swapper.cpp

#include <sys/time.h>
#include <sys/resource.h>

#include <x86intrin.h>

#include <cassert>
#include <cstdint>
#include <cstdlib>
#include <byteswap.h>
#include <iostream>

#define get_time(output) \
  output = __rdtsc();

namespace D3
{
// std::byteswap from C++23
    template <typename T> constexpr T byteswap(T n) {
        T m;
        for (size_t i = 0; i < sizeof(T); i++)
            reinterpret_cast<uint8_t *>(&m)[i] = reinterpret_cast<uint8_t *>(&n)[sizeof(T) - 1 - i];
        return m;
    }
}

int main(void)
{
    if(setpriority(PRIO_PROCESS, 0, -20) != 0)
    {
         std::cout << "Must be run with superuser privileges" << std::endl;
         return 0;
    }

    constexpr const uint64_t swaps_per_test = 0xFFFFFF;

    uint16_t a = 0;
    uint32_t b = 0;
    uint64_t c = 0;
    uint64_t count = 0;
    uint64_t start, stop;
    uint64_t records[6];

    std::cout << "Performing " << swaps_per_test << " swaps per test." << std::endl;

    count = swaps_per_test;
    get_time(start);
    do {
      a = D3::byteswap(start^a);
    } while(--count);
    get_time(stop);

    std::cerr << "value 16: " << a << std::endl;

    records[0] = stop - start;
    std::cout << "D3::byteswap<uint16_t>() took " << records[0] << " cycles, an average of " << (double(records[0]) / swaps_per_test) << " cycles per call" << std::endl;

    count = swaps_per_test;
    get_time(start);
    do {
      b = D3::byteswap(start^b);
    } while(--count);
    get_time(stop);

    std::cerr << "value 32: " << b << std::endl;

    records[1] = stop - start;
    std::cout << "D3::byteswap<uint32_t>() took " << records[1] << " cycles, an average of " << (double(records[1]) / swaps_per_test) << " cycles per call" << std::endl;

    count = swaps_per_test;
    get_time(start);
    do {
      c = D3::byteswap(start^c);
    } while(--count);
    get_time(stop);

    std::cerr << "value 64: " << c << std::endl;

    records[2] = stop - start;
    std::cout << "D3::byteswap<uint64_t>() took " << records[2] << " cycles, an average of " << (double(records[2]) / swaps_per_test) << " cycles per call" << std::endl;

    count = swaps_per_test;
    get_time(start);
    do {
      a = bswap_16(start^a);
    } while(--count);
    get_time(stop);

    std::cerr << "value 16: " << a << std::endl;

    records[3] = stop - start;
    std::cout << "bswap_16() took " << records[3] << " cycles, an average of " << (double(records[3]) / swaps_per_test) << " cycles per call" << std::endl;

    count = swaps_per_test;
    get_time(start);
    do {
      b = bswap_32(start^b);
    } while(--count);
    get_time(stop);

    std::cerr << "value 32: " << b << std::endl;

    records[4] = stop - start;
    std::cout << "bswap_32() took " << records[4] << " cycles, an average of " << (double(records[4]) / swaps_per_test) << " cycles per call" << std::endl;

    count = swaps_per_test;
    get_time(start);
    do {
      c = bswap_64(start^c);
    } while(--count);
    get_time(stop);

    std::cerr << "value 32: " << c << std::endl;

    records[5] = stop - start;
    std::cout << "bswap_64() took " << records[5] << " cycles, an average of " << (double(records[5]) / swaps_per_test) << " cycles per call" << std::endl;

    std::cout << "bswap_16 was " << (float(records[0])/float(records[3])) << " times faster." << std::endl;
    std::cout << "bswap_32 was " << (float(records[1])/float(records[4])) << " times faster." << std::endl;
    std::cout << "bswap_64 was " << (float(records[2])/float(records[5])) << " times faster." << std::endl;

    return 0;
}