From 1b6a04ca5504955c571d1c97504fb45ea0befee4 Mon Sep 17 00:00:00 2001 From: Valentin Popov Date: Mon, 8 Jan 2024 01:21:28 +0400 Subject: Initial vendor packages Signed-off-by: Valentin Popov --- vendor/simd-adler32/.cargo-checksum.json | 1 + vendor/simd-adler32/CHANGELOG.md | 12 ++ vendor/simd-adler32/Cargo.toml | 67 +++++++ vendor/simd-adler32/LICENSE.md | 21 +++ vendor/simd-adler32/README.md | 131 +++++++++++++ vendor/simd-adler32/src/hash.rs | 156 ++++++++++++++++ vendor/simd-adler32/src/imp/avx2.rs | 214 +++++++++++++++++++++ vendor/simd-adler32/src/imp/avx512.rs | 242 ++++++++++++++++++++++++ vendor/simd-adler32/src/imp/mod.rs | 23 +++ vendor/simd-adler32/src/imp/neon.rs | 241 ++++++++++++++++++++++++ vendor/simd-adler32/src/imp/scalar.rs | 69 +++++++ vendor/simd-adler32/src/imp/sse2.rs | 233 +++++++++++++++++++++++ vendor/simd-adler32/src/imp/ssse3.rs | 219 ++++++++++++++++++++++ vendor/simd-adler32/src/imp/wasm.rs | 217 ++++++++++++++++++++++ vendor/simd-adler32/src/lib.rs | 310 +++++++++++++++++++++++++++++++ 15 files changed, 2156 insertions(+) create mode 100644 vendor/simd-adler32/.cargo-checksum.json create mode 100644 vendor/simd-adler32/CHANGELOG.md create mode 100644 vendor/simd-adler32/Cargo.toml create mode 100644 vendor/simd-adler32/LICENSE.md create mode 100644 vendor/simd-adler32/README.md create mode 100644 vendor/simd-adler32/src/hash.rs create mode 100644 vendor/simd-adler32/src/imp/avx2.rs create mode 100644 vendor/simd-adler32/src/imp/avx512.rs create mode 100644 vendor/simd-adler32/src/imp/mod.rs create mode 100644 vendor/simd-adler32/src/imp/neon.rs create mode 100644 vendor/simd-adler32/src/imp/scalar.rs create mode 100644 vendor/simd-adler32/src/imp/sse2.rs create mode 100644 vendor/simd-adler32/src/imp/ssse3.rs create mode 100644 vendor/simd-adler32/src/imp/wasm.rs create mode 100644 vendor/simd-adler32/src/lib.rs (limited to 'vendor/simd-adler32') diff --git a/vendor/simd-adler32/.cargo-checksum.json b/vendor/simd-adler32/.cargo-checksum.json new file mode 100644 index 0000000..b25bff0 --- /dev/null +++ b/vendor/simd-adler32/.cargo-checksum.json @@ -0,0 +1 @@ +{"files":{"CHANGELOG.md":"4f762219d91ec3f7b6f4f25c6c4f4a83a5c3a0db1503827d9e1541c5fb8b62a1","Cargo.toml":"1df396cf8d69ee6f333bae303d4051c9f189167fa1ce07791becc5d3f2bd14fe","LICENSE.md":"42a35170233e83e18856792e748de4c1ce4a63b2afce9a370c89ef3fe23f9f2d","README.md":"a0568aab174ff15c2252898c7520ce0bd539cfb75d799238acd34dec44ee772c","src/hash.rs":"b5d0aea3ad486cccb67a156606e932b857cf0db5014393df50f290a495836d17","src/imp/avx2.rs":"46a83702bfe182278851d75f83b916a9fe0cc1ff5d1244d1912960ee1fcd82b2","src/imp/avx512.rs":"6e96adfe3099319529c486a186c71c85d92e93b29ea73558b84d65c618367858","src/imp/mod.rs":"9e6a1e69c40b03922b66a939b60e4217886a8aed3757af8d5a1e168c888121dd","src/imp/neon.rs":"2edad3b532b3384c83607a8adc6a1a46b4428708d02cb213f6afd333f60cfc99","src/imp/scalar.rs":"950c9f1202110da14d4ce725591fb2bdab36b73e500d2654e3a75c1a69a4ca5d","src/imp/sse2.rs":"fabefe43151e989a8a24910c682ae2661c30ee39363f0903311e6da60e192658","src/imp/ssse3.rs":"12f719d17bd9044c374f217312ea094625f22436e27bd12a28c60469f9448608","src/imp/wasm.rs":"3b49cde90bde87548e5169c061bf1abdde328710f75cde807d0921130aa232f1","src/lib.rs":"94525f007da93fbb09bbcdc164bc86a13a9526464203eb508ce65e5752c1a3f2"},"package":"d66dc143e6b11c1eddc06d5c423cfc97062865baf299914ab64caa38182078fe"} \ No newline at end of file diff --git a/vendor/simd-adler32/CHANGELOG.md b/vendor/simd-adler32/CHANGELOG.md new file mode 100644 index 0000000..95b141e --- /dev/null +++ b/vendor/simd-adler32/CHANGELOG.md @@ -0,0 +1,12 @@ +# Changelog + +## 0.3.3 - 2021-04-14 + +### Features + +- **from_checksum**: add `Adler32::from_checksum` + +### Performance Improvements + +- **scalar**: improve scalar performance by 90-600% + - Defer modulo until right before u16 overflow diff --git a/vendor/simd-adler32/Cargo.toml b/vendor/simd-adler32/Cargo.toml new file mode 100644 index 0000000..cdb8e0b --- /dev/null +++ b/vendor/simd-adler32/Cargo.toml @@ -0,0 +1,67 @@ +# THIS FILE IS AUTOMATICALLY GENERATED BY CARGO +# +# When uploading crates to the registry Cargo will automatically +# "normalize" Cargo.toml files for maximal compatibility +# with all versions of Cargo and also rewrite `path` dependencies +# to registry (e.g., crates.io) dependencies. +# +# If you are reading this file be aware that the original Cargo.toml +# will likely look very different (and much more reasonable). +# See Cargo.toml.orig for the original contents. + +[package] +edition = "2018" +name = "simd-adler32" +version = "0.3.7" +authors = ["Marvin Countryman "] +exclude = ["bench"] +description = "A SIMD-accelerated Adler-32 hash algorithm implementation." +readme = "README.md" +keywords = [ + "simd", + "avx2", + "ssse3", + "adler", + "adler32", +] +categories = [ + "algorithms", + "no-std", +] +license = "MIT" +repository = "https://github.com/mcountryman/simd-adler32" + +[profile.release] +opt-level = 2 +debug = 2 + +[[bench]] +name = "alts" +path = "bench/alts.rs" +harness = false + +[[bench]] +name = "variants" +path = "bench/variants.rs" +harness = false + +[dev-dependencies.adler] +version = "1.0.2" + +[dev-dependencies.adler32] +version = "1.2.0" + +[dev-dependencies.criterion] +version = "0.3" + +[dev-dependencies.rand] +version = "0.8" + +[features] +const-generics = [] +default = [ + "std", + "const-generics", +] +nightly = [] +std = [] diff --git a/vendor/simd-adler32/LICENSE.md b/vendor/simd-adler32/LICENSE.md new file mode 100644 index 0000000..9bd65ce --- /dev/null +++ b/vendor/simd-adler32/LICENSE.md @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) [2021] [Marvin Countryman] + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. diff --git a/vendor/simd-adler32/README.md b/vendor/simd-adler32/README.md new file mode 100644 index 0000000..4eeec46 --- /dev/null +++ b/vendor/simd-adler32/README.md @@ -0,0 +1,131 @@ +

simd-adler32

+

+ + docs.rs badge + + + crates.io badge + + + mit license badge + +

+ +A SIMD-accelerated Adler-32 hash algorithm implementation. + +## Features + +- No dependencies +- Support `no_std` (with `default-features = false`) +- Runtime CPU feature detection (when `std` enabled) +- Blazing fast performance on as many targets as possible (currently only x86 and x86_64) +- Default to scalar implementation when simd not available + +## Quick start + +> Cargo.toml + +```toml +[dependencies] +simd-adler32 = "*" +``` + +> example.rs + +```rust +use simd_adler32::Adler32; + +let mut adler = Adler32::new(); +adler.write(b"rust is pretty cool, man"); +let hash = adler.finish(); + +println!("{}", hash); +// 1921255656 +``` + +## Support + +**CPU Features** + +| impl | arch | feature | +| ---- | ---------------- | ------- | +| βœ… | `x86`, `x86_64` | avx512 | +| βœ… | `x86`, `x86_64` | avx2 | +| βœ… | `x86`, `x86_64` | ssse3 | +| βœ… | `x86`, `x86_64` | sse2 | +| 🚧 | `arm`, `aarch64` | neon | +| βœ… | `wasm32` | simd128 | + +**MSRV** `1.36.0`\*\* + +Minimum supported rust version is tested before a new version is published. [**] Feature +`const-generics` needs to disabled to build on rustc versions `<1.51` which can be done +by updating your dependency definition to the following. + +> Cargo.toml + +```toml +[dependencies] +simd-adler32 = { version "*", default-features = false, features = ["std"] } +``` + +## Performance + +Benchmarks listed display number of randomly generated bytes (10k / 100k) and library +name. Benchmarks sources can be found under the [bench](/bench) directory. Crates used for +comparison are [adler](https://crates.io/crates/adler) and +[adler32](https://crates.io/crates/adler32). + +> Windows 10 Pro - Intel i5-8300H @ 2.30GHz + +| name | avg. time | avg. thrpt | +| ----------------------- | --------------- | ------------------ | +| **10k/simd-adler32** | **212.61 ns** | **43.805 GiB/s** | +| 10k/wuffs | 3843 ns | 2.63 GiB/s\* | +| 10k/adler32 | 4.8084 us | 1.9369 GiB/s | +| 10k/adler | 17.979 us | 530.43 MiB/s | +| ----------------------- | --------------- | ------------------ | +| **100k/simd-adler32** | **2.7951 us** | **33.320 GiB/s** | +| 100k/wuffs | 34733 ns | 2.6814 GiB/s\* | +| 100k/adler32 | 48.488 us | 1.9207 GiB/s | +| 100k/adler | 178.36 us | 534.69 MiB/s | + +\* wuffs ran using mingw64/gcc, ran with `wuffs bench -ccompilers=gcc -reps=1 -iterscale=300 std/adler32`. + +> MacBookPro16,1 - Intel i9-9880H CPU @ 2.30GHz + +| name | avg. time | avg. thrpt | +| ----------------------- | --------------- | ------------------ | +| **10k/simd-adler32** | **200.37 ns** | **46.480 GiB/s** | +| 10k/adler32 | 4.1516 us | 2.2433 GiB/s | +| 10k/adler | 10.220 us | 933.15 MiB/s | +| ----------------------- | --------------- | ------------------ | +| **100k/simd-adler32** | **2.3282 us** | **40.003 GiB/s** | +| 100k/adler32 | 41.130 us | 2.2643 GiB/s | +| 100k/adler | 83.776 us | 534.69 MiB/s | + +## Safety + +This crate contains a significant amount of `unsafe` code due to the requirement of `unsafe` +for simd intrinsics. Fuzzing is done on release and debug builds prior to publishing via +`afl`. Fuzzy tests can be found under [fuzz](/fuzz) the directory. + +## Resources + +- [LICENSE](./LICENSE.md) - MIT +- [CHANGELOG](./CHANGELOG.md) + +## Credits + +Thank you to the