Runtime: Optimize System.Buffers using arm64 intrinsics

[BruceForstall]

This item tracks the conversion of the System.Buffers class to use arm64 intrinsics.

Related: #33308

[msftbot[bot]]

Tagging subscribers to this area: @tannergooding
Notify danmosemsft if you want to be subscribed.

[john-h-k]

I can give this a shot

[john-h-k]

These ARM64 intrinsics are missing to make the code "worthwhile":

Tracked by #32512: Done!

UnsignedSaturatingSubtract - vqsubq_u8 - UQSUB

Tracked by #33398:

UnsignedShiftRight - vshrq_n_u8 - USHR
UnsignedShiftLeft - vshlq_n_u8 - USHL

Tracked by #1277: Done!

TableLookup - vqtbl4q_u8 - TBL
TableLookupExtension - vqtbx4q_u8 - TBX

These helper methods could be accelerated to help it and are tracked in #33496:

Vector128.Create(byte)

These intrinsics could be introduced to maximise perf:

Load 3 Vector128<T> - vld3q_u8 - LD3
Store 3 Vector128<T> - vst3q_u8 - ST3
Load 4 Vector128<T> - vld4q_u8 - LD4
Store 4 Vector128<T> - vst4q_u8 - ST4

[john-h-k]

https://github.com/john-h-k/runtime/commit/adeea3662d12c2cc06a241736c34cb6e4f22f446

Draft commit, most that can be done until necessary intrinsics are added

[tannergooding]

Thanks @john-h-k, as per our conversation on Discord, I'll ping you once the necessary intrinsics are added so you can make further progress.

[BruceForstall]

@tannergooding @john-h-k The shift intrinsics went in with https://github.com/dotnet/runtime/pull/36830. There have been many improvements to System.Runtime.Intrinsics optimization in https://github.com/dotnet/runtime/issues/33496. Is this work now unblocked?

[tannergooding]

Yes, this should be unblocked now.

[BruceForstall]

@john-h-k Are you planning to try and finish this for .NET 5, now that it's unblocked?

[tannergooding]

If not, we'll have 3-4 people free to help tackle the remaining issues from https://github.com/dotnet/runtime/issues/33308 on Monday.

CC. @jeffhandley

[john-h-k]

Yes, I should be. I just need to benchmark the 2 possible impls I've got locally, either on my old Pi2b or I can get a friend to do it on newer device

[BruceForstall]

@tannergooding @jeffhandley What's the plan for this for .NET 5?

[jeffhandley]

@tannergooding is planning to carry this one over the finish line ahead of RC1.

[jeffhandley]

We could not get this completed in time for 5.0.0; moved to 6.0.0.