Hello World Parallel Bench Derived From OpenMP Hello World Example

hpx_for_each.cpp

#include <hpx/hpx.hpp>
#include <hpx/hpx_init.hpp>
#include <hpx/parallel/algorithms/for_each.hpp>

#include <stdio.h>
#include <chrono>
#include <cstddef>
#include <iterator>
#include <vector>
#include "stopwatch.hpp"

using namespace std;
using namespace std::chrono;

int A[3000];

template<class ExecutionPolicy>
void test(ExecutionPolicy&& exec) {
    hpx::parallel::for_each(std::forward<ExecutionPolicy>(exec), begin(A), end(A), [](int& r) {
        const int j = static_cast<int>(&r - &A[0]);
        int red = 0;
        for (int i=0; i<100000; i++) {
            red += static_cast<int>(sin(i+j) + cos(i+j));
        }
        r = red;
    });
}

///////////////////////////////////////////////////////////////////////////////
#pragma optimize("", off)
int hpx_main(boost::program_options::variables_map& vm)
{
    stopwatch sw;
    sw.start();
    for (int idx = 0; idx < 1; ++idx)
        test(hpx::parallel::execution::par_unseq);
    sw.stop();
    printf("A[70] is %d\n", A[70]);
	sw.print("parallel for_each");
    return hpx::finalize();
}
#pragma optimize("", on)

///////////////////////////////////////////////////////////////////////////////
int main(int argc, char* argv[])
{
    return hpx::init(argc, argv);
}

openmp.cpp

#include <math.h>
#include <chrono>
#include "stopwatch.hpp"

int A[3000];

void test() {
    #pragma omp parallel for
    for (int j=0; j<3000; j++) {
        int red = 0;
        for (int i=0; i<100000; i++) {
            red += static_cast<int>(sin(i+j) + cos(i+j));
        }
        A[j] = red;
    }
}

#pragma optimize("", off)

#include <stdio.h>
int main() {
    stopwatch sw;
    sw.start();
    for (int idx = 0; idx < 1; ++idx)
        test();
    sw.stop();
    printf("A[70] is %d\n", A[70]);
	sw.print("openmp");
}

#pragma optimize("", on)

parallel_for_each.cpp

#include <math.h>
#include <stdio.h>
#include <chrono>
#include <execution>
#include <iterator>
#include "stopwatch.hpp"

using namespace std;
using namespace std::chrono;
using namespace std::execution;

int A[3000];

template<class ExecutionPolicy>
void test(ExecutionPolicy&& exec) {
    for_each(std::forward<ExecutionPolicy>(exec), begin(A), end(A), [](int& r) {
        const int j = static_cast<int>(&r - &A[0]);
        int red = 0;
        for (int i=0; i<100000; i++) {
            red += static_cast<int>(sin(i+j) + cos(i+j));
        }
        r = red;
    });
}

#pragma optimize("", off)

int main() {
    stopwatch sequential;
    stopwatch parallel;

    puts("Testing sequential");
    fill(begin(A), end(A), int{});
    sequential.start();
    test(execution::seq);
    sequential.stop();
    printf("A[70] is %d\n", A[70]);
    sequential.print("Sequential");

    puts("Testing parallel");
    fill(begin(A), end(A), int{});
    parallel.start();
    test(execution::par);
    parallel.stop();
    printf("A[70] is %d\n", A[70]);
    parallel.print("Parallel");

    printf("Speedup: %fx\n", parallel.compare(sequential));
}

#pragma optimize("", on)

Results

cl /Zi /std:c++latest /MD /EHsc /O2 /openmp .\openmp.cpp
cl /Zi /std:c++latest /MD /EHsc /O2 .\parallel_for_each_comparative.cpp
cl /D_HAS_AUTO_PTR_ETC=1 /std:c++latest /MD /EHsc /O2 /IC:\Users\bion\Desktop\vcpkg-export-20170608-100432\installed\x86-windows\include .\hpx_for_each.cpp /link /libpath:C:\Users\bion\Desktop\vcpkg-export-20170608-100432\installed\x86-windows\lib hpx.lib hpx_init.lib boost_program_options-vc140-mt-1_64.lib boost_system-vc140-mt-1_64.lib boost_timer-vc140-mt-1_64.lib boost_date_time-vc140-mt-1_64.lib boost_thread-vc140-mt-1_64.lib


PS TEST WCFB01 C:\Users\bion\Desktop>.\parallel_for_each_comparative.exe
Testing sequential
A[70] is 1
Sequential took 2874ms

Testing parallel
A[70] is 1
Parallel took 300ms

Speedup: 9.574882x
PS TEST WCFB01 C:\Users\bion\Desktop>.\openmp.exe
A[70] is 1
openmp took 303ms

PS TEST WCFB01 C:\Users\bion\Desktop>.\hpx_for_each.exe
A[70] is 1
parallel for_each took 2918ms

PS TEST WCFB01 C:\Users\bion\Desktop>

stopwatch.hpp

#pragma once

#include <stdio.h>
#include <chrono>

class stopwatch {
    std::chrono::high_resolution_clock::time_point startTime;
    std::chrono::high_resolution_clock::time_point stopTime;

public:
    stopwatch()
        : startTime()
        , stopTime()
    { }

    stopwatch(const stopwatch&) = delete;
    stopwatch& operator=(const stopwatch&) = delete;

    void start() {
        startTime = std::chrono::high_resolution_clock::now();
    }

    void stop() {
        stopTime = std::chrono::high_resolution_clock::now();
    }

    std::chrono::high_resolution_clock::duration time_taken() const {
        return stopTime - startTime;
    }

    double compare(const stopwatch& s) {
        return static_cast<double>(s.time_taken().count()) / time_taken().count();
    }

    void print(const char * const action) const {
        printf("%s took %lldms\n\n", action,
            static_cast<long long>(std::chrono::duration_cast<std::chrono::milliseconds>(time_taken()).count()));
    }
};