The crate versions are synchronized now. But they should be independent for further maintenance.

Related:

• #1
• #5

https://github.com/Nugine/simd-benches/actions/runs/8495251262

Related:

• rust-lang/stdarch#148
• rust-lang/stdarch#1210
• rust-lang/stdarch#1236
• rust-lang/stdarch#1235
• rust-lang/rust#102220

now vsimd support integer number only, but it in 3D math it represent floating number, please consider support it as possible ?

Project Status:

• Maintaining but not developing.
• I'm thinking a new SIMD design based on custom codegen, but have no idea for now.
• To avoid frequent version bumping, I plan to stay at v0.8 for one year.
• I'll come back!

Docs:

• More examples
• API docs translation
• Add some explanations and examples for encode and decode.
  #36

Tests:

• Fuzz tests
• #23
• Coverage report

Needs design:

• #5
• #11
• AVX512
• Refactor hex-simd APIs

Experimental

• #24

Issues

• #7
• #14
• #15
• #22

Improve throughput by using multi-threading (rayon).

In most cases, the simd functions are fast enough that we may not benefit from multi-threading. And some functions are not divisible.

Crates with parallel api:

• https://github.com/uhmarcel/rbase64

The arch::is_aarch64_feature_detected! feature has been stabilized as of rust 1.60.0: https://github.com/rust-lang/rust/blob/master/RELEASES.md#version-1600-2022-04-07

And armv8 neon intrinsics for aarch64 are stabilized as of rust 1.59.0: https://github.com/rust-lang/rust/blob/master/RELEASES.md#version-1590-2022-02-24

The hex simd features in particular should be possible to use without the unstable flag.

Awesome work on this library btw, very useful!

For all multiversion symbols, analyze code size

Related:

• #22

Upstream:

• #7
• rust-lang/rust#102220
• rust-lang/rust#102738

Hi Nugine,
I'm planning to add a new charset support to simd-base32, and is stuck with the alsw decoder implementation.
Given that the charset is:

const BASE32CROCKFORD_CHARSET: &[u8; 32] = b"0123456789ABCDEFGHJKMNPQRSTVWXYZ";

I have the decode function as:

    #[inline]
    const fn decode(c: u8) -> u8 {
        match c {
            b'0'..=b'9' => c - b'0',
            b'A'..=b'H' => c - b'A' + 10,
            b'J'..=b'K' => c - b'J' + 18,
            b'M'..=b'N' => c - b'M' + 20,
            b'P'..=b'T' => c - b'P' + 22,
            b'V'..=b'Z' => c - b'V' + 27,
            _ => 0xff,
        }
    }

But it is unclear to me how the check_hash function is implemented. Would you please add some documentation to the implementation and the verification of such functions?

Thanks!

base64-simd and other simd crates need more examples for common use cases.

Consider refactoring all the APIs for further extensions.

Related:

• marshallpierce/rust-base64#205

Updated by GitHub Action automatically

Summary 2023-01-13

• base64-simd is the fastest base64 crate.
• hex-simd is the fastest hex crate.

Current supported instruction sets (2022-11-18):

• AVX2
• SSE4.1
• SSSE3
• SSE2
• ARM NEON
• AArch64 NEON
• WASM SIMD128

Upstream blockers: (2022-11-18)

• rust-lang/portable-simd#303
• rust-lang/portable-simd#316

base64-simd

这么好的东西 github 默认搜索不到啊

Related:

• #36

https://llvm.org/docs/CommandGuide/FileCheck.html

Keep tracking performance

• x86_64 benchmark (GitHub Actions)
• aarch64 benchmark (CircleCI)
• A dashboard (GitHub Actions)

Currently forgiving mode handles no padding properly:

decoding SGVsbG8sIHdvcmxkIQ = Hello, world!

but decoding to little padding SGVsbG8sIHdvcmxkIQ= errors.

It would be great if forgiving mode appends the correct padding. Similarly it could also ignore extra padding:

aGVsbG8== = hello

Would it be possible to add a simple encode/decode method pair?

e.g:

encode_to_slice(source:&[u8],dest:&mut [u8])->Result(usize, Error)

decode_to_slice(source: &[u8], dest[&mut[u8])->Result(usize, Error)

so the caller does not need to call sizing additionally to find out how much was written. It just makes it a little more fluid to use.

This obviously assumes the dest buffers are large enough to hold the output (as with the other encode/decode methods)

Upstream:

• llvm/llvm-project#59081
• rust-lang/stdarch#1217

https://rust.godbolt.org/z/qjx6znnq4

rustc 1.67.0-nightly (83356b78c 2022-11-17)
binary: rustc
commit-hash: 83356b78c4ff3e7d84e977aa6143793545967301
commit-date: 2022-11-17
host: x86_64-unknown-linux-gnu
release: 1.67.0-nightly
LLVM version: 15.0.4

https://developer.arm.com/documentation/den0018/a/NEON-Instruction-Set-Architecture/Alignment

Related:

• #19

Upstream:

• #7
• rust-lang/rust#102220
• rust-lang/rust#102664
• rust-lang/rust#102709

Related crates:

• base64-simd
• hex-simd
• base32-simd

API docs in English and Chinese