Issue

Currently accscalar_t for float is double on the CPU:

https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/AccumulateType.h#L34

I suggest to have a small discussion whether it'd be better to switch to float.

Motivation

This is prompted by @wanchaol asking about a UndefinedBehaviorSanitizer: float-cast-overflow error, but I think there are three possible reasons to consider changing this:

• Consistency with GPU,
• support platforms on which float is faster (e.g. arm32)
• get rid of UB.

I seem to recall @apaszke preferring the current behaviour a year ago or so back (in the context of #6855, which always dispatched a double auxilliary function on CPU or so).

Pitch

Switch to float generally.

Alternatives

Switch to float only for specific platforms (e.g. arm32), somehow get rid of the UB warning.

Additional context

We (e.g. @ljk53 and me) might be interested in changing this for Android specifically if we don't generally.