contributors of the following projects. + +- [adler](https://github.com/jonas-schievink/adler) +- [adler32](https://github.com/remram44/adler32-rs) +- [crc32fast](https://github.com/srijs/rust-crc32fast) +- [wuffs](https://github.com/google/wuffs) +- [chromium](https://bugs.chromium.org/p/chromium/issues/detail?id=762564) +- [zlib](https://zlib.net/) + +## Contributing + +Feel free to submit a issue or pull request. :smile: diff --git a/vendor/simd-adler32/src/hash.rs b/vendor/simd-adler32/src/hash.rs new file mode 100644 index 0000000..558542b --- /dev/null +++ b/vendor/simd-adler32/src/hash.rs @@ -0,0 +1,156 @@ +use crate::{Adler32, Adler32Hash}; + +impl Adler32Hash for &[u8] { + fn hash(&self) -> u32 { + let mut hash = Adler32::new(); + + hash.write(self); + hash.finish() + } +} + +impl Adler32Hash for &str { + fn hash(&self) -> u32 { + let mut hash = Adler32::new(); + + hash.write(self.as_bytes()); + hash.finish() + } +} + +#[cfg(feature = "const-generics")] +impl Adler32Hash for [u8; SIZE] { + fn hash(&self) -> u32 { + let mut hash = Adler32::new(); + + hash.write(self); + hash.finish() + } +} + +macro_rules! array_impl { + ($s:expr, $($size:expr),+) => { + array_impl!($s); + $(array_impl!{$size})* + }; + ($size:expr) => { + #[cfg(not(feature = "const-generics"))] + impl Adler32Hash for [u8; $size] { + fn hash(&self) -> u32 { + let mut hash = Adler32::new(); + + hash.write(self); + hash.finish() + } + } + }; +} + +array_impl!( + 0, + 1, + 2, + 3, + 4, + 5, + 6, + 7, + 8, + 9, + 10, + 11, + 12, + 13, + 14, + 15, + 16, + 17, + 18, + 19, + 20, + 21, + 22, + 23, + 24, + 25, + 26, + 27, + 28, + 29, + 30, + 31, + 32, + 33, + 34, + 35, + 36, + 37, + 38, + 39, + 40, + 41, + 42, + 43, + 44, + 45, + 46, + 47, + 48, + 49, + 50, + 51, + 52, + 53, + 54, + 55, + 56, + 57, + 58, + 59, + 60, + 61, + 62, + 63, + 64, + 65, + 66, + 67, + 68, + 69, + 70, + 71, + 72, + 73, + 74, + 75, + 76, + 77, + 78, + 79, + 80, + 81, + 82, + 83, + 84, + 85, + 86, + 87, + 88, + 89, + 90, + 91, + 92, + 93, + 94, + 95, + 96, + 97, + 98, + 99, + 100, + 1024, + 1024 * 1024, + 1024 * 1024 * 1024, + 2048, + 4096 +); diff --git a/vendor/simd-adler32/src/imp/avx2.rs b/vendor/simd-adler32/src/imp/avx2.rs new file mode 100644 index 0000000..c16cc99 --- /dev/null +++ b/vendor/simd-adler32/src/imp/avx2.rs @@ -0,0 +1,214 @@ +use super::Adler32Imp; + +/// Resolves update implementation if CPU supports avx2 instructions. +pub fn get_imp() -> Option { + get_imp_inner() +} + +#[inline] +#[cfg(all(feature = "std", any(target_arch = "x86", target_arch = "x86_64")))] +fn get_imp_inner() -> Option { + if std::is_x86_feature_detected!("avx2") { + Some(imp::update) + } else { + None + } +} + +#[inline] +#[cfg(all( + target_feature = "avx2", + not(all(feature = "std", any(target_arch = "x86", target_arch = "x86_64"))) +))] +fn get_imp_inner() -> Option { + Some(imp::update) +} + +#[inline] +#[cfg(all( + not(target_feature = "avx2"), + not(all(feature = "std", any(target_arch = "x86", target_arch = "x86_64"))) +))] +fn get_imp_inner() -> Option { + None +} + +#[cfg(all( + any(target_arch = "x86", target_arch = "x86_64"), + any(feature = "std", target_feature = "avx2") +))] +mod imp { + const MOD: u32 = 65521; + const NMAX: usize = 5552; + const BLOCK_SIZE: usize = 32; + const CHUNK_SIZE: usize = NMAX / BLOCK_SIZE * BLOCK_SIZE; + + #[cfg(target_arch = "x86")] + use core::arch::x86::*; + #[cfg(target_arch = "x86_64")] + use core::arch::x86_64::*; + + pub fn update(a: u16, b: u16, data: &[u8]) -> (u16, u16) { + unsafe { update_imp(a, b, data) } + } + + #[inline] + #[target_feature(enable = "avx2")] + unsafe fn update_imp(a: u16, b: u16, data: &[u8]) -> (u16, u16) { + let mut a = a as u32; + let mut b = b as u32; + + let chunks = data.chunks_exact(CHUNK_SIZE); + let remainder = chunks.remainder(); + for chunk in chunks { + update_chunk_block(&mut a, &mut b, chunk); + } + + update_block(&mut a, &mut b, remainder); + + (a as u16, b as u16) + } + + #[inline] + unsafe fn update_chunk_block(a: &mut u32, b: &mut u32, chunk: &[u8]) { + debug_assert_eq!( + chunk.len(), + CHUNK_SIZE, + "Unexpected chunk size (expected {}, got {})", + CHUNK_SIZE, + chunk.len() + ); + + reduce_add_blocks(a, b, chunk); + + *a %= MOD; + *b %= MOD; + } + + #[inline] + unsafe fn update_block(a: &mut u32, b: &mut u32, chunk: &[u8]) { + debug_assert!