When there are multiple workers, if any model setting is updated in the root worker (relative_movement, relative_jacobian, adapt_movement_scale, or the initial frame), then only one of the workers will consume the update.
Right now this is incorrectly implemented using a single process pipe. I'm unclear if python allows multiple pipe pairs to be instantiated by a host process, but I either need to add one pipe per process or one unique queue per process, and have the broadcast_* methods on CUDAWorkerPool individually message each process via those channels.
When there are multiple workers, if any model setting is updated in the root worker (relative_movement, relative_jacobian, adapt_movement_scale, or the initial frame), then only one of the workers will consume the update.
Right now this is incorrectly implemented using a single process pipe. I'm unclear if python allows multiple pipe pairs to be instantiated by a host process, but I either need to add one pipe per process or one unique queue per process, and have the broadcast_* methods on CUDAWorkerPool individually message each process via those channels.