用SSE加速CPU蒙皮计算

朱雀 2022-07-14 10:59 234阅读 0赞

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][]

[Optimized CPU-based Skinning for 3D Games]: http://software.intel.com/en-us/articles/optimized-cpu-based-skinning-for-3d-games/
[Fast Skinning]: http://cache-www.intel.com/cd/00/00/29/37/293750_293750.pdf
[The Skeleton Assembly Line]: http://cache-www.intel.com/cd/00/00/29/37/293749_293749.pdf