SIMD: Почему SSE RGB преобразование цвета в YUV примерно с той же скоростью, что и реализация в C ++?

Я только что попытался оптимизировать преобразователь RGB в YUV420. Использование таблицы поиска привело к увеличению скорости, как и использование арифметики с фиксированной запятой. Однако я ожидал реальной выгоды, используя инструкции SSE. Моя первая попытка привела к тому, что код стал медленнее, и после объединения всех операций скорость примерно такая же, как и у исходного кода. Что-то не так в моей реализации или инструкции SSE просто не подходят для поставленной задачи?

Далее следует фрагмент исходного кода:

#define RRGB24YUVCI2_00   0.299
#define RRGB24YUVCI2_01   0.587
#define RRGB24YUVCI2_02   0.114
#define RRGB24YUVCI2_10  -0.147
#define RRGB24YUVCI2_11  -0.289
#define RRGB24YUVCI2_12   0.436
#define RRGB24YUVCI2_20   0.615
#define RRGB24YUVCI2_21  -0.515
#define RRGB24YUVCI2_22  -0.100

void RealRGB24toYUV420Converter::Convert(void* pRgb, void* pY, void* pU, void* pV)
{
  yuvType* py = (yuvType *)pY;
  yuvType* pu = (yuvType *)pU;
  yuvType* pv = (yuvType *)pV;
  unsigned char* src = (unsigned char *)pRgb;

  /// Y have range 0..255, U & V have range -128..127.
  double u,v;
  double r,g,b;

  /// Step in 2x2 pel blocks. (4 pels per block).
  int xBlks = _width >> 1;
  int yBlks = _height >> 1;
  for(int yb = 0; yb < yBlks; yb++)
  for(int xb = 0; xb < xBlks; xb++)
  {
    int chrOff = yb*xBlks + xb;
    int lumOff = (yb*_width + xb) << 1;
    unsigned char* t    = src + lumOff*3;

    /// Top left pel.
    b = (double)(*t++);
    g = (double)(*t++);
    r = (double)(*t++);
    py[lumOff] = (yuvType)RRGB24YUVCI2_RANGECHECK_0TO255((int)(0.5 + RRGB24YUVCI2_00*r + RRGB24YUVCI2_01*g + RRGB24YUVCI2_02*b));

    u = RRGB24YUVCI2_10*r + RRGB24YUVCI2_11*g + RRGB24YUVCI2_12*b;
    v = RRGB24YUVCI2_20*r + RRGB24YUVCI2_21*g + RRGB24YUVCI2_22*b;

    /// Top right pel.
    b = (double)(*t++);
    g = (double)(*t++);
    r = (double)(*t++);
    py[lumOff+1] = (yuvType)RRGB24YUVCI2_RANGECHECK_0TO255((int)(0.5 + RRGB24YUVCI2_00*r + RRGB24YUVCI2_01*g + RRGB24YUVCI2_02*b));

    u += RRGB24YUVCI2_10*r + RRGB24YUVCI2_11*g + RRGB24YUVCI2_12*b;
    v += RRGB24YUVCI2_20*r + RRGB24YUVCI2_21*g + RRGB24YUVCI2_22*b;

    lumOff += _width;
    t = t + _width*3 - 6;
    /// Bottom left pel.
    b = (double)(*t++);
    g = (double)(*t++);
    r = (double)(*t++);
    py[lumOff] = (yuvType)RRGB24YUVCI2_RANGECHECK_0TO255((int)(0.5 + RRGB24YUVCI2_00*r + RRGB24YUVCI2_01*g + RRGB24YUVCI2_02*b));

    u += RRGB24YUVCI2_10*r + RRGB24YUVCI2_11*g + RRGB24YUVCI2_12*b;
    v += RRGB24YUVCI2_20*r + RRGB24YUVCI2_21*g + RRGB24YUVCI2_22*b;

    /// Bottom right pel.
    b = (double)(*t++);
    g = (double)(*t++);
    r = (double)(*t++);
    py[lumOff+1] = (yuvType)RRGB24YUVCI2_RANGECHECK_0TO255((int)(0.5 + RRGB24YUVCI2_00*r + RRGB24YUVCI2_01*g + RRGB24YUVCI2_02*b));

    u += RRGB24YUVCI2_10*r + RRGB24YUVCI2_11*g + RRGB24YUVCI2_12*b;
    v += RRGB24YUVCI2_20*r + RRGB24YUVCI2_21*g + RRGB24YUVCI2_22*b;

    /// Average the 4 chr values.
    int iu = (int)u;
    int iv = (int)v;
    if(iu < 0) ///< Rounding.
      iu -= 2;
    else
      iu += 2;
    if(iv < 0) ///< Rounding.
      iv -= 2;
    else
      iv += 2;

    pu[chrOff] = (yuvType)( _chrOff + RRGB24YUVCI2_RANGECHECK_N128TO127(iu/4) );
    pv[chrOff] = (yuvType)( _chrOff + RRGB24YUVCI2_RANGECHECK_N128TO127(iv/4) );
  }//end for xb & yb...
}//end Convert.

