Mario Lezcano Casado

We would accept a fix following https://github.com/pytorch/pytorch/issues/92141#issuecomment-1382241971

What about the last two files? Also, cc @fritzo for confirmation on whether we can use `log1p` on distributions. I would say it'd be very much desirable to use it...

This is the only PR left to close https://github.com/pytorch/pytorch/issues/91122. PTAL @ngimel

That path is vectorised, so it used vectorised CPU operations. Have a look at how they are implemented within `aten/src/ATen/cpu/vec`. Fixing those while keeping a not-too-bad performance is going to...

You'd still need to fix the AVX2 and AVX512 implementations of div accordingly. And sure, you can use hypot there, that may be faster. When you do that, it'd still...

At any rate, I'd still suggest first merging this PR, and then fixing the vectorised path on a follow-up PR. If what you are proposing is to implement `div` as...

This PR already fixed some operations, so you can remove the xfails on those! Also, it seems to be failing in Windows, so we need to fix those.

To see the different tests failures, see https://hud.pytorch.org/pr/92539 and select the commit you'd want to inspect.

mac builds are complaining as well. It looks like this is a "non-standard extension" of the standard, as [`std::abs` is not defined to be constexpr](https://en.cppreference.com/w/cpp/numeric/math/fabs) (ugh). Could you check whether...

I'd keep the `std::abs` for gcc as it's the most used compiler for PyTorch and it generates the fastest intel instruction. Otherwise, yep, let's fall back to the other implementation.