sum.cpp

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

long long sum[4] __attribute__((aligned(32))) = {1LL, 2LL, 3LL, 4LL};
long long four[4] __attribute__((aligned(32))) = {4LL, 4LL, 4LL, 4LL};

int main() {
  struct timeval start, end;
  gettimeofday(&start, NULL);
  __asm__ (
    "vpxor %%ymm0, %%ymm0, %%ymm0;"
    "vmovdqa %0, %%ymm1;"
    "vmovdqa %1, %%ymm2;"
    "movl $250000000, %%ecx;"
    "LOOP:"
      "vpaddq %%ymm0, %%ymm1, %%ymm0;"
      "vpaddq %%ymm1, %%ymm2, %%ymm1;"
      "dec %%ecx;"
      "jnz LOOP;"
    "vmovdqa %%ymm0, %0;"
    "vzeroall;"
      :: "m"(sum), "m"(four) : "memory"
    );
  gettimeofday(&end, NULL);

  printf("sum: %lld, time used: %f 's", 
  sum[0] + sum[1] + sum[2] + sum[3],
  end.tv_sec - start.tv_sec + (end.tv_usec - start.tv_usec) / 1000000.0);
}

// clang++ -o sum.exe sum.cpp -mavx2
// sum: 500000000500000000, time used: 0.069361 's