Hello again

My game is coming along. I now have a humanoid with 120 Bones. It has 8 Mesh parts, each part is only influenced by maximum 30 Bones of the total skeleton. I do know how to animate this but I am unsure which way I should do it.

My problem is that as of now a vertex may index to bone0 and bone119, or any in between. Which means I have to map out which Bones influence each Mesh Part, copy those bones to a local Matrix Palette for that Mesh Part and THEN I have to convert the bone indices for each vertex to the new local palette.

Instead of doing this Skinned Model Instancing came to mind: http://www.ionixxgam...lInstancing.pdf

This paper describes how Textures can hold a large number of bones encoded in them.

My thought is then to code the entire animated skeleton to a texture, pass it to the shader(60 times a second), and use it as an array of many bones (120+).

Is this worth the effort, can it hurt performance? Are there a better ways?


I welcome all comments.

Would it be reasonable or even possible to try it with DX9?

Here is how I imagine I would split up the Animated Skeleton the "old" way


D3DXMATRIX* pAnimatedBones = new D3DXMATRIX[120];
/*
Combine Transforms and populate pAnimatedBones
*/
struct MeshPart
{
D3DXMATRIX* pLocalBones;
};

//Another place
MeshPart Torso;
Torso.pLocalBones = new D3DXMATRIX[30];

for(int i=0;i<30;i++)
{
Torso.pLocalBones = pAnimatedBones[GetMappedIndex(i,Torso)];
}

//During init vertices will have their bone indices remapped


Would it be reasonable or even possible to try it with DX9?

Here is how I imagine I would split up the Animated Skeleton the "old" way


#define hTorso 3

D3DXMATRIX* pAnimatedBones = new D3DXMATRIX[120];
/*
Combine Transforms and populate pAnimatedBones
*/
struct MeshPart
{
D3DXMATRIX* pLocalBones;
};

//Another place
MeshPart Torso;
Torso.pLocalBones = new D3DXMATRIX[30];

for(int i=0;i<30;i++)
{
Torso.pLocalBones = pAnimatedBones[GetMappedIndex(i,hTorso)];
}

//During init vertices will have their bone indices remapped


Sorry for the bump but I need some pointers. I want to implement this so it runs in DX9.0c/WinXP.

I create a texture using the D3DFMT_A8R8G8B8 flag in managed pool. 120 Bones = 480 pixels. Room for 8 bone hierarchies per row. So animated texture width is 3840 pixels, and height is 4096. This gives me room for 32768 instances on a 3840x4096 texture.

A constant register is updated with the SkeletonID before each DrawEntity.

This pseudo code demonstrates how I copy matrices into the texture:

//Once
D3DXVECTOR4 *pAnimatedMatrices = new D3DXVECTOR4[3840][4096]
//Update matrices

//Each Render Frame
hr = pAnimTex->LockRect(0,&rc,NULL,0);
if(hr==D3D_OK){
CopyMemory(rc.pBits,pAnimatedMatrices,3840*4096*4);
}
pAnimTex->UnlockRect(0);


Before drawing I set Texture:

SetTexture(D3DVERTEXTEXTURESAMPLER0, pAnimTex);



HLSL:

const uniform float2 SkeletonID;
sampler2D tex;
float4x4 Bone(tex2D(tex, SkeletonID),tex2D(tex, SkeletonID+float2(4,0)),tex2D(tex, SkeletonID+float2(8,0)),tex2D(tex, SkeletonID+float2(12,0)));
//Transform vertex



I am treading in deep waters here. Am I going in the right direction? Any shortcomings besides the huge amount of Draw Calls and DX9 instead of DX10 etc..?