Speed up TensorIterator::compute_strides a little (#22779)
Summary:
For three 1-D operands, compute_strides now takes 298 instructions instead
of 480. (Saves ~36 ns). We'll want to make Tensor::sizes(), strides(), and
element_size() trivially inlinable to speed this up more.
(Using PMCTest from https://www.agner.org/optimize/ to measure instructions retired)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/22779
Differential Revision: D16223595
Pulled By: colesbury
fbshipit-source-id: e4730755f29a0aea9cbc82c2d376a8e6a0c7bce8
