Видео

Thanks! In a conventional rasterizer, skinning can be computed in the vertex shader. This is highly efficient since only one mesh is needed on the GPU, and during rendering, each character instance maps its vertices to the skeleton matrices. However, in a raytracer (to my knowledge), there is no vertex shader. Therefore, each character instance requires its own mesh and BLAS. This means that, similar to the rasterizer, I compute per instance the current animation pose and his skin matrices and then i invoke a compute shader to transform the static mesh into the current pose. Afterward, the BLAS is updated. Since this process takes time for many character instances, I only update each instance every x frames. The main issue is that each character instance needs a separate copy of the mesh, which limits the number of characters that can be displayed by the GPU's memory. As a result, large crowds are not possible with this approach. However, since the camera would be far away in such scenarios, alternative techniques can be used to bypass this limitation.