Subgroups 🏎️

[FL33TW00D]

Looks like subgroups are going to land imminently on wgpu as per here: gfx-rs/wgpu#5301

Integrating these should be fun. Few thoughts on what should be done here:

1. We need some kind of native vs web feature flag, since subgroups haven't shipped on web yet.
2. We need to identify which kernels can benefit (hint. anything with a reduction)
3. Write up some benchmarks on https://github.com/FL33TW00D/wgpu-bench

References:
https://fleetwood.dev/posts/layernorm-as-fast-as-possible
https://github.com/FL33TW00D/wgpu-bench/blob/master/kernels/layernorm/welford_vec4.wgsl#L47

[AmineDiro]

In addition to reduce ops, Subgroups can also be useful for warp aware compute like warptiling

[FL33TW00D]

In addition to reduce ops, Subgroups can also be useful for warp aware compute like warptiling

Yeah I think the reduction would do nicely in the new GEMV kernels 🔥

[FL33TW00D]

Completed in #220