( + chunk.len() <= CHUNK_SIZE, + "Unexpected chunk size (expected <= {}, got {})", + CHUNK_SIZE, + chunk.len() + ); + + for byte in reduce_add_blocks(a, b, chunk) { + *a += *byte as u32; + *b += *a; + } + + *a %= MOD; + *b %= MOD; + } + + #[inline(always)] + unsafe fn reduce_add_blocks<'a>(a: &mut u32, b: &mut u32, chunk: &'a [u8]) -> &'a [u8] { + if chunk.len() < BLOCK_SIZE { + return chunk; + } + + let blocks = chunk.chunks_exact(BLOCK_SIZE); + let blocks_remainder = blocks.remainder(); + + let one_v = _mm256_set1_epi16(1); + let zero_v = _mm256_setzero_si256(); + let weights = get_weights(); + + let mut p_v = _mm256_set_epi32(0, 0, 0, 0, 0, 0, 0, (*a * blocks.len() as u32) as _); + let mut a_v = _mm256_setzero_si256(); + let mut b_v = _mm256_set_epi32(0, 0, 0, 0, 0, 0, 0, *b as _); + + for block in blocks { + let block_ptr = block.as_ptr() as *const _; + let block = _mm256_loadu_si256(block_ptr); + + p_v = _mm256_add_epi32(p_v, a_v); + + a_v = _mm256_add_epi32(a_v, _mm256_sad_epu8(block, zero_v)); + let mad = _mm256_maddubs_epi16(block, weights); + b_v = _mm256_add_epi32(b_v, _mm256_madd_epi16(mad, one_v)); + } + + b_v = _mm256_add_epi32(b_v, _mm256_slli_epi32(p_v, 5)); + + *a += reduce_add(a_v); + *b = reduce_add(b_v); + + blocks_remainder + } + + #[inline(always)] + unsafe fn reduce_add(v: __m256i) -> u32 { + let sum = _mm_add_epi32(_mm256_castsi256_si128(v), _mm256_extracti128_si256(v, 1)); + let hi = _mm_unpackhi_epi64(sum, sum); + + let sum = _mm_add_epi32(hi, sum); + let hi = _mm_shuffle_epi32(sum, crate::imp::_MM_SHUFFLE(2, 3, 0, 1)); + + let sum = _mm_add_epi32(sum, hi); + + _mm_cvtsi128_si32(sum) as _ + } + + #[inline(always)] + unsafe fn get_weights() -> __m256i { + _mm256_set_epi8( + 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, + 24, 25, 26, 27, 28, 29, 30, 31, 32, + ) + } +} + +#[cfg(test)] +mod tests { + use rand::Rng; + + #[test] + fn zeroes() { + assert_sum_eq(&[]); + assert_sum_eq(&[0]); + assert_sum_eq(&[0, 0]); + assert_sum_eq(&[0; 100]); + assert_sum_eq(&[0; 1024]); + assert_sum_eq(&[0; 1024 * 1024]); + } + + #[test] + fn ones() { + assert_sum_eq(&[]); + assert_sum_eq(&[1]); + assert_sum_eq(&[1, 1]); + assert_sum_eq(&[1; 100]); + assert_sum_eq(&[1; 1024]); + assert_sum_eq(&[1; 1024 * 1024]); + } + + #[test] + fn random() { + let mut random = [0; 1024 * 1024]; + rand::thread_rng().fill(&mut random[..]); + + assert_sum_eq(&random[..1]); + assert_sum_eq(&random[..100]); + assert_sum_eq(&random[..1024]); + assert_sum_eq(&random[..1024 * 1024]); + } + + /// Example calculation from https://en.wikipedia.org/wiki/Adler-32. + #[test] + fn wiki() { + assert_sum_eq(b"Wikipedia"); + } + + fn assert_sum_eq(data: &[u8]) { + if let Some(update) = super::get_imp() { + let (a, b) = update(1, 0, data); + let left = u32::from(b) << 16 | u32::from(a); + let right = adler::adler32_slice(data); + + assert_eq!(left, right, "len({})", data.len()); + } + } +} diff --git a/vendor/simd-adler32/src/imp/avx512.rs b/vendor/simd-adler32/src/imp/avx512.rs new file mode 100644 index 0000000..ebb32fa --- /dev/null +++ b/vendor/simd-adler32/src/imp/avx512.rs @@ -0,0 +1,242 @@ +use super::Adler32Imp; + +/// Resolves update implementation if CPU supports avx512f and avx512bw instructions. +pub fn get_imp() -> Option { + get_imp_inner() +} + +#[inline] +#[cfg(all( + feature = "std", + feature = "nightly", + any(target_arch = "x86", target_arch = "x86_64") +))] +fn get_imp_inner() -> Option { + let has_avx512f = std::is_x86_feature_detected!("avx512f"); + let has_avx512bw = std::is_x86_feature_detected!("avx512bw"); + + if has_avx512f && has_avx512bw { + Some(imp::update) + } else { + None + } +} + +#[inline] +#[cfg(all( + feature = "nightly", + all(target_feature = "avx512f", target_feature = "avx512bw"), + not(all(feature = "std", any(target_arch = "x86", target_arch = "x86_64"))) +))] +fn get_imp_inner() -> Option { + Some(imp::update) +} + +#[inline] +#[cfg(all( + not(all(feature = "nightly", target_feature = "avx512f", target_feature = "avx512bw")), + not(all( + feature = "std", + feature = "nightly", + any(target_arch = "x86", target_arch = "x86_64") + )) +))] +fn get_imp_inner() -> Option { + None +} + +#[cfg(all( + feature = "nightly", + any(target_arch = "x86", target_arch = "x86_64"), + any( + feature = "std", + all(target_feature = "avx512f", target_feature = "avx512bw") + ) +))] +mod imp { + const MOD: u32 = 65521; + const NMAX: usize = 5552; + const BLOCK_SIZE: usize = 64; + const CHUNK_SIZE: usize = NMAX / BLOCK_SIZE * BLOCK_SIZE; + + #[cfg(target_arch = "x86")] + use core::arch::x86::*; + #[cfg(target_arch = "x86_64")] + use core::arch::x86_64::*; + + pub fn update(a: u16, b: u16, data: &[u8]) -> (u16, u16) { + unsafe { update_imp(a, b, data) } + } + + #[inline] + #[target_feature(enable = "avx512f")] + #[target_feature(enable = "avx512bw")] + unsafe fn update_imp(a: u16, b: u16, data: &[u8]) -> (u16, u16) { + let mut a = a as u32; + let mut b = b as u32; + + let chunks = data.chunks_exact(CHUNK_SIZE); + let remainder = chunks.remainder(); + for chunk in chunks { + update_chunk_block(&mut a, &mut b, chunk); + } + + update_block(&mut a, &mut b, remainder); + + (a as u16, b as u16) + } + + #[inline] + unsafe fn update_chunk_block(a: &mut u32, b: &mut u32, chunk: &[u8]) { + debug_assert_eq!