Implement Neon SIMD

[CryZe]

This ports the WASM algorithm over to Aarch64 / ARM Neon. Rust itself isn't quite ready yet, but this should start compiling soon.

Blocked on: rust-lang/stdarch#1230

[pthariensflame]

I think those transmutes should probably be vld1qs?

[CryZe]

They totally could be, but ptr::read_unaligned also should work just fine across the board (and is platform independent, so that makes the code easier to port to a different platform). Also this issue where vld* instructions are supposedly broken concerns me: rust-lang/stdarch#1227

[hkratz]

Actually only the interleaved loads are affected (vld[234]*), using vld1* should be fine as those use ptr::read_unaligned under the hood.

[kaffeemonster]

Maybe you want to take a look at my adler32 SIMD implementation for NEON from way back then:
https://github.com/kaffeemonster/zlib/blob/adler32_vec/arm/adler32.c

It also contains some tricks which can be used in other SIMD (x86), mainly vector_chop, a modulo-approx. you can do on most SIMD instruction set. This way you can leave the data in the vector register and don't have to do a full reduction, copy to normal register set, do a modulo, and back to vector register. Also since you then have often 4 or more single sums, your inner loop count is not limited to NMAX (5552).

[peterdk]

Is there any progress or status on this? I need adler32 to be fast on arm64. Is this PR not deemed ok? Any plans to finish the support for NEON? I would love to contribute, but no SIMD experience.

[mcountryman]

Is there any progress or status on this? I need adler32 to be fast on arm64. Is this PR not deemed ok? Any plans to finish the support for NEON? I would love to contribute, but no SIMD experience.

I'll try running the build again when I have time. Last a checked rust support for some of the operations we need was lacking although this was a while ago

[mcountryman]

bump

[CryZe]

I'll quickly remove all the nightly conditions that shouldn't apply anymore to aarch64.

[CryZe]

Seems like vdotq_u32 is not stable yet somehow.

[CryZe]

Actually the error seems incorrect. The function is stable, but requires dotprod in addition to neon.

[mcountryman]

Actually the error seems incorrect. The function is stable, but requires dotprod in addition to neon.

It's been a bit since I looked at this code, but if I remember correct this would require adding an additional #cfg clause to allow compilation on non-dotprod,neon targets.

[CryZe]

Mmh, turns out that the dotprod feature is indeed still unstable:

https://github.com/rust-lang/stdarch/blob/master/crates/stdarch-gen/src/main.rs#L1590

[peterdk]

Would this dotprod be supported on Android ARM64 devices? Or is it a very obscure instruction? Would it impact performance much if a device doesn't have it? (cq can we still use NEON for the rest?)

[mcountryman]

Would this dotprod be supported on Android ARM64 devices? Or is it a very obscure instruction? Would it impact performance much if a device doesn't have it? (cq can we still use NEON for the rest?)

At a glance @CryZe comment indicates we could get it working when compiling with nightly. The alternative is to port the c implementations mentioned above and hope that the intrinsics are stable.

[peterdk]

I am willing to offer a bounty of 50 Euro to get this working on ARM64 Android. One for the implementer (@CryZe ?) and one for the owner of this crate.

[peterdk]

I now resorted to using libdeflater-sys that wraps the libdeflate C++ library, and provides SIMD accelerated adler32 besides deflate and crc32. It has a dotproduct check and implementations for both with and without dot prod, while using NEON intrinsics. It did speed up my app on Android with 33% compared to using nightly simd-adler32. For now I leave it at that, maybe I find time and motivation in the future to extract that adler32 part from the library into a rust one, to make it somewhat more safe. But it's SIMD, so I doubt much improvements can be done in that regard.