I am looking for either an algorithm or a library that can generate
matrix palettes for vertex shader skinning. e.g., given a maximum
palette size that can fit in a vertex shader, I want to create N
matrix palettes that can transform my skinned mesh, and I'd like for
the number of palettes generated to be as small as possible for
optimal performance. (fewer draw calls and matrix uploads to the
vertex shader)

I know that ID3DXSkinInfo::ConvertToIndexedBlendedMesh can do this,
however, going through Direct3D is a last resort. Has anyone seen
anything like this?