#include <cinttypes>
#include <cstdlib>
#include <cstring>
#include <ctime>
#include <iostream>

std::uint32_t tbl[0x1000000];
const char b64alpha[] = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";

void base64b(const char source[], std::size_t size) {
    std::cout << "Run bitwise \n";
    auto result = new char[size*4/3];
    struct timespec start;
    clock_gettime(CLOCK_MONOTONIC, &start);
    for (auto counter = 0; counter < 100; ++counter) {
      for (auto i = 0, j = 0; i < size; i +=3, j += 4) {
        auto block = (std::uint8_t(source[i]) << 16) |
                      (std::uint8_t(source[i+1]) << 8) |
                      std::uint8_t(source[i+2]);
        result[j+0] = b64alpha[(block>>18)&0x3f];
        result[j+1] = b64alpha[(block>>12)&0x3f];
        result[j+2] = b64alpha[(block>>6)&0x3f];
        result[j+3] = b64alpha[block&0x3f];
      }
      asm volatile("" :: "m"(result));
    }
    struct timespec finish;
    clock_gettime(CLOCK_MONOTONIC, &finish);
    auto delta = std::int64_t(finish.tv_nsec) + std::int64_t(finish.tv_sec) * 1000000000 -
                 std::int64_t(start.tv_nsec) - std::int64_t(start.tv_sec) * 1000000000;
    std::cout << (double(delta) / 1000000000) << "s \n";
    (std::cout << "Base64: ").write(result, 40) << "\n";
    delete[] result;
}

void base64t(const char source[], std::size_t size) {
    std::cout << "Run tabular \n";
    auto result = new char[size*4/3];
    struct timespec start;
    clock_gettime(CLOCK_MONOTONIC, &start);
    for (auto counter = 0; counter < 100; ++counter) {
      for (auto i = 0, j = 0; i < size; i +=3, j += 4) {
        uint32_t block;
        std::memcpy(&block, &source[i], sizeof(block));
        block &= 0xffffff;
        std::memcpy(&result[j], &tbl[block], sizeof(uint32_t));
      }
      asm volatile("" :: "m"(result));
    }
    struct timespec finish;
    clock_gettime(CLOCK_MONOTONIC, &finish);
    auto delta = std::int64_t(finish.tv_nsec) + std::int64_t(finish.tv_sec) * 1000000000 -
                 std::int64_t(start.tv_nsec) - std::int64_t(start.tv_sec) * 1000000000;
    std::cout << (double(delta) / 1000000000) << "s \n";
    (std::cout << "Base64: ").write(result, 40) << "\n";
    delete[] result;
}

char source[3*10*1024*1024];
int main() {
  std::cout << "Prepare table\n";
  for (auto i = 0; i < 0x1000000; i++) {
    auto block = ((i & 0xff) << 16) | (i & 0xff00) | (i >> 16);

    tbl[i] = std::uint32_t(b64alpha[(block>>18)&0x3f]) |
             std::uint32_t(b64alpha[(block>>12)&0x3f])<<8 |
             std::uint32_t(b64alpha[(block>>6)&0x3f])<<16 |
             std::uint32_t(b64alpha[block&0x3f])<<24;
  }

  const char alpha[] = "abcdefghijklmnopqrstuvwxyz0123456789 {}\":";
  for (auto i = 0; i < sizeof(source); ++i) {
    source[i] = alpha[rand() % (sizeof(alpha) - 1)];
  }
  (std::cout << "Text source data:").write(source, 30) << "\n";
  base64b(source, sizeof(source));
  base64t(source, sizeof(source));
}