6

我刚刚尝试优化 RGB 到 YUV420 转换器。使用查找表可以提高速度，使用定点算法也是如此。但是，我期待使用 SSE 指令获得真正的收益。我第一次尝试导致代码变慢，在链接所有操作之后，它的速度与原始代码大致相同。我的实现是否有问题，或者 SSE 指令不适合手头的任务？

一段原始代码如下：

#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.

这是使用 SSE 的版本

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...
}

这是我在 SSE2 的第一次尝试，所以也许我错过了什么？仅供参考，我正在使用 Visual Studio 2008 在 Windows 平台上工作。

4

3 回答 3

9

几个问题：

  • 您正在使用未对齐的负载 - 这些非常昂贵（除了在 Nehalem aka Core i5/Core i7 上） - 至少是对齐负载成本的 2 倍 - 如果您在负载之后有大量计算，则可以摊销成本但在这种情况你比较少。您可以通过使这些 16 字节对齐并使用对齐的负载来解决来自 bgr1、bgr2 等的负载的问题。[更好的是，根本不要使用这些中间数组——将数据直接从内存加载到 SSE 寄存器，并使用 SIMD 进行所有洗牌等——见下文]

  • 您将在标量代码和 SIMD 代码之间来回切换 - 就性能而言，标量代码可能是主要部分，因此任何 SIMD 收益都会被此淹没 - 您确实需要在循环中完成所有工作使用 SIMD 指令（即摆脱标量代码）

于 2011-01-28T14:19:03.037 回答
1

您可以使用内联汇编指令而不是 insintrics。它可能会稍微提高代码的速度。但是内联汇编是特定于编译器的。无论如何，正如 Paul R 在回答中所述，您必须使用对齐的数据才能达到全速。但是数据对齐是编译器特定的事情:)

如果您可以更改编译器，您可以尝试适用于 Windows 的 Intel 编译器。我怀疑它会好得多，尤其是对于内联汇编代码，但它绝对值得一看。

于 2011-01-28T14:45:39.490 回答
0

我发现您的方法存在一些问题：

  1. C++ 版本从指针 t 加载到“double r,g,b”，并且很可能，编译器已将这些优化为直接加载到 FP 寄存器，即“double r,g,b”在运行时存在于寄存器中时间。但是在您的版本中，您加载到“float bgr0/1/2/3”，然后调用_mm_loadu_ps。如果“float bgr0/1/2/3”在内存中，我不会感到惊讶，这意味着您对内存有额外的读写操作。

  2. 您使用的是内在函数而不是内联汇编。这些 __m128 变量中的一些（如果不是全部）可能仍在内存中。同样，额外的内存读取和写入。

  3. 大多数工作可能是在 RRGB24YUVCI2_*() 中完成的，您并没有尝试优化这些。

您没有对齐任何变量，但这只是对额外内存访问的额外惩罚，请先尝试消除这些。

您最好的选择是找到一个现有的、优化的 RGB/YUV 转换库并使用它。

于 2011-04-22T20:34:12.437 回答