I continue to describe the skeletal animation algorithm as it is implemented in the Flame Steel Engine.
Since this is the most complex algorithm I’ve ever implemented, there may be errors in the development notes. In the previous article about this algorithm, I made a mistake: the bone array is passed to the shader for each mesh separately, not for the entire model.
Hierarchy of nodes
For the algorithm to work correctly, the model must contain a connection between the bones (graph). Let’s imagine a situation in which two animations are played simultaneously – a jump and raising the right hand. The jump animation must raise the model along the Y axis, while the animation of raising the hand must take this into account and rise together with the model in the jump, otherwise the hand will remain in place on its own.
Let’s describe the node connection for this case – the body contains a hand. When the algorithm is processed, the bone graph will be read, all animations will be taken into account with correct connections. In the model’s memory, the graph is stored separately from all animations, only to reflect the connectivity of the model’s bones.
Interpolation on CPU
In the previous article I described the principle of rendering skeletal animation – “transformation matrices are passed from the CPU to the shader at each rendering frame.”
Each rendering frame is processed on the CPU, for each bone of the mesh the engine gets the final transformation matrix using interpolation of position, rotation, magnification. During the interpolation of the final bone matrix, the node tree is traversed for all active node animations, the final matrix is multiplied with the parents, then sent to the vertex shader for rendering.
Vectors are used for position interpolation and magnification, quaternions are used for rotation, since they are very easy to interpolate (SLERP) unlike Euler angles, and they are also very easy to represent as a transformation matrix.
How to Simplify Implementation
To simplify debugging of vertex shader operation, I added simulation of vertex shader operation on CPU using macro FSGLOGLNEWAGERENDERER_CPU_BASED_VERTEX_MODS_ENABLED. Video card manufacturer NVIDIA has a utility for debugging shader code Nsight, perhaps it can also simplify development of complex algorithms of vertex/pixel shaders, however I never had a chance to check its functionality, simulation on CPU was enough.
In the next article I plan to describe mixing several animations, fill in the remaining gaps.
Sources
https://www.youtube.com/watch?v= f3Cr8Yx3GGA