( + chunk.len(), + CHUNK_SIZE, + "Unexpected chunk size (expected {}, got {})", + CHUNK_SIZE, + chunk.len() + ); + + reduce_add_blocks(a, b, chunk); + + *a %= MOD; + *b %= MOD; + } + + #[inline] + unsafe fn update_block(a: &mut u32, b: &mut u32, chunk: &[u8]) { + debug_assert!( + chunk.len() <= CHUNK_SIZE, + "Unexpected chunk size (expected <= {}, got {})", + CHUNK_SIZE, + chunk.len() + ); + + for byte in reduce_add_blocks(a, b, chunk) { + *a += *byte as u32; + *b += *a; + } + + *a %= MOD; + *b %= MOD; + } + + #[inline(always)] + unsafe fn reduce_add_blocks<'a>(a: &mut u32, b: &mut u32, chunk: &'a [u8]) -> &'a [u8] { + if chunk.len() < BLOCK_SIZE { + return chunk; + } + + let blocks = chunk.chunks_exact(BLOCK_SIZE); + let blocks_remainder = blocks.remainder(); + + let one_v = _mm512_set1_epi16(1); + let zero_v = _mm512_setzero_si512(); + let weights = get_weights(); + + let p_v = (*a * blocks.len() as u32) as _; + let mut p_v = _mm512_set_epi32(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, p_v); + let mut a_v = _mm512_setzero_si512(); + let mut b_v = _mm512_set_epi32(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, *b as _); + + for block in blocks { + let block_ptr = block.as_ptr() as *const _; + let block = _mm512_loadu_si512(block_ptr); + + p_v = _mm512_add_epi32(p_v, a_v); + + a_v = _mm512_add_epi32(a_v, _mm512_sad_epu8(block, zero_v)); + let mad = _mm512_maddubs_epi16(block, weights); + b_v = _mm512_add_epi32(b_v, _mm512_madd_epi16(mad, one_v)); + } + + b_v = _mm512_add_epi32(b_v, _mm512_slli_epi32(p_v, 6)); + + *a += reduce_add(a_v); + *b = reduce_add(b_v); + + blocks_remainder + } + + #[inline(always)] + unsafe fn reduce_add(v: __m512i) -> u32 { + let v: [__m256i; 2] = core::mem::transmute(v); + + reduce_add_256(v[0]) + reduce_add_256(v[1]) + } + + #[inline(always)] + unsafe fn reduce_add_256(v: __m256i) -> u32 { + let v: [__m128i; 2] = core::mem::transmute(v); + let sum = _mm_add_epi32(v[0], v[1]); + let hi = _mm_unpackhi_epi64(sum, sum); + + let sum = _mm_add_epi32(hi, sum); + let hi = _mm_shuffle_epi32(sum, crate::imp::_MM_SHUFFLE(2, 3, 0, 1)); + + let sum = _mm_add_epi32(sum, hi); + let sum = _mm_cvtsi128_si32(sum) as _; + + sum + } + + #[inline(always)] + unsafe fn get_weights() -> __m512i { + _mm512_set_epi8( + 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, + 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, + 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, + ) + } +} + +#[cfg(test)] +mod tests { + use rand::Rng; + + #[test] + fn zeroes() { + assert_sum_eq(&[]); + assert_sum_eq(&[0]); + assert_sum_eq(&[0, 0]); + assert_sum_eq(&[0; 100]); + assert_sum_eq(&[0; 1024]); + assert_sum_eq(&[0; 1024 * 1024]); + } + + #[test] + fn ones() { + assert_sum_eq(&[]); + assert_sum_eq(&[1]); + assert_sum_eq(&[1, 1]); + assert_sum_eq(&[1; 100]); + assert_sum_eq(&[1; 1024]); + assert_sum_eq(&[1; 1024 * 1024]); + } + + #[test] + fn random() { + let mut random = [0; 1024 * 1024]; + rand::thread_rng().fill(&mut random[..]); + + assert_sum_eq(&random[..1]); + assert_sum_eq(&random[..100]); + assert_sum_eq(&random[..1024]); + assert_sum_eq(&random[..1024 * 1024]); + } + + /// Example calculation from https://en.wikipedia.org/wiki/Adler-32. + #[test] + fn wiki() { + assert_sum_eq(b"Wikipedia"); + } + + fn assert_sum_eq(data: &[u8]) { + if let Some(update) = super::get_imp() { + let (a, b) = update(1, 0, data); + let left = u32::from(b) << 16 | u32::from(a); + let right = adler::adler32_slice(data); + + assert_eq!(left, right, "len({})", data.len()); + } + } +} diff --git a/vendor/simd-adler32/src/imp/mod.rs b/vendor/simd-adler32/src/imp/mod.rs new file mode 100644 index 0000000..957b50a --- /dev/null +++ b/vendor/simd-adler32/src/imp/mod.rs @@ -0,0 +1,23 @@ +pub mod avx2; +pub mod avx512; +pub mod scalar; +pub mod sse2; +pub mod ssse3; +pub mod wasm; + +pub type Adler32Imp = fn(u16, u16, &[u8]) -> (u16, u16); + +#[inline] +#[allow(non_snake_case)] +pub const fn _MM_SHUFFLE(z: u32, y: u32, x: u32, w: u32) -> i32 { + ((z << 6) | (y << 4) | (x << 2) | w) as i32 +} + +pub fn get_imp() -> Adler32Imp { + avx512::get_imp() + .or_else(avx2::get_imp) + .or_else(ssse3::get_imp) + .or_else(sse2::get_imp) + .or_else(wasm::get_imp) + .unwrap_or(scalar::update) +} diff --git a/vendor/simd-adler32/src/imp/neon.rs