c++ – C ++ vs Haskell: performance

Question:

I compare the performance. This problem is taken as an example. C ++ solution

#include <iostream>
#include <limits>
#include <cmath>
#include <string>
#include <cstdlib>
#include <boost/multiprecision/cpp_int.hpp>
#include <boost/rational.hpp>

using namespace boost;

typedef boost::multiprecision::cpp_int LL;

const char MSG_ERR[] = "Error\n";
const size_t MSG_ERR_LEN = std::strlen(MSG_ERR);

rational<LL> *x;

rational<LL> fun(const rational<LL> & y, const rational<LL> & z) {
    rational<LL> ret = 108 - (815 - (1500 / z))/y;
    return ret;
}

int main(int argn, char *argv[]) {
    if(argn != 2) {
        std::cout.write(MSG_ERR, MSG_ERR_LEN);
        return 1;
    }
    int n = std::atoi(argv[1]) + 1;
    if(n < 1) {
        std::cout.write(MSG_ERR, MSG_ERR_LEN);
        return 1;
    }
    x = new rational<LL>[n];
    x[0] = 4;
    x[1] = rational<LL>(17,4);
    for(int i = 2; i < n; ++i)
        x[i] = fun(x[i-1], x[i-2]);
    std::cout << x[n-1] << std::endl;
    return 0;
}

Haskell:

import Data.Ratio
import System.Environment

f y z = 108 - (815 - 1500 / z) / y

xlist = 4 : (17 % 4) : zipWith f (tail xlist) xlist

main = do
     args <- getArgs
     case args of
                [arg1] -> do
                            let x = read arg1
                            putStrLn $ show $ xlist !! x
                _      -> error "ERR: Main.Args"

Results: C ++:

$ clang++37 -O3 -I/usr/local/include -o CPP t1.cpp 
$ time ./CPP 10000 > /dev/null
       39,64 real        39,62 user         0,00 sys

Haskell:

$ ghc -O2 t1.hs -o HS
Linking HS ...
$ time ./HS 10000 > /dev/null
        6,73 real         6,73 user         0,00 sys

How can you explain such a significant difference? Boost? Laziness?

Answer:

Such a significant difference can be explained by the fact that boost not optimized as much as possible, although the result is much better than writing long arithmetic by hand.

Now let's look at what potential we can squeeze out of С++ . For this we will use the GMP library, take this code:

#include <cstddef>
#include <iostream>
#include <vector>
#include <gmpxx.h>

const char MSG_ERR[] = "Error\n";
const size_t MSG_ERR_LEN = std::strlen(MSG_ERR);

mpq_class fun(const mpq_class & y, const mpq_class & z) {
    mpq_class ret = 108 - (815 - (1500 / z))/y;
    return ret;
}

int main(int argc, char *argv[])
{
    if(argc != 2)
    {
        std::cout.write(MSG_ERR, MSG_ERR_LEN);
        return 1;
    }

    int n = std::atoi(argv[1]) + 1;

    if(n < 1)
    {
        std::cout.write(MSG_ERR, MSG_ERR_LEN);
        return 1;
    }

    std::vector<mpq_class> v;
    v.push_back(mpq_class(4, 1));
    v.push_back(mpq_class(17, 4));

    for (int i = 2; i < n; ++i)
    {
        v.push_back(fun(v[i - 1], v[i - 2]));
    }
    std::cout << v.rbegin()->get_num() << " / " << v.rbegin()->get_den() << std::endl;

    return 0;
}

And let's collect all the results in a heap, on my machine I got the following result:

Haskell

> ghc -O2 test.hs -o HS
> time ./HS 10000 > /dev/null 
real    0m7.491s
user    0m7.420s
sys     0m0.004s

C++, boost 

> g++-5 -O2 -o CPP test.cpp
> time ./CPP 10000 > /dev/null
real    0m22.272s
user    0m22.256s
sys     0m0.012s

> clang++ -O2 -o CPP-CLANG test.cpp
> time ./CPP-CLANG 10000 > /dev/null
real    0m20.638s
user    0m20.628s
sys     0m0.008s

C++, gmp

> g++-5 -O2 -std=c++14 -lgmpxx -lgmp -o CPP-GMP testgmp.cpp
> time ./CPP-GMP 10000 > /dev/null
real    0m1.418s
user    0m1.404s
sys     0m0.016s

We made sure that boost doesn't give us the best performance.

Post Views: 5
Scroll to Top