用SSE加速CPU蒙皮计算

http://blog.csdn.net/garuda/article/details/6539271

我们知道现在绝大多数情况下角色动画的蒙皮计算是放在GPU中计算的。
但是仍然有一些特殊的场合我们需要在CPU端使用蒙皮计算的结果，比如涉及到
布料的物理模拟的时候。这时我们需要在CPU端计算蒙皮。
为了节约宝贵的CPU计算的时间，我们需要用SSE对CPU计算蒙皮进行加速。

 

顶点结构如下：
struct Vertex
{
 float3 Pos;
 float3 Normal;
 int  n;   //该顶点蒙到了几个骨骼上
 int  BoneId[4];
 float Weight[3];
 .... //切线等等
};

 

对于每个顶点的计算过程是先由BoneId和Weight计算出变换矩阵，然后用矩阵变换
顶点的位置，法线等。其中求变换矩阵是对多个矩阵加权求和的过程，这一步很适合
使用SSE加速。
每个顶点可能蒙到1-4个骨骼上，也就是每个顶点可能需要对1-4个矩阵加权求和。
1个骨骼的时候不用计算。2-4个骨骼的时候，我们分别写3个函数对应这3种情况。

 

__forceinline void LoadFourFloats(float* a0, __m128& res)
{
 res = _mm_load_ps(a0);
}

__forceinline void StoreFourFloats(float* a0, const __m128& src)
{
 _mm_store_ps(a0, src);
}

__forceinline void MulMatrixFloat(__m128& mo0, __m128& mo1, __m128& mo2,
          const __m128& mi0, const __m128& mi1, const __m128& mi2,
          float w)
{
 __m128 xmm;

 xmm = _mm_load_ss(&w);
 xmm = _mm_shuffle_ps(xmm,xmm,0);

 // Multiply matrix 1 by weight 1.
 mo0 = _mm_mul_ps(xmm, mi0);
 mo1 = _mm_mul_ps(xmm, mi1);
 mo2 = _mm_mul_ps(xmm, mi2);
}

__forceinline void Collapse2MatSSE(float* pM1, float* pM2, 
           float W1, float W2, float* pR)
{
 __m128 xmm1, xmm2, xmm3;
 __m128 xmm4, xmm5, xmm6;

 // Load matrix 1.
 LoadFourFloats(pM1 + 0, xmm1);
 LoadFourFloats(pM1 + 4, xmm2);
 LoadFourFloats(pM1 + 8, xmm3);
 MulMatrixFloat(xmm1, xmm2, xmm3, xmm1, xmm2, xmm3, W1);

 // Load matrix 2.
 LoadFourFloats(pM2 + 0, xmm4);
 LoadFourFloats(pM2 + 4, xmm5);
 LoadFourFloats(pM2 + 8, xmm6);
 MulMatrixFloat(xmm4, xmm5, xmm6, xmm4, xmm5, xmm6, W2);

 // Add matrix 1 to matrix 2.
 xmm1 = _mm_add_ps(xmm1, xmm4);
 xmm2 = _mm_add_ps(xmm2, xmm5);
 xmm3 = _mm_add_ps(xmm3, xmm6);

 StoreFourFloats(pR + 0, xmm1);
 StoreFourFloats(pR + 4, xmm2);
 StoreFourFloats(pR + 8, xmm3);
}

__forceinline void Collapse3MatSSE(float* pM1, float* pM2, float* pM3,
           float W1, float W2, float W3, float* pR)
{
 __m128 xmm1, xmm2, xmm3;
 __m128 xmm4, xmm5, xmm6;

 // Load matrix 1.
 LoadFourFloats(pM1 + 0, xmm1);
 LoadFourFloats(pM1 + 4, xmm2);
 LoadFourFloats(pM1 + 8, xmm3);
 MulMatrixFloat(xmm1, xmm2, xmm3, xmm1, xmm2, xmm3, W1);

 // Load matrix 2.
 LoadFourFloats(pM2 + 0, xmm4);
 LoadFourFloats(pM2 + 4, xmm5);
 LoadFourFloats(pM2 + 8, xmm6);
 MulMatrixFloat(xmm4, xmm5, xmm6, xmm4, xmm5, xmm6, W2);

 // Add matrix 1 to matrix 2.
 xmm1 = _mm_add_ps(xmm1, xmm4);
 xmm2 = _mm_add_ps(xmm2, xmm5);
 xmm3 = _mm_add_ps(xmm3, xmm6);

 // Load matrix 2.
 LoadFourFloats(pM3 + 0, xmm4);
 LoadFourFloats(pM3 + 4, xmm5);
 LoadFourFloats(pM3 + 8, xmm6);
 MulMatrixFloat(xmm4, xmm5, xmm6, xmm4, xmm5, xmm6, W3);

 // Add matrix 1 to matrix 2.
 xmm1 = _mm_add_ps(xmm1, xmm4);
 xmm2 = _mm_add_ps(xmm2, xmm5);
 xmm3 = _mm_add_ps(xmm3, xmm6);

