CPU: feature flag u8 and f32 pipelines

[nicoburns]

Adds feature flags to vello_cpus u8 and f32 pipelines so that you can save on binary size if you're only planning on using one of them. Disabling the f32 pipeline knocks ~60kb off the binary size for me ~250kb -> ~190kb.

[LaurenzV]

I’m not sure locking enum variants behind a feature flag is a good idea because it might break existing matching code downstream if someone activates one of the features.

[nicoburns]

I’m not sure locking enum variants behind a feature flag is a good idea because it might break existing matching code downstream if someone activates one of the features.

Yeah, good point. We could mark the enum as non-exhaustive. But perhaps it's better to just keep both variants but ignore the value passed if only one pipeline is enabled? The enum values are "phrased" to be hint-like anyway...

[LaurenzV]

Yes I think ignoring is better.

[nicoburns]

@DJMcNab Any ideas on how to setup CI for this:

• Two new features u8_pipeline and f32_pipeline.
• Only exist for vello_cpu crate
• Should ideally be a compile error if neither of them is specified

Perhaps we need to exclude vello_cpu from the workspace cargo-hack and test it separately?

[DJMcNab]

I think it should be able to be treated in the same way that we treat std and libm.

[nicoburns]

I have fixed CI by enabling both pipelines everywhere we're not explicitly checking that it compiles with just one of them enabled, and I've added a compile_error! for the case where neither pipeline is enabled.

[LaurenzV]

It seems like there are still failures

[nicoburns]

It seems like there are still failures

Ah, I think the compile_error! when neither feature is enabled reintroduced CI failures. Fixed.

[LaurenzV]

I'm wondering whether it make sense to only enable the u8 pipeline by default. IMO, that's pretty much always the one you should use. But not sure.

[nicoburns]

I'm wondering whether it make sense to only enable the u8 pipeline by default. IMO, that's pretty much always the one you should use. But not sure.

I would agree with this. I've made the f32_pipeline feature non-default.

[LaurenzV]

A bit annoying to make this work with CI, but it is what it is. 😄

[nicoburns]

A bit annoying to make this work with CI, but it is what it is. 😄

There is an --at-least-one-of flag to cargo hack that looks like it would make it nicer. But that requires changing to --feature-powerset rather than using --each-feature which seems like it would be a much bigger change.