b/vendor/simd-adler32/src/imp/neon.rs new file mode 100644 index 0000000..8398b6d --- /dev/null +++ b/vendor/simd-adler32/src/imp/neon.rs @@ -0,0 +1,241 @@ +use super::Adler32Imp; + +/// Resolves update implementation if CPU supports avx512f and avx512bw instructions. +pub fn get_imp() -> Option { + get_imp_inner() +} + +#[inline] +#[cfg(all(feature = "std", feature = "nightly", target_arch = "arm"))] +fn get_imp_inner() -> Option { + if std::is_arm_feature_detected("neon") { + Some(imp::update) + } else { + None + } +} + +#[inline] +#[cfg(all(feature = "std", feature = "nightly", target_arch = "aarch64"))] +fn get_imp_inner() -> Option { + if std::is_aarch64_feature_detected("neon") { + Some(imp::update) + } else { + None + } +} + +#[inline] +#[cfg(all( + feature = "nightly", + target_feature = "neon", + not(all(feature = "std", any(target_arch = "arm", target_arch = "aarch64"))) +))] +fn get_imp_inner() -> Option { + Some(imp::update) +} + +#[inline] +#[cfg(all( + not(target_feature = "neon"), + not(all( + feature = "std", + feature = "nightly", + any(target_arch = "arm", target_arch = "aarch64") + )) +))] +fn get_imp_inner() -> Option { + None +} + +#[cfg(all( + feature = "nightly", + any(target_arch = "arm", target_arch = "aarch64"), + any(feature = "std", target_feature = "neon") +))] +mod imp { + const MOD: u32 = 65521; + const NMAX: usize = 5552; + const BLOCK_SIZE: usize = 64; + const CHUNK_SIZE: usize = NMAX / BLOCK_SIZE * BLOCK_SIZE; + + #[cfg(target_arch = "aarch64")] + use core::arch::aarch64::*; + #[cfg(target_arch = "arm")] + use core::arch::arm::*; + + pub fn update(a: u16, b: u16, data: &[u8]) -> (u16, u16) { + unsafe { update_imp(a, b, data) } + } + + #[inline] + #[target_feature(enable = "neon")] + unsafe fn update_imp(a: u16, b: u16, data: &[u8]) -> (u16, u16) { + let mut a = a as u32; + let mut b = b as u32; + + let chunks = data.chunks_exact(CHUNK_SIZE); + let remainder = chunks.remainder(); + for chunk in chunks { + update_chunk_block(&mut a, &mut b, chunk); + } + + update_block(&mut a, &mut b, remainder); + + (a as u16, b as u16) + } + + #[inline] + unsafe fn update_chunk_block(a: &mut u32, b: &mut u32, chunk: &[u8]) { + debug_assert_eq!( + chunk.len(), + CHUNK_SIZE, + "Unexpected chunk size (expected {}, got {})", + CHUNK_SIZE, + chunk.len() + ); + + reduce_add_blocks(a, b, chunk); + + *a %= MOD; + *b %= MOD; + } + + #[inline] + unsafe fn update_block(a: &mut u32, b: &mut u32, chunk: &[u8]) { + debug_assert!( + chunk.len() <= CHUNK_SIZE, + "Unexpected chunk size (expected <= {}, got {})", + CHUNK_SIZE, + chunk.len() + ); + + for byte in reduce_add_blocks(a, b, chunk) { + *a += *byte as u32; + *b += *a; + } + + *a %= MOD; + *b %= MOD; + } + + #[inline(always)] + unsafe fn reduce_add_blocks<'a>(a: &mut u32, b: &mut u32, chunk: &'a [u8]) -> &'a [u8] { + if chunk.len() < BLOCK_SIZE { + return chunk; + } + + let blocks = chunk.chunks_exact(BLOCK_SIZE); + let blocks_remainder = blocks.remainder(); + + let one_v = _mm512_set1_epi16(1); + let zero_v = _mm512_setzero_si512(); + let weights = get_weights(); + + let p_v = (*a * blocks.len() as u32) as _; + let mut p_v = _mm512_set_epi32(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, p_v); + let mut a_v = _mm512_setzero_si512(); + let mut b_v = _mm512_set_epi32(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, *b as _); + + for block in blocks { + let block_ptr = block.as_ptr() as *const _; + let block = _mm512_loadu_si512(block_ptr); + + p_v = _mm512_add_epi32(p_v, a_v); + + a_v = _mm512_add_epi32(a_v, _mm512_sad_epu8(block, zero_v)); + let mad = _mm512_maddubs_epi16(block, weights); + b_v = _mm512_add_epi32(b_v, _mm512_madd_epi16(mad, one_v)); + } + + b_v = _mm512_add_epi32(b_v, _mm512_slli_epi32(p_v, 6)); + + *a += reduce_add(a_v); + *b = reduce_add(b_v); + + blocks_remainder + } + + #[inline(always)] + unsafe fn reduce_add(v: __m512i) -> u32 { + let v: [__m256i; 2] = core::mem::transmute(v); + + reduce_add_256(v[0]) + reduce_add_256(v[1]) + } + + #[inline(always)] + unsafe fn reduce_add_256(v: __m256i) -> u32 { + let v: [__m128i; 2] = core::mem::transmute(v); + let sum = _mm_add_epi32(v[0], v[1]); + let hi = _mm_unpackhi_epi64(sum, sum); + + let sum = _mm_add_epi32(hi, sum); + let hi = _mm_shuffle_epi32(sum, crate::imp::_MM_SHUFFLE(2, 3, 0, 1)); + + let sum = _mm_add_epi32(sum, hi); + let sum = _mm_cvtsi128_si32(sum) as _; + + sum + } + + #[inline(always)] + unsafe fn get_weights() -> __m512i { + _mm512_set_epi8( + 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, + 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, + 