hardshrink_cpu and hardshrink_backward_cpu refactoring with at::native::cpu_kernel

[pbelevich]

No description provided.

[fmassa]

We can potentially get better performance tha is to vectorization if you move this code to the cpu folder, see binary ops there for an example

[VitalyFedyunin]

Please include performance comparison and measurements strategy as part of PR description.

[facebook-github-bot]

@pbelevich has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[pbelevich]

data = torch.randn(1000000)
timeit hardshrink(data)

CPU_tensor_apply2: 4.88 ms ± 213 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
cpu_kernel:         223 µs ± 2.72 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
OMP_NUM_THREADS=1:  3.9 ms ± 192 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
cpu_kernel_vec:    40.1 µs ± 7.64 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
OMP_NUM_THREADS=1:  336 µs ± 2.85 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

[VitalyFedyunin]

Please undo submodule changes

[pbelevich]

@pytorchbot retest this please

[pbelevich]

@pytorchbot retest this please

[VitalyFedyunin]

Everything looks good, you may or may not apply additional changes request before landing.

[pbelevich]

@pytorchbot retest this please

[facebook-github-bot]

@pbelevich is landing this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[facebook-github-bot]

@pbelevich merged this pull request in bcfa023.