th::cuda_bonedeform::1.0OpenCL when CUDA is not there
Each GPU node has a GPU Backend setting: Automatic (the default) runs CUDA on an NVIDIA RTX 20 series card or newer, and OpenCL on any other card, with a note on the node. Inside a rig, the mesh stays on the graphics card from bone deform into delta mush.
- Runs both ways: joint capture, bone deform, delta mush, elastic delta mush, blendshapes, lattice, wrap, shrinkwrap, rivet, normals, subdivide, wrinkles, Solidify (ARAP), attribute copy, attribute transfer and attribute math.
- GeForce RTX 20 and GTX 16 cards now run the CUDA path too.
- A card that can run neither now stops with a readable message on the node instead of a wrong result.
- Checked on NVIDIA's OpenCL against CUDA: the results match to within 0.00001 units, and no node is slower.



















































