PyTorch ThroughputBenchmark: fix inaccuracy in number of iterations reporting (#22292)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/22292
as we do atomic fetch_add to validate if a thread should
finish, we should not take the last iteration into account. As a
result total number of iterations should be exactly the same as user
sets via config.num_iters
Now when running a unit test I see exact number of iterations reported
Differential Revision: D16023963
fbshipit-source-id: 3b12ee17276628ecd7b0979f28cd6deb777a1543