Test for Precision with Vectorization

gistfile1.c

#include <stdio.h>
#include <xmmintrin.h>

int
main()
{
  int i;
  float fi;
  union {
    int i;
    float f;
  } sum;
  __m128 vi, vsum, vinc;
  vinc = _mm_set1_ps(0.04f);

  sum.f = 0.0f;
  fi = 0.01f;
  for (i = 0; i < 100; i++) {
    sum.f += fi;
    fi += 0.01f;
  }
  printf("From 1 to 100:\t%f (%08x.)\n", sum.f, sum.i);

  sum.f = 0.0f;
  fi = 1.0f;
  for (i = 0; i < 100; i++) {
    sum.f += fi;
    fi -= 0.01f;
  }
  printf("From 100 to 1:\t%f (%08x.)\n", sum.f, sum.i);

  vsum = _mm_setzero_ps();
  vi = _mm_set_ps(0.01f, 0.02f, 0.03f, 0.04f);
  for (i = 0; i < 100 / 4; i++) {
    vsum = _mm_add_ps(vsum, vi);
    vi = _mm_add_ps(vi, vinc);
  }
  vsum = _mm_add_ps(_mm_unpacklo_ps(vsum ,vsum),
                    _mm_unpackhi_ps(vsum ,vsum));
  vsum = _mm_add_ps(_mm_unpacklo_ps(vsum ,vsum),
                    _mm_unpackhi_ps(vsum ,vsum));
  _mm_store1_ps(&sum.f, vsum);
  printf("SIMD incre:\t%f (%08x.)\n", sum.f, sum.i);

  vsum = _mm_setzero_ps();
  vi = _mm_set_ps(1.0f, 0.99f, 0.98f, 0.97f);
  for (i = 0; i < 100 / 4; i++) {
    vsum = _mm_add_ps(vsum, vi);
    vi = _mm_sub_ps(vi, vinc);
  }
  vsum = _mm_add_ps(_mm_unpacklo_ps(vsum ,vsum),
                    _mm_unpackhi_ps(vsum ,vsum));
  vsum = _mm_add_ps(_mm_unpacklo_ps(vsum ,vsum),
                    _mm_unpackhi_ps(vsum ,vsum));
  _mm_store1_ps(&sum.f, vsum);
  printf("SIMD decre:\t%f (%08x.)\n", sum.f, sum.i);

  return 0;
}