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, + ) + } +} + +#[cfg(test)] +mod tests { + use rand::Rng; + + #[test] + fn zeroes() { + assert_sum_eq(&[]); + assert_sum_eq(&[0]); + assert_sum_eq(&[0, 0]); + assert_sum_eq(&[0; 100]); + assert_sum_eq(&[0; 1024]); + assert_sum_eq(&[0; 1024 * 1024]); + } + + #[test] + fn ones() { + assert_sum_eq(&[]); + assert_sum_eq(&[1]); + assert_sum_eq(&[1, 1]); + assert_sum_eq(&[1; 100]); + assert_sum_eq(&[1; 1024]); + assert_sum_eq(&[1; 1024 * 1024]); + } + + #[test] + fn random() { + let mut random = [0; 1024 * 1024]; + rand::thread_rng().fill(&mut random[..]); + + assert_sum_eq(&random[..1]); + assert_sum_eq(&random[..100]); + assert_sum_eq(&random[..1024]); + assert_sum_eq(&random[..1024 * 1024]); + } + + /// Example calculation from https://en.wikipedia.org/wiki/Adler-32. + #[test] + fn wiki() { + assert_sum_eq(b"Wikipedia"); + } + + fn assert_sum_eq(data: &[u8]) { + if let Some(update) = super::get_imp() { + let (a, b) = update(1, 0, data); + let left = u32::from(b) << 16 | u32::from(a); + let right = adler::adler32_slice(data); + + assert_eq!(left, right, "len({})", data.len()); + } + } +} diff --git a/vendor/simd-adler32/src/imp/scalar.rs b/vendor/simd-adler32/src/imp/scalar.rs new file mode 100644 index 0000000..558813e --- /dev/null +++ b/vendor/simd-adler32/src/imp/scalar.rs @@ -0,0 +1,69 @@ +const MOD: u32 = 65521; +const NMAX: usize = 5552; + +pub fn update(a: u16, b: u16, data: &[u8]) -> (u16, u16) { + let mut a = a as u32; + let mut b = b as u32; + + let chunks = data.chunks_exact(NMAX); + let remainder = chunks.remainder(); + + for chunk in chunks { + for byte in chunk { + a = a.wrapping_add(*byte as _); + b = b.wrapping_add(a); + } + + a %= MOD; + b %= MOD; + } + + for byte in remainder { + a = a.wrapping_add(*byte as _); + b = b.wrapping_add(a); + } + + a %= MOD; + b %= MOD; + + (a as u16, b as u16) +} + +#[cfg(test)] +mod tests { + #[test] + fn zeroes() { + assert_eq!(adler32(&[]), 1); + assert_eq!(adler32(&[0]), 1 | 1 << 16); + assert_eq!(adler32(&[0, 0]), 1 | 2 << 16); + assert_eq!(adler32(&[0; 100]), 0x00640001); + assert_eq!(adler32(&[0; 1024]), 0x04000001); + assert_eq!(adler32(&[0; 1024 * 1024]), 0x00f00001); + } + + #[test] + fn ones() { + assert_eq!(adler32(&[0xff; 1024]), 0x79a6fc2e); + assert_eq!(adler32(&[0xff; 1024 * 1024]), 0x8e88ef11); + } + + #[test] + fn mixed() { + assert_eq!(adler32(&[1]), 2 | 2 << 16); + assert_eq!(adler32(&[40]), 41 | 41 << 16); + + assert_eq!(adler32(&[0xA5; 1024 * 1024]), 0xd5009ab1); + } + + /// Example calculation from https://en.wikipedia.org/wiki/Adler-32. + #[test] + fn wiki() { + assert_eq!(adler32(b"Wikipedia"), 0x11E60398); + } + + fn adler32(data: &[u8]) -> u32 { + let (a, b) = super::update(1, 0, data); + + u32::from(b) << 16 | u32::from(a) + } +} diff --git a/vendor/simd-adler32/src/imp/sse2.rs b/vendor/simd-adler32/src/imp/sse2.rs new file mode 100644 index 0000000..b76df52 --- /dev/null +++ b/vendor/simd-adler32/src/imp/sse2.rs @@ -0,0 +1,233 @@ +use super::Adler32Imp; + +/// Resolves update implementation if CPU supports sse2 instructions. +pub fn get_imp() -> Option { + get_imp_inner() +} + +#[inline] +#[cfg(all(feature = "std", any(target_arch = "x86", target_arch = "x86_64")))] +fn get_imp_inner() -> Option { + if std::is_x86_feature_detected!("sse2") { + Some(imp::update) + } else { + None + } +} + +#[inline] +#[cfg(all( + target_feature = "sse2", + not(all(feature = "std", any(target_arch = "x86", target_arch = "x86_64"))) +))] +fn get_imp_inner() -> Option { + Some(imp::update) +} + +#[inline] +#[cfg(all( + not(target_feature = "sse2"), + not(all(feature = "std", any(target_arch = "x86", target_arch = "x86_64"))) +))] +fn get_imp_inner() -> Option { + None +} + +#[cfg(all( + any(target_arch = "x86", target_arch = "x86_64"), + any(feature = "std", target_feature = "sse2") +))] +mod imp { + const MOD: u32 = 65521; + const NMAX: usize = 5552; + const BLOCK_SIZE: usize = 32; + const CHUNK_SIZE: usize = NMAX / BLOCK_SIZE * BLOCK_SIZE; + + #[cfg(target_arch = "x86")] + use core::arch::x86::*; + #[cfg(target_arch = "x86_64")] + use core::arch::x86_64::*; + + pub fn update(a: u16, b: u16, data: &[u8]) -> (u16, u16) { + unsafe { update_imp(a, b, data) } + } + + #[inline] + #[target_feature(enable = "sse2")] + unsafe fn update_imp(a: u16, b: u16, data: &[u8]) -> (u16, u16) { + let mut a = a as u32; + let mut b = b as u32; + + let chunks = data.chunks_exact(CHUNK_SIZE); + let remainder = chunks.remainder(); + for chunk in chunks { + update_chunk_block(&mut a, &mut b, chunk); + } + + update_block(&mut a, &mut b, remainder); + + (a as u16, b as u16) + } + + unsafe fn update_chunk_block(a: &mut u32, b: &mut u32, chunk: &[u8]) { + debug_assert_eq!