gistfile1.txt

#include <assert.h>
#include <stdlib.h>
#include <stdio.h>
#include <stdint.h>
#include <sys/time.h>
// #include <smmintrin.h>
// #include <emmintrin.h>

#define SAMPLES 5
#define SIZE (1024*1024*1024)
uint64_t array[SIZE/sizeof(uint64_t)];

#define to_bw(usec) ((SIZE/(1024.0 * 1024.0 * 1024.0)) / (usec/(1000.0*1000.0)))

void write_memory() {
  //  __m128 const vals = _mm_set1_ps(1.0);
  size_t i;
  float* fbase = (float*) array;
  for (i = 0; i < SIZE / sizeof(float); i += sizeof(float)) {
    // _mm_stream_ps(&fbase[i], vals);
    fbase[i] = 0;
  }
}

void read_memory() {
  float accum = 0;
  size_t i;
  float* fbase = (float*) array;
  for (i = 0; i < SIZE / sizeof(float); i += sizeof(float)) {
    accum += fbase[i]; 
  }

  // This is unlikely, and we want to make sure the reads are not optimized
  // away.
  assert(accum != (float) 0xDEADBEEF);
}

void timeit(void (*function)(), char* name) {
  size_t min_usec = 9999999999;
  size_t i;
  for (i = 0; i < SAMPLES; i++) {
    struct timeval before, after, total;

    gettimeofday(&before, NULL);
    function();
    gettimeofday(&after, NULL);

    timersub(&after, &before, &total);
    assert(total.tv_sec == 0);

    if (total.tv_usec < min_usec) {
      min_usec = total.tv_usec;
    }
  }

  double usec = min_usec;
  printf("%s: %.2f ms\n", name, usec / 1000.0);
  printf("%s: %.2f GiB/s\n", name, to_bw(usec));
}

int main() {
  // Memory must be 16 byte aligned.
  assert(((uintptr_t) array) % 16 == 0);
  assert((SIZE % 16) == 0);

  timeit(read_memory, "read");
  timeit(write_memory, "write");

  return 0;
}