I am testing the RigidBody dynamic of Bullet3 especially in the OpenCL module. A simple sphere scene with ~12k spheres sampled on a mesh and a plane as ground.
The simulation is only ~12k spheres free falling with gravity -980 m/s.
(Yah, my scene is in 0.01 metric-scale.) Broadphase, narrowphase and pipeline setups are all identical to the shipped example in version 2.87. There are some wild driftings on batches of spheres as shown in the attached video. This happens when I was using the GPU (1080Ti) to do the simulation. When I use the CPU (5930k) with the same setup in Bullet3's OpenCL implementation.
Anyone knows the reason? (Would it be my GPU is naughty to trick me up
)