We currently sample in bulk, but do not fetch features in bulk. This does not take full advantage of PCI-e bandwidth and, in some cases, the memory capacity of the GPU. This should be added as an option, and should possibly be the default means of feature fetching if performance is good enough.
Implementing this will require implementation of a new buffer that stores fetched features for a set of bulk samples.
We currently sample in bulk, but do not fetch features in bulk. This does not take full advantage of PCI-e bandwidth and, in some cases, the memory capacity of the GPU. This should be added as an option, and should possibly be the default means of feature fetching if performance is good enough.
Implementing this will require implementation of a new buffer that stores fetched features for a set of bulk samples.