Port all and any full reductions to structured kernels.

[ysiraichi]

Small BC-breaking change for torch.all/any interaction with uint8 and zero-dimensional tensors

Before this PR:

>>> torch.all(torch.tensor(42, dtype=torch.uint8))
tensor(1, dtype=torch.uint8)
>>> torch.all(torch.tensor(42, dtype=torch.uint8), dim=0)
tensor(42, dtype=torch.uint8)

After this PR:

>>> torch.all(torch.tensor(42, dtype=torch.uint8))
tensor(1, dtype=torch.uint8)
>>> torch.all(torch.tensor(42, dtype=torch.uint8), dim=0)
tensor(1, dtype=torch.uint8)

This behavior is now consistent between torch.all(x) and torch.all(x, dim=0)

Stack from ghstack:

• Port all and any full reductions to structured kernels. #64642 Port all and any full reductions to structured kernels.

Tracking issue: #55070

This PR creates out overloads for both all and any kernels (full reduction overload),
and ports them to structured kernels.

Differential Revision: D30867354

cc @ezyang @gchanan

[facebook-github-bot]

🔗 Helpful links

• 🧪 See artifacts and rendered test results at hud.pytorch.org/pr/64642
• 📄 Preview docs built from this PR
• 📄 Preview C++ docs built from this PR
• 🔧 Opt-in to CIFlow to control what jobs run on your PRs

💊 CI failures summary and remediations

As of commit 7a0a3e9 (more details on the Dr. CI page):

---

💚 💚 Looks good so far! There are no failures yet. 💚 💚

---

This comment was automatically generated by Dr. CI (expand for details).

Follow this link to opt-out of these comments for your Pull Requests.

Please report bugs/suggestions to the (internal) Dr. CI Users group.

Click here to manually regenerate this comment.

[ysiraichi]

CI failure not related (see #64595).

[ezyang]

How come you're allowed to delete the calls to _dimreduce_return_trivial ?

[ezyang]

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

[ysiraichi]

How come you're allowed to delete the calls to _dimreduce_return_trivial ?

As far as I understand, that function does two things:

• resize the result tensor (the meta function takes care of it)
• fill the result tensor with the appropriate value (the _<name> function takes care of it)

There was one case missing: for self.numel() == 1 && self.ndimension() == 0. Therefore, I added the following to the _<name> functions: iter.numel() == 1 (I don't think we need to check for the dimension here, since we already did the resizing).

With this, I believe the implementation is equivalent.

[ysiraichi]

@ezyang @bdhirsh
Apparently, all and all.dim (same for any) had slightly different behavior for 0-dimensional tensors. Which should be the correct result when executing torch.all(torch.tensor(42, dtype=torch.uint8))?

(a)	(b)
tensor(1, dtype=torch.uint8)	tensor(42, dtype=torch.uint8)

Before this PR, all(tensor) returned (a), while all(tensor, dim=0) returned (b). In order to be consistent, I'm defaulting to (a) (feels right), but have no strong feelings about this one. What do you think?

[ezyang]

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

[ezyang]

I agree with your analysis. I'll mark this as BC-breaking just to be safe tho.

[ezyang]

@ngimel pointed out some relevant prior art #47878 but I think it doesn't apply here (as you are not changing the dtype)

[ngimel]

Right, and without changing dtype I agree that 1 is more consistent than non-1.

[facebook-github-bot]

@ezyang merged this pull request in 54d060a.