Updating accel structure from GPU

[cat-state]

Hi, is it possible to efficiently update the positions of procedural primitives in the accel structure from device side kernels? In the examples I only see building the structure up front once and then reusing it. My use-case is to reuse the fast ray-aabb tests for collision detection and I have a realtime gpu physics simulation id port to luisa if that were possible :)

[Mike-Leo-Smith]

You can launch a kernel to update the AABB buffer and then encode into the stream a ProceduralPrimitive::build() command and an Accel::build() command.

[cat-state]

oh ofc! thank you for the fast response :)
before i try porting everything - do you know how many AABBs would be possible to update at 100Hz or so (on best hardware possible)?

[Mike-Leo-Smith]

I don’t have the specific data for now but I think several million won’t be a problem. I’ll run a quick test today or tomorrow.

[Mike-Leo-Smith]

I have run the tests on my 4060 laptop (70W TDP for GPU), where I randomly set the center of AABBs every frame and update the accel. For 100Hz update, if the accel is created with the AccelOption::allow_update flag on, the number is over 4M; if the option is off (so that the acceleration structure is forced to rebuild at every frame for better quality), the number is over 1M. I guess on better hardware like a desktop 4090, the numbers would further grow several times.