( + chunk.len(), + CHUNK_SIZE, + "Unexpected chunk size (expected {}, got {})", + CHUNK_SIZE, + chunk.len() + ); + + reduce_add_blocks(a, b, chunk); + + *a %= MOD; + *b %= MOD; + } + + unsafe fn update_block(a: &mut u32, b: &mut u32, chunk: &[u8]) { + debug_assert!( + chunk.len() <= CHUNK_SIZE, + "Unexpected chunk size (expected <= {}, got {})", + CHUNK_SIZE, + chunk.len() + ); + + for byte in reduce_add_blocks(a, b, chunk) { + *a += *byte as u32; + *b += *a; + } + + *a %= MOD; + *b %= MOD; + } + + #[inline(always)] + unsafe fn reduce_add_blocks<'a>(a: &mut u32, b: &mut u32, chunk: &'a [u8]) -> &'a [u8] { + if chunk.len() < BLOCK_SIZE { + return chunk; + } + + let blocks = chunk.chunks_exact(BLOCK_SIZE); + let blocks_remainder = blocks.remainder(); + + let zero_v = _mm_setzero_si128(); + let weight_hi_v = get_weight_hi(); + let weight_lo_v = get_weight_lo(); + + let mut p_v = _mm_set_epi32(0, 0, 0, (*a * blocks.len() as u32) as _); + let mut a_v = _mm_setzero_si128(); + let mut b_v = _mm_set_epi32(0, 0, 0, *b as _); + + for block in blocks { + let block_ptr = block.as_ptr() as *const _; + let left_v = _mm_loadu_si128(block_ptr); + let right_v = _mm_loadu_si128(block_ptr.add(1)); + + p_v = _mm_add_epi32(p_v, a_v); + + a_v = _mm_add_epi32(a_v, _mm_sad_epu8(left_v, zero_v)); + let mad = maddubs(left_v, weight_hi_v); + b_v = _mm_add_epi32(b_v, mad); + + a_v = _mm_add_epi32(a_v, _mm_sad_epu8(right_v, zero_v)); + let mad = maddubs(right_v, weight_lo_v); + b_v = _mm_add_epi32(b_v, mad); + } + + b_v = _mm_add_epi32(b_v, _mm_slli_epi32(p_v, 5)); + + *a += reduce_add(a_v); + *b = reduce_add(b_v); + + blocks_remainder + } + + #[inline(always)] + unsafe fn maddubs(a: __m128i, b: __m128i) -> __m128i { + let a_lo = _mm_unpacklo_epi8(a, _mm_setzero_si128()); + let a_hi = _mm_unpackhi_epi8(a, _mm_setzero_si128()); + + let b_lo = _mm_unpacklo_epi8(b, _mm_setzero_si128()); + let b_hi = _mm_unpackhi_epi8(b, _mm_setzero_si128()); + + let lo = _mm_madd_epi16(a_lo, b_lo); + let hi = _mm_madd_epi16(a_hi, b_hi); + + _mm_add_epi32(lo, hi) + } + + #[inline(always)] + unsafe fn reduce_add(v: __m128i) -> u32 { + let hi = _mm_unpackhi_epi64(v, v); + let sum = _mm_add_epi32(hi, v); + let hi = _mm_shuffle_epi32(sum, crate::imp::_MM_SHUFFLE(2, 3, 0, 1)); + + let sum = _mm_add_epi32(sum, hi); + + _mm_cvtsi128_si32(sum) as _ + } + + #[inline(always)] + unsafe fn get_weight_lo() -> __m128i { + _mm_set_epi8(1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16) + } + + #[inline(always)] + unsafe fn get_weight_hi() -> __m128i { + _mm_set_epi8( + 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, + ) + } +} + +#[cfg(test)] +mod tests { + use rand::Rng; + + #[test] + fn zeroes() { + assert_sum_eq(&[]); + assert_sum_eq(&[0]); + assert_sum_eq(&[0, 0]); + assert_sum_eq(&[0; 100]); + assert_sum_eq(&[0; 1024]); + assert_sum_eq(&[0; 1024 * 1024]); + } + + #[test] + fn ones() { + assert_sum_eq(&[]); + assert_sum_eq(&[1]); + assert_sum_eq(&[1, 1]); + assert_sum_eq(&[1; 100]); + assert_sum_eq(&[1; 1024]); + assert_sum_eq(&[1; 1024 * 1024]); + } + + #[test] + fn random() { + let mut random = [0; 1024 * 1024]; + rand::thread_rng().fill(&mut random[..]); + + assert_sum_eq(&random[..1]); + assert_sum_eq(&random[..100]); + assert_sum_eq(&random[..1024]); + assert_sum_eq(&random[..1024 * 1024]); + } + + /// Example calculation from https://en.wikipedia.org/wiki/Adler-32. + #[test] + fn wiki() { + assert_sum_eq(b"Wikipedia"); + } + + fn assert_sum_eq(data: &[u8]) { + if let Some(update) = super::get_imp() { + let (a, b) = update(1, 0, data); + let left = u32::from(b) << 16 | u32::from(a); + let right = adler::adler32_slice(data); + + assert_eq!(left, right, "len({})", data.len()); + } + } +} diff --git a/vendor/simd-adler32/src/imp/ssse3.rs b/vendor/simd-adler32/src/imp/ssse3.rs new file mode 100644 index 0000000..2602d47 --- /dev/null +++ b/vendor/simd-adler32/src/imp/ssse3.rs @@ -0,0 +1,219 @@ +use super::Adler32Imp; + +/// Resolves update implementation if CPU supports ssse3 instructions. +pub fn get_imp() -> Option { + get_imp_inner() +} + +#[inline] +#[cfg(all(feature = "std", any(target_arch = "x86", target_arch = "x86_64")))] +fn get_imp_inner() -> Option { + if std::is_x86_feature_detected!("ssse3") { + Some(imp::update) + } else { + None + } +} + +#[inline] +#[cfg(all( + target_feature = "ssse3", + not(all(feature = "std", any(target_arch = "x86", target_arch = "x86_64"))) +))] +fn get_imp_inner() -> Option { + Some(imp::update) +} + +#[inline] +#[cfg(all( + not(target_feature = "ssse3"), + not(all(feature = "std", any(target