The c++ extension tutorial uses torch::PackedTensorAccessor32 and states
This is important as using the 64-bit variant (
PackedAccessor64) can make the kernel slower
But the github code it links to uses torch::PackedTensorAccessor. The code is supposed to be the same in both. Which one is correct?
Update: based on compile warnings it looks like the tutorial is correct and the github code is out of date.