This could use a more detailed description, if there is more context around this change.

@jeffreyksmithjr Thanks, I updated the description. I don't have much more to say -- this is a pretty general performance improvement that is self-explanatory.

@VitalyFedyunin Thanks, this is really good to know. I will try to look into how to make many CPU functions to the native/cpu folder.

Do you happen to know where I can systematically learn these kinds of information for PyTorch internals? It would be quite helpful if such a document is available.

We are working on better doc now, so far you can look at aten/src/ATen/native/cpu/README.md

Also it is reasonable to attach benchmark methodology and results, especially if you are working on performance related tasks.

@VitalyFedyunin I looked into the code again. I kind of what to change the implementation to make use of arange_out instead of implementing once again, and leave all possible future optimization (if it will ever happen) to arange_out. I have changed the current code to follow this thought. What do you think?

But I only did this for stride == 1 -- I looked into the code of range_cpu_out, and it seems to assume that stride is always 1. I'm a bit confused: Is it same to assume stride to be 1, or arange_out didn't do it correctly?

@VitalyFedyunin Do you mind if you could take a look at the recent commits and land the PR? Thanks.

Sure you can, go ahead with it. Even previous state was fine. I even recommend to remove this ==1 check, as It is implemented incorrectly. What I meant by previous comment is something like:

if (r__stride == 1) {
 at::parallel_for(0, n, internal::GRAIN_SIZE,
                    [&r__data, &r__stride_0](int64_t p_begin, int64_t p_end) {
      for(int64_t i = p_begin; i < p_end; i++)
        r__data[i] = static_cast<scalar_t>(i);
    }); 
}
else {
 at::parallel_for(0, n, internal::GRAIN_SIZE,
                    [&r__data, &r__stride_0](int64_t p_begin, int64_t p_end) {
      for(int64_t i = p_begin; i < p_end; i++)
        r__data[i*r__stride_0] = static_cast<scalar_t>(i);
    });
}

Which helps to avoid multiplication.

But previous version with just:

 at::parallel_for(0, n, internal::GRAIN_SIZE,
                    [&r__data, &r__stride_0](int64_t p_begin, int64_t p_end) {
      for(int64_t i = p_begin; i < p_end; i++)
        r__data[i*r__stride_0] = static_cast<scalar_t>(i);
    });

is also good

Sure, restored.

@pytorchbot rebase this please

@VitalyFedyunin merged this pull request in 0702b5f.