) + abs(pointsA[i].second - pointsB[j].second) < dist)
                    dist = abs(pointsA[i].first - pointsB[j].first) + abs(pointsA[i].second - pointsB[j].second);
            }
        }
        printf("%lld\n", dist);
    }
}

Rust变种是:

use std::io;
use std::io::BufReader;
use std::io::BufRead;

fn read_array(stdin: &mut BufReader<io::Stdin>, array_len: usize, points: &mut Vec<(i32, i32)>) {
    let mut line = String::new();
    for _ in 0..array_len {
        line.clear();
        stdin.read_line(&mut line).unwrap();
        let mut item = line.split_whitespace();
        let x = item.next().unwrap().parse().unwrap();
        let y = item.next().unwrap().parse().unwrap();
        points.push((x, y));
    }
}

fn manhattan_dist(a: &(i32, i32), b: &(i32, i32)) -> u32 {
    ((a.0 - b.0).abs() + (a.1 - b.1).abs()) as u32
}

fn main() {
    let mut line = String::new();
    let mut stdin = BufReader::new(io::stdin());
    stdin.read_line(&mut line).unwrap();
    let n_iters = line.trim_right().parse::<usize>().unwrap();
    let mut points_a = Vec::with_capacity(10000);
    let mut points_b = Vec::with_capacity(10000);
    for _ in 0..n_iters {
        line.clear();
        stdin.read_line(&mut line).unwrap();
        let set_len = line.trim_right().parse::<usize>().unwrap();
        points_a.clear();
        points_b.clear();
        read_array(&mut stdin, set_len, &mut points_a);
        read_array(&mut stdin, set_len, &mut points_b);
        let mut dist = u32::max_value();
        for i in points_a.iter() {
            for j in points_b.iter() {
                dist = std::cmp::min(manhattan_dist(i, j), dist);
            }
        }
        println!("{}", dist);
    }
}

然后,我使用Python脚本生成数据:

import random

ITER = 100
N = 10000
MAX_INT = 1000000

print("%d" % ITER)

for _ in range(0, ITER):
    print("%d" % N)
    for _ in range(0, N):
        print(random.randrange(-MAX_INT, MAX_INT + 1), random.randrange(1, MAX_INT + 1))
    for _ in range(0, N):
        print(random.randrange(-MAX_INT, MAX_INT + 1), random.randrange(-MAX_INT, 0))

并分别使用g++ -Ofast -march=native和编译了两个变体rustc -C opt-level=3.时间是:

C++

real    0m7.789s
user    0m7.760s
sys     0m0.020s

锈

real    0m28.589s
user    0m28.570s
sys     0m0.010s

为什么我的Rust代码比C++变体慢四倍？我正在使用Rust 1.12.0-beta.1.

我添加了时间测量:

let now = SystemTime::now();
line.clear();
stdin.read_line(&mut line).unwrap();
let set_len = line.trim_right().parse::<usize>().unwrap();
points_a.clear();
points_b.clear();
read_array(&mut stdin, set_len, &mut points_a);
read_array(&mut stdin, set_len, &mut points_b);
io_time += now.elapsed().unwrap();

let now = SystemTime::now();
let mut dist = u32::max_value();
for i in points_a.iter() {
    for j in points_b.iter() {
        dist = std::cmp::min(manhattan_dist(i, j), dist);
    }
}
calc_time += now.elapsed().unwrap();

和writeln!(&mut std::io::stderr(), "io_time: {}, calc_time: {}", io_time.as_secs(), calc_time.as_secs()).unwrap();打印io_time: 0, calc_time: 27.

我每晚都试过rustc 1.13.0-nightly (e9bc1bac8 2016-08-24):

$ time ./test_rust < data.txt  > test3_res
io_time: 0, calc_time: 19

real    0m19.592s
user    0m19.560s
sys     0m0.020s
$ time ./test1 < data.txt  > test1_res

real    0m7.797s
user    0m7.780s
sys     0m0.010s

所以它在我的Core i7上差不多是2.7倍.

Answer 1

差别当然是-march=native......有点.Rust有这个-C target_cpu=native,但这并没有给出相同的速度优势.这是因为LLVM不愿意在这种情况下进行矢量化,而GCC则不然.您可能会注意到,使用Clang(一种也使用LLVM的C++编译器)也会产生相对较慢的代码.

为了鼓励LLVM进行矢量化,可以将主循环移动到单独的函数中.或者,您可以使用本地块.如果你仔细编写代码

let dist = {
    let mut dist = i32::max_value();
    for &(a, b) in &points_a[..n] {
        for &(c, d) in &points_b[..n] {
            dist = std::cmp::min(((a - c).abs() + (b - d).abs()), dist);
        }
    }
    dist
} as u32;

Rust和C++之间的区别几乎可以忽略不计(~4%).

Answer 2

你在C++中看到的绝大部分性能都归功于旗帜-march=native.

这个标志不是Rust的等价标志--release.它采用特定于它是在编译的CPU CPU指令,所以数学特别是将是方式更快.

删除该标志会使C++代码处于19秒.

然后是C++代码中存在的不安全现象.没有选中任何输入.Rust代码检查它,你使用.unwrap()- unwrap具有性能成本,有断言,然后是展开所需的代码等.

使用if lets而不是raw unwraps,或者在可能的情况下忽略结果,会再次降低Rust代码.

锈:22秒

C++:19秒

3秒来自哪里？一些游戏让我相信它的println!对比printf,但我没有C++代码的硬数字.我可以说的是,当我在基准测试之外执行打印时,Rust代码会下降到13秒.

TLDR:您的编译器标志不同,您的C++代码不安全.