scaleneg.cpp

#include <iostream>
#include <cmath>
#include <cstdlib>
#include <ctime>

#define FP double

using std::cout;
using std::endl;

constexpr size_t num_FPs = 8 * 8192;

FP scaleneg_copysign(FP a, FP b)
{
    FP c = std::copysign(1.0, a);
    c = (c + 1) * 0.5;
    return (a * b * (1.0-c)) + (a * c);
}

FP scaleneg_signbit(FP a, FP b)
{
    int c = std::signbit(a);
    return (a * b * c) + (a * (1.0-c));
}

FP scaleneg_signbit_float(FP a, FP b)
{
    FP c = std::signbit(a);
    return (a * b * c) + (a * (1.0-c));
}

FP scaleneg_naive(FP a, FP b)
{
    return a < 0 ? a * b : a;
}

// between -5 and 5
FP drand() { return ((FP)rand() / RAND_MAX - 0.5) * 10; }

FP nums[num_FPs];

using scaleneg_t = FP (*)(FP, FP);

void time_func(scaleneg_t scaleneg)
{
    clock_t begin = std::clock();

    FP accum = 0.;
    for (size_t i = 0; i < num_FPs; ++i) {
        accum += scaleneg(nums[i], nums[i]);
    }

    clock_t end = std::clock();
    cout << "\tTime: " << ((FP)(end - begin) / CLOCKS_PER_SEC) << " [accum=" << accum << "]\n";
}

int main()
{
    // fixed-point floating point
    cout << std::fixed;

    for (size_t i = 0; i < num_FPs; ++i) {
        nums[i] = drand();
    }

    for (size_t i = 0; i < 3; ++i) {
        cout << "Run " << (i+1) << endl;
        cout << "scaleneg_signbit:" << endl;
        time_func(&scaleneg_signbit);
        cout << "scaleneg_naive:" << endl;
        time_func(&scaleneg_naive);
        cout << "scaleneg_copysign:" << endl;
        time_func(&scaleneg_copysign);
        cout << "scaleneg_signbit_float:" << endl;
        time_func(&scaleneg_signbit_float);
    }

    return 0;
}

Results:

gcc 7

Run 1
scaleneg_signbit:
	Time: 0.000114 [accum=353518.582899]
scaleneg_naive:
	Time: 0.000283 [accum=353518.582899]
scaleneg_copysign:
	Time: 0.000102 [accum=353518.582899]
scaleneg_signbit_float:
	Time: 0.000117 [accum=353518.582899]
Run 2
scaleneg_signbit:
	Time: 0.000115 [accum=353518.582899]
scaleneg_naive:
	Time: 0.000303 [accum=353518.582899]
scaleneg_copysign:
	Time: 0.000106 [accum=353518.582899]
scaleneg_signbit_float:
	Time: 0.000108 [accum=353518.582899]
Run 3
scaleneg_signbit:
	Time: 0.000126 [accum=353518.582899]
scaleneg_naive:
	Time: 0.000342 [accum=353518.582899]
scaleneg_copysign:
	Time: 0.000104 [accum=353518.582899]
scaleneg_signbit_float:
	Time: 0.000126 [accum=353518.582899]

AppleClang 9.0

Run 1
scaleneg_signbit:
	Time: 0.000332 [accum=353518.582899]
scaleneg_naive:
	Time: 0.000062 [accum=353518.582899]
scaleneg_copysign:
	Time: 0.000114 [accum=353518.582899]
scaleneg_signbit_float:
	Time: 0.000347 [accum=353518.582899]
Run 2
scaleneg_signbit:
	Time: 0.000323 [accum=353518.582899]
scaleneg_naive:
	Time: 0.000059 [accum=353518.582899]
scaleneg_copysign:
	Time: 0.000105 [accum=353518.582899]
scaleneg_signbit_float:
	Time: 0.000347 [accum=353518.582899]
Run 3
scaleneg_signbit:
	Time: 0.000351 [accum=353518.582899]
scaleneg_naive:
	Time: 0.000058 [accum=353518.582899]
scaleneg_copysign:
	Time: 0.000103 [accum=353518.582899]
scaleneg_signbit_float:
	Time: 0.000346 [accum=353518.582899]

VS 2017

num_FPs was increased to 8192 * 8192 because of the lower precision of the default clock.

Run 1
scaleneg_signbit:
        Time: 0.348000 [accum=363539847.782539]
scaleneg_naive:
        Time: 0.490000 [accum=363539847.782539]
scaleneg_copysign:
        Time: 1.200000 [accum=363539847.782539]
scaleneg_signbit_float:
        Time: 0.361000 [accum=363539847.782539]
Run 2
scaleneg_signbit:
        Time: 0.355000 [accum=363539847.782539]
scaleneg_naive:
        Time: 0.531000 [accum=363539847.782539]
scaleneg_copysign:
        Time: 1.186000 [accum=363539847.782539]
scaleneg_signbit_float:
        Time: 0.354000 [accum=363539847.782539]
Run 3
scaleneg_signbit:
        Time: 0.357000 [accum=363539847.782539]
scaleneg_naive:
        Time: 0.484000 [accum=363539847.782539]
scaleneg_copysign:
        Time: 1.197000 [accum=363539847.782539]
scaleneg_signbit_float:
        Time: 0.359000 [accum=363539847.782539]

Best:

• VS 2017: scaleneg_signbit
• AppleClang: scaleneg_naive
• gcc 7: scaleneg_copysign (though scaleneg_signbit is nearly the same)