Ok. Not doing blending at the moment. 1 joint per vertex only
This is what I have so far:
- vertices transformed by the inverse of it''s bone''s reference matrix
- vertex normals
- vertex texcoords
- triangles (indices into the vertex array)
- joints
- position and rotation in a matrix - this is the one by the inverse of which the vertices are multiplied
- each joint has it''s local transformation matrix
- and it''s vertex transformation matrix
- each joint also has keyframes with rotations stored in them
What I am doing for each bone each frame is:
1. Interpolate between 2 rotation keyframes
2. Create a matrix from the interpolated rotation
3. Postmultiply this bone''s local matrix with the one created in step 2
4. Go up the hierarcy and postmultiply the parent''s matrix with the child''s and put the result in the vertex transformation matrix
I end up with one vertex transformation matrix per bone.
Now for each vertex I just multiply it by it''s bone''s vertex transformation matrix.
Then I assign vertex array pointers and draw. This works ok, but it has some drawbacks:
1. Vertex X Matrix transformation is done on the CPU
2. I constantly have to modify the vertex array
Now I was thinking about doing this in vertex shaders.
I could have a static vertex array pretransformed by the inverse matrix. Then I would only have to calculate each bone''s vertex transformation matrix on the CPU (from the keyframes). I would upload all bones'' matrices into the vertex shader''s constant registers. I have 10 bones, each has a 4x4 matrix, but only the first 3 lines are needed. So that means 30 registers. Could be done. Now the shader would do the matrix multiplication to transform the vertices and output the result.
This would mean unloading the CPU for other tasks and also save quite a lot of AGP bandwidth, because I wouldn''t have to change the vertex array all the time.
Am I on the right track?
