Support for overflow options in Integer cases

[arhik]

We need to support overflow options in integer case.

[AntonOresten]

Is there precedence for this in cutile-python? How might it work?

On a related note, I know at least that many operations like exp have e.g. flush_to_zero parameters for floats, but it's not currently exposed.

[arhik]

Is there precedence for this in cutile-python? How might it work?

I didn't look into cutile-python source code yet. I only followed tile-ir docs so far. And I didn't see any mention of explicit behaviour control over float cases in tile-ir docs. For signed and unsigned cases though we can provide explicit no_wrap or none and compiler do optimization in certain cases. Thought we might have to provide an option to control this integer type overflow behavior based on user use cases.

Another reason I raised is to hint on the nuance that we simply ignored overflow in integers case, which defaults to none, in PR #37.

I know at least that many operations like exp have e.g. flush_to_zero parameters for floats, but it's not currently exposed.

We may not have control over those yet, since I didn't see any mention of them in docs or docs are not updated yet.

[maleadt]

I just added an ct.@fpmode in #172 which could serve as a basis for ct.@intmode with overflow flags, however, AFAIU these are only optimizer hints, and don't actually control overflow behavior: https://docs.nvidia.com/cuda/tile-ir/latest/sections/operations.html#explicit-overflow-annotations

[maleadt]

Yeah I don't think this needs significant API surface. Let's see what cuTile Python does; for now I think we can close this.