gistfile1.cpp

  __m128 Uint32_Float_Round(__m128i in)
  {
    // Generate mask when input elements are > 2^31 - 1
    __m128i mask = _mm_srai_epi32(in, 31);
    
    // Version of the input shifted down one bit logically
    __m128i t   = _mm_srli_epi32(in, 1);
    // Add lowest bit of each word to round.
    __m128i in1 = _mm_or_epi32(t, _mm_and_si128(in, _mm_set1_epi32(1)));

    // Select either shifted or unshifted version based on mask
    __m128 a = _mm_or_si128(_mm_and_si128(mask, in1),
                            _mm_andnot_si128(mask, in));

    // Convert to float
    __m128 b = _mm_cvtepi32_ps(a);

    // Correct magnitude of large elements by adding in the right elements again.
    __m128 c = _mm_and_ps(mask, b);

    __m128 result = _mm_add_ps(c, b);

    return result;
  }