Cache prefetch example

main.cpp

#include <array>
#include <chrono>
#include <iostream>


// Width of each dimension
constexpr size_t N = 200;

// Number of iterations for the test
constexpr size_t NUM_ITERATIONS = 100;

// Data type we'll use for testing
using DataType = short;

// 1D array index from a 3D coordinate
int index(int x, int y, int z) {
    return N*(N*y + z) + x;
}

void test1() {
    using namespace std::chrono;

    const DataType data[N*N*N] = {0};
    int sum = 0;

    const auto start = high_resolution_clock::now();
    for(int iteration = 0; iteration < NUM_ITERATIONS; ++iteration) {
        for(int x = 0; x < N; ++x) {
            for(int z = 0; z < N; ++z) {
                for(int y = 0; y < N; ++y) {
                    if(x > 0 && x + 1 < N) {
                        sum += data[index(x - 1, y, z)] / 2;
                        sum += data[index(x, y, z)];
                        sum += data[index(x + 1, y, z)] / 2;
                    } else {
                        sum += data[index(x, y, z)];
                    }
                }
            }
        }
    }

    const auto elapsed = high_resolution_clock::now() - start;
    const auto ms = duration_cast<milliseconds>(elapsed).count();
    std::cout << sum << " in " << ((1.0 * ms) / NUM_ITERATIONS)  << "ms per loop\n";
}

void test2() {
    using namespace std::chrono;

    const DataType data[N*N*N] = {0};
    int sum = 0;

    const auto start = high_resolution_clock::now();
    for(int iteration = 0; iteration < NUM_ITERATIONS; ++iteration) {
        for(int x = 0; x < N; ++x) {
            for(int z = 0; z < N; ++z) {
                for(int y = 0; y < N; ++y) {
                    if(x > 0 && x + 1 < N) {
                        sum += data[index(x - 1, y, z)] / 2;
                        sum += data[index(x, y, z)];
                        sum += data[index(x + 1, y, z)] / 2;
                    } else {
                        sum += data[index(x, y, z)];
                    }

                    const int prefetch_index = index(x, y + 20, z);
                    __builtin_prefetch(data + prefetch_index, 0, 0);
                }
            }
        }
    }

    const auto elapsed = high_resolution_clock::now() - start;
    const auto ms = duration_cast<milliseconds>(elapsed).count();
    std::cout << sum << " in " << ((1.0 * ms) / NUM_ITERATIONS)  << "ms per loop\n";
}

int main(int argc, char **argv) {
    test1();
    test2();
    return 0;
}

test.sh-session

$ clang++ -O2 -std=c++1y main.cpp
$ ./a.out
0 in 145.29ms per loop
0 in 84.02ms per loop