Debate: Should `never_partition_all()` should be the default?

[mcourteaux]

I find myself literring my code with this directive everywhere. Now that we have it, and we all were able to use it for a while, I am tempted to argue that it should be the default. Loop partitioning is quite an extensive transformation that ...

• is surprising for newcomers (especially when dealing with TailStrategies where the thing you expected is fully transformed away by duplicating the code: e.g., the expected logic for ShiftInwards (i.e. some min(loop_var, image_bound - vectorsize)) can fully disappear due to it)
• doing a lot of bad codegen, as the Auto strategy is not smart enough yet; especially on GPU.
• increases code size
• increases compile times
• is often unnecessary

Overall, I think that loop partitioning can be beneficial, but should be a conscious choice of the developer of the pipeline. My personal preference would go to loop partitioning being disabled by default; but as an alternative, I can see a lot of benefit into making the Auto strategy a lot smarter into deciding when to, and when not (which would change it from actually being EverywherePossible).

Opinions? Perhaps something for the dev-meeting?

[steven-johnson]

Was this discussed, and if so, were there any conclusions?

[mcourteaux]

I unfortunately could be there during the last dev-meeting. I don't know if it was discussed yet.

[abadams]

We can discuss it today, but it would tank the performance of a whole lot of existing production code by turning vector loads into vector gathers.

[abadams]

At the dev meeting we discussed making a target flag that changes the default

[mcourteaux]

it would tank the performance of a whole lot of existing production code by turning vector loads into vector gathers.

That would potentially be a good heuristic for the Auto strategy: does it turn load/stores into gathers/scatters. A simple TailStrategy::ShiftInwards would not suffer from loads/stores being turned into gather/scatters.