Optimize zero length input (#31602)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/31602
Pull Request resolved: https://github.com/pytorch/glow/pull/3943
Zero length input is something we hit fairly frequently in practice. Previous handling of global TensorPool involves two locks per input (acquire and reclaim). Here we use a specialized anchor tensor to host zero length input. Note that it is only padded to max sequence length. If necessary, an easy extension can be added to pad to max `InputPlaceholder.getType().size()`.
Reviewed By: jfix71
Differential Revision: D19192467
fbshipit-source-id: cafdc1eb7bf9b9d6ead04a0243b0be838f6b71cd