А вот версия с использованием SSE

const float fRRGB24YUVCI2_00 = 0.299;
const float fRRGB24YUVCI2_01 = 0.587;
const float fRRGB24YUVCI2_02 = 0.114;
const float fRRGB24YUVCI2_10 = -0.147;
const float fRRGB24YUVCI2_11 = -0.289;
const float fRRGB24YUVCI2_12 = 0.436;
const float fRRGB24YUVCI2_20 = 0.615;
const float fRRGB24YUVCI2_21 = -0.515;
const float fRRGB24YUVCI2_22 = -0.100;

void RealRGB24toYUV420Converter::Convert(void* pRgb, void* pY, void* pU, void* pV)
{
   __m128 xmm_y = _mm_loadu_ps(fCOEFF_0);
   __m128 xmm_u = _mm_loadu_ps(fCOEFF_1);
   __m128 xmm_v = _mm_loadu_ps(fCOEFF_2);

   yuvType* py = (yuvType *)pY;
   yuvType* pu = (yuvType *)pU;
   yuvType* pv = (yuvType *)pV;
   unsigned char* src = (unsigned char *)pRgb;

   /// Y have range 0..255, U & V have range -128..127.
   float bgr1[4];
   bgr1[3] = 0.0;
   float bgr2[4];
   bgr2[3] = 0.0;
   float bgr3[4];
   bgr3[3] = 0.0;
   float bgr4[4];
   bgr4[3] = 0.0;

   /// Step in 2x2 pel blocks. (4 pels per block).
   int xBlks = _width >> 1;
   int yBlks = _height >> 1;
   for(int yb = 0; yb < yBlks; yb++)
     for(int xb = 0; xb < xBlks; xb++)
     {
       int       chrOff = yb*xBlks + xb;
       int       lumOff = (yb*_width + xb) << 1;
       unsigned char* t    = src + lumOff*3;

       bgr1[2] = (float)*t++;
       bgr1[1] = (float)*t++;
       bgr1[0] = (float)*t++;
       bgr2[2] = (float)*t++;
       bgr2[1] = (float)*t++;
       bgr2[0] = (float)*t++;
       t = t + _width*3 - 6;
       bgr3[2] = (float)*t++;
       bgr3[1] = (float)*t++;
       bgr3[0] = (float)*t++;
       bgr4[2] = (float)*t++;
       bgr4[1] = (float)*t++;
       bgr4[0] = (float)*t++;
       __m128 xmm1 = _mm_loadu_ps(bgr1);
       __m128 xmm2 = _mm_loadu_ps(bgr2);
       __m128 xmm3 = _mm_loadu_ps(bgr3);
       __m128 xmm4 = _mm_loadu_ps(bgr4);

       // Y
       __m128 xmm_res_y = _mm_mul_ps(xmm1, xmm_y);
       py[lumOff] = (yuvType)RRGB24YUVCI2_RANGECHECK_0TO255((xmm_res_y.m128_f32[0] + xmm_res_y.m128_f32[1] + xmm_res_y.m128_f32[2] ));
       // Y
       xmm_res_y = _mm_mul_ps(xmm2, xmm_y);
       py[lumOff + 1] = (yuvType)RRGB24YUVCI2_RANGECHECK_0TO255((xmm_res_y.m128_f32[0]    + xmm_res_y.m128_f32[1] + xmm_res_y.m128_f32[2] ));
       lumOff += _width;
       // Y
       xmm_res_y = _mm_mul_ps(xmm3, xmm_y);
       py[lumOff] = (yuvType)RRGB24YUVCI2_RANGECHECK_0TO255((xmm_res_y.m128_f32[0] + xmm_res_y.m128_f32[1] + xmm_res_y.m128_f32[2] ));
       // Y
       xmm_res_y = _mm_mul_ps(xmm4, xmm_y);
       py[lumOff+1] = (yuvType)RRGB24YUVCI2_RANGECHECK_0TO255((xmm_res_y.m128_f32[0] + xmm_res_y.m128_f32[1] + xmm_res_y.m128_f32[2] ));

       // U
       __m128 xmm_res = _mm_add_ps(
                          _mm_add_ps(_mm_mul_ps(xmm1, xmm_u), _mm_mul_ps(xmm2, xmm_u)),
                          _mm_add_ps(_mm_mul_ps(xmm3, xmm_u), _mm_mul_ps(xmm4, xmm_u))
                       );

       float fU  = xmm_res.m128_f32[0] + xmm_res.m128_f32[1] + xmm_res.m128_f32[2];

       // V
       xmm_res = _mm_add_ps(
      _mm_add_ps(_mm_mul_ps(xmm1, xmm_v), _mm_mul_ps(xmm2, xmm_v)),
      _mm_add_ps(_mm_mul_ps(xmm3, xmm_v), _mm_mul_ps(xmm4, xmm_v))
      );
       float fV  = xmm_res.m128_f32[0] + xmm_res.m128_f32[1] + xmm_res.m128_f32[2];

       /// Average the 4 chr values.
       int iu = (int)fU;
       int iv = (int)fV;
       if(iu < 0) ///< Rounding.
         iu -= 2;
       else
         iu += 2;
       if(iv < 0) ///< Rounding.
         iv -= 2;
       else
         iv += 2;

       pu[chrOff] = (yuvType)( _chrOff + RRGB24YUVCI2_RANGECHECK_N128TO127(iu >> 2) );
       pv[chrOff] = (yuvType)( _chrOff + RRGB24YUVCI2_RANGECHECK_N128TO127(iv >> 2) );
     }//end for xb & yb...
}

Это одна из моих первых попытки SSE2, так что, возможно, я что-то упускаю? К вашему сведению, я работаю на платформе Windows с использованием Visual Studio 2008.