测试alignment对内存访问性能的影响

alignment.cc

#include <iostream>
#include <stdint.h>

template<typename T>
inline void copy_byte(uint8_t* dst, uint8_t* src, size_t size ) {
  T* sdst = (T*)dst;
  T* ssrc = (T*)src;
  size /= sizeof(T);
  clock_t s = clock();
  while (size-- > 0)
    *sdst++ = *ssrc++;
  clock_t e = clock();
  std::cout << "size:" << sizeof(T) << "," << (e - s) * 1.0/(CLOCKS_PER_SEC) << std::endl;
}

int main(int argc, char** argv) {
  const size_t kSize = 1024*1024*100;
  uint8_t* dst = (uint8_t*) malloc(kSize);
  uint8_t* src = (uint8_t*) malloc(kSize);
  std::cout << (void*)dst << std::endl;

  for (int i = 0; i < 8; i++) {
  std::cout << (void*)(dst + i) << std::endl;

    clock_t s = clock();
    memcpy(dst + i, src + i, kSize - i);
    clock_t e = clock();
    std::cout << (e - s) * 1.0/CLOCKS_PER_SEC << std::endl;

    copy_byte<uint8_t>(dst + i, src + i, kSize - i);
    copy_byte<uint16_t>(dst + i, src + i, kSize - i);
    copy_byte<uint32_t>(dst + i, src + i, kSize - i);
    copy_byte<size_t>(dst + i, src + i, kSize - i);
  }
  // memcpy(dst + 1, src + 1, kSize - 1);
  return 0;
}

/*
 caoshiweideMacBook-Pro:test caoshiwei$ g++ -O3 alignment.cc
 caoshiweideMacBook-Pro:test caoshiwei$ ./a.out
 0x10652f000
 0x10652f000
 0.099874
 size:1,0.017212
 size:2,0.016584
 size:4,0.025358
 size:8,0.017103
 0x10652f001
 0.015442
 size:1,0.025877
 size:2,0.016838
 size:4,0.017774
 size:8,0.021104
 0x10652f002
 0.01942
 size:1,0.025366
 size:2,0.019621
 size:4,0.025841
 size:8,0.018191
 0x10652f003
 0.020936
 size:1,0.022729
 size:2,0.023116
 size:4,0.029354
 size:8,0.022287
 0x10652f004
 0.022226
 size:1,0.023343
 size:2,0.023158
 size:4,0.017101
 size:8,0.022718
 0x10652f005
 0.021892
 size:1,0.023646
 size:2,0.025684
 size:4,0.019284
 size:8,0.026392
 0x10652f006
 0.019575
 size:1,0.021858
 size:2,0.024489
 size:4,0.025159
 size:8,0.023368
 0x10652f007
 0.018965
 size:1,0.026968
 size:2,0.01899
 size:4,0.025701
 size:8,0.020535
*/

alignment1.cc

#include <iostream>
#include <stdint.h>

int main(int argc, char** argv) {
  const size_t kSize = 1024*1024;
  uint8_t* mem1 = (uint8_t*) malloc(kSize*sizeof(double));
  uint8_t* mem2 = (uint8_t*) malloc(kSize*sizeof(double));
  std::cout << (void*) mem1 << "," << (void*) mem2 << std::endl;

  for (int i = 0; i < kSize*sizeof(double); i++) {
    mem1[i] = 0;
    mem2[i] = 0;
  }

  double sum = 0;
  for (int align = 0; align < 8; align++) {
    double* nums = (double*)(mem1 + align);
    double* numb = (double*)(mem2 + align);

    for (int i = 0; i < kSize - 1; i++) {
      nums[i] = i*random();
      numb[i] = i*random();
    }

    clock_t s = clock();
    for (int i = 0; i < kSize - 1; i++) {
      nums[i] = nums[i] * numb[i];
      sum  += nums[i];
    }
    clock_t e = clock();
    std::cout << "align:" << align << "," << (e - s) * 1.0/(CLOCKS_PER_SEC) << std::endl;
    std::cout << sum << std::endl;
  }
  return 0;
}
/*
caoshiweideMacBook-Pro:test caoshiwei$ ./a.out
0x10fd9d000,0x11059e000
align:0,0.00296
4.42795e+35
align:1,0.001538
8.86286e+35
align:2,0.003423
1.32945e+36
align:3,0.001611
1.77246e+36
align:4,0.001402
2.21554e+36
align:5,0.00242
2.65889e+36
align:6,0.001776
3.10164e+36
align:7,0.002441
3.5454e+36
 */
/*
swcao@swcao-linux:~$ ./a.out 
0x7f5358757010,0x7f5326756010
align:0,0.38
4.43033e+41
align:1,0.37
8.86186e+41
align:2,0.38
1.32928e+42
align:3,0.37
1.77234e+42
align:4,0.38
2.21538e+42
align:5,0.37
2.65848e+42
align:6,0.38
3.10148e+42
align:7,0.38
3.54458e+42

*/