 StoreFourFloats(pR + 0, xmm1);
 StoreFourFloats(pR + 4, xmm2);
 StoreFourFloats(pR + 8, xmm3);
}

__forceinline void Collapse4MatSSE(float* pM1, float* pM2, float* pM3, float* pM4,
           float W1, float W2, float W3, float W4, float* pR)
{
 __m128 xmm1, xmm2, xmm3;
 __m128 xmm4, xmm5, xmm6;

 // Load matrix 1.
 LoadFourFloats(pM1 + 0, xmm1);
 LoadFourFloats(pM1 + 4, xmm2);
 LoadFourFloats(pM1 + 8, xmm3);
 MulMatrixFloat(xmm1, xmm2, xmm3, xmm1, xmm2, xmm3, W1);

 // Load matrix 2.
 LoadFourFloats(pM2 + 0, xmm4);
 LoadFourFloats(pM2 + 4, xmm5);
 LoadFourFloats(pM2 + 8, xmm6);
 MulMatrixFloat(xmm4, xmm5, xmm6, xmm4, xmm5, xmm6, W2);

 // Add matrix 1 to matrix 2.
 xmm1 = _mm_add_ps(xmm1, xmm4);
 xmm2 = _mm_add_ps(xmm2, xmm5);
 xmm3 = _mm_add_ps(xmm3, xmm6);

 LoadFourFloats(pM3 + 0, xmm4);
 LoadFourFloats(pM3 + 4, xmm5);
 LoadFourFloats(pM3 + 8, xmm6);
 MulMatrixFloat(xmm4, xmm5, xmm6, xmm4, xmm5, xmm6, W3);

 xmm1 = _mm_add_ps(xmm1, xmm4);
 xmm2 = _mm_add_ps(xmm2, xmm5);
 xmm3 = _mm_add_ps(xmm3, xmm6);

 LoadFourFloats(pM4 + 0, xmm4);
 LoadFourFloats(pM4 + 4, xmm5);
 LoadFourFloats(pM4 + 8, xmm6);
 MulMatrixFloat(xmm4, xmm5, xmm6, xmm4, xmm5, xmm6, W4);

 xmm1 = _mm_add_ps(xmm1, xmm4);
 xmm2 = _mm_add_ps(xmm2, xmm5);
 xmm3 = _mm_add_ps(xmm3, xmm6);

 StoreFourFloats(pR + 0, xmm1);
 StoreFourFloats(pR + 4, xmm2);
 StoreFourFloats(pR + 8, xmm3);
}

计算顶点蒙皮矩阵的过程如下

// 为了使用sse优化，这个矩阵必须是16字节对齐的。
// release 编译时编译器会自动保证这一点，但
// 要debug正确运行必须加上内存对齐的声明.
__declspec(align(16)) Matrix4 matObj;

for (int i = 0; i < vertCount; i++, pVertex++)
{
 if (pVertex->n == 1)
 {
  matObj = *BoneMatrixPalette[pWeight->nBones[0]];
 }
 else if (pVertex->n == 2)
 {
  float* pMat0 = BoneMatrixPalette[pVertex->nBones[0]]->ToFloatPtr();
  float* pMat1 = BoneMatrixPalette[pVertex->nBones[1]]->ToFloatPtr();

  Collapse2MatSSE(pMat0, pMat1,pVertex->Weight[0], pVertex->Weight[1], matObj.ToFloatPtr());

 }
 else if (pVertex->n == 3)
 {
  float* pMat0 = BoneMatrixPalette[pWeight->nBones[0]]->ToFloatPtr();
  float* pMat1 = BoneMatrixPalette[pWeight->nBones[1]]->ToFloatPtr();
  float* pMat2 = BoneMatrixPalette[pWeight->nBones[2]]->ToFloatPtr();

  Collapse3MatSSE(pMat0, pMat1, pMat2,pVertex->Weight[0], pVertex->Weight[1], 
   pVertex->Weight[2], matObj.ToFloatPtr());
 }
 else if (pVertex->n == 4)
 {
  float* pMat0 = BoneMatrixPalette[pVertex->nBones[0]]->ToFloatPtr();
  float* pMat1 = BoneMatrixPalette[pVertex->nBones[1]]->ToFloatPtr();
  float* pMat2 = BoneMatrixPalette[pVertex->nBones[2]]->ToFloatPtr();
  float* pMat3 = BoneMatrixPalette[pVertex->nBones[3]]->ToFloatPtr();

  Collapse4MatSSE(pMat0, pMat1, pMat2, pMat3, pVertex->Weight[0], pVertex->Weight[1],
   pVertex->Weight[2], pVertex->Weight[3], matObj.ToFloatPtr());
 }
 else
 {
  assert(0);
 }
}

得到matObj后分别对等点位置、法线等做变换就可以了。这一过程仍然可以用SSE加速，但这需要顶点
的位置、法线等均是16字节对齐的。这需要较大改动，因此我们没有做。
经profile，加速后CPU蒙皮的速度提升了1倍。

 

主要参考

Optimized CPU-based Skinning for 3D Games

 

下面两篇来自id，更加变态的优化

Fast Skinning

The Skeleton Assembly Line