Runtime: Optimize System.Collections.BitArray using arm64 intrinsics

Created on 6 Mar 2020 · 6Comments · Source: dotnet/runtime

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

arch-arm64 area-System.Collections

Source

BruceForstall

👍1

All 6 comments

I think I can put up a PR for this, but we seem to be lacking a couple instructions that would be nice to have:

Unloading values of single SIMD register where the vector contains a single element (e.g. Vector64 returned from AddAcross) to non-vector registers - vduph_laneq_s16 etc?
Unloading values of single SIMD register to a memory location ("Store") - vst1q_s16 etc?

Gnbrkm41 on 13 Mar 2020

@Gnbrkm41

Unloading values of single SIMD register where the vector contains a single element (e.g. Vector64 returned from AddAcross) to non-vector registers - vduph_laneq_s16 etc?

I believe this should be supported via ToScalar() method. cc @tannergooding to confirm

Unloading values of single SIMD register to a memory location ("Store") - vst1q_s16 etc?

I am working on these https://github.com/dotnet/runtime/pull/33461 https://github.com/dotnet/runtime/pull/33535

echesakovMSFT on 13 Mar 2020

👍1

Ah, did not notice ToScalar extension method 🙂. I did manage to work around using a different approach which probably is better.

It still would be nice to have those exposed as intrinsics for feature parity, though.

Gnbrkm41 on 13 Mar 2020

It still would be nice to have those exposed as intrinsics for feature parity, though.

For some cases, there isn't a corresponding intrinsic on the native side. Where there is, we also expose it there (for example int Sse2.ConvertToInt32(Vector128<int> value) on the x86 side, which corresponds to MOVD reg/m32, xmm).