Khai Nguyen Question

#include <emmintrin.h>
#include <stdlib.h>
#include <stdio.h>
#include <time.h>

#define NEAR 10
#define T1 100
#define T2 1000
#define T3 10000

int func(int Di) {
  if (Di <= -T3) return  -4;
  if (Di <= -T2) return  -3;
  if (Di <= -T1) return  -2;
  if (Di < -NEAR)  return  -1;
  if (Di <= NEAR) return   0;
  if (Di < T1)   return   1;
  if (Di < T2)   return   2;
  if (Di < T3)   return   3;
  return  4;
}

int funcX(int Di) {
  int cmp_m3 = -(Di > -T3);
  int cmp_m2 = -(Di > -T2);
  int cmp_m1 = -(Di > -T1);
  int cmp_p0 = -(Di > NEAR);
  int reduce_true = cmp_m3 + cmp_m2 + cmp_m1 + cmp_p0;
  int cmp_m0 = -(Di < -NEAR);
  int cmp_p1 = -(Di < T1);
  int cmp_p2 = -(Di < T2);
  int cmp_p3 = -(Di < T3);
  int reduce_false = cmp_p3 + cmp_p2 + cmp_p1 + cmp_m0;
  return reduce_false - reduce_true;
}

__m128i func4(__m128i D) {
  __m128i cmp_m3 = _mm_cmpgt_epi32(D, _mm_set1_epi32(-T3));
  __m128i cmp_m2 = _mm_cmpgt_epi32(D, _mm_set1_epi32(-T2));
  __m128i cmp_m1 = _mm_cmpgt_epi32(D, _mm_set1_epi32(-T1));
  __m128i cmp_p0 = _mm_cmpgt_epi32(D, _mm_set1_epi32(NEAR));
  __m128i reduce_true = _mm_add_epi32(_mm_add_epi32(cmp_m3, cmp_m2), _mm_add_epi32(cmp_m1, cmp_p0));
  __m128i cmp_m0 = _mm_cmplt_epi32(D, _mm_set1_epi32(-NEAR));
  __m128i cmp_p1 = _mm_cmplt_epi32(D, _mm_set1_epi32(T1));
  __m128i cmp_p2 = _mm_cmplt_epi32(D, _mm_set1_epi32(T2));
  __m128i cmp_p3 = _mm_cmplt_epi32(D, _mm_set1_epi32(T3));
  __m128i reduce_false = _mm_add_epi32(_mm_add_epi32(cmp_p3, cmp_p2), _mm_add_epi32(cmp_p1, cmp_m0));
  return _mm_sub_epi32(reduce_false, reduce_true);
}

const int COUNT = 16<<10;

static union {
  int data_i[4 * COUNT];
  __m128i data_r[COUNT];
};

static union {
  int res_i[4];
  __m128i res_r;
};

int main() {
  for (int i = 0; i < 4*COUNT; i++)
    data_i[i] = rand() % 32768 - 16768;

  for (int i = 0; i < COUNT; i++) {
    int corr[4];
    for (int j = 0; j < 4; j++)
      corr[j] = func(data_i[4*i+j]);

    res_r = func4(data_r[i]);
    for (int j = 0; j < 4; j++)
      if (corr[j] != res_i[j])
        printf("Wrong!");

    for (int j = 0; j < 4; j++)
      res_i[j] = funcX(data_i[4*i+j]);
    for (int j = 0; j < 4; j++)
      if (corr[j] != res_i[j])
        printf("Wrong!");
  }

  {
    int start = clock();
    int sum = 0;
    for (int q = 0; q < 10000; q++) {
      for (int i = 0; i < 4*COUNT; i++)
        sum += func(data_i[i]);
    }
    printf("Time = %0.3lf   (%d)\n", double(clock() - start) / CLOCKS_PER_SEC, sum);
  }

  {
    int start = clock();
    __m128i sum = _mm_setzero_si128();
    for (int q = 0; q < 10000; q++) {
      for (int i = 0; i < COUNT; i++)
        sum = _mm_add_epi32(sum, func4(data_r[i]));
    }
    res_r = sum;
    int rsum = res_i[0] + res_i[1] + res_i[2] + res_i[3];
    printf("Time = %0.3lf   (%d)\n", double(clock() - start) / CLOCKS_PER_SEC, rsum);
  }

  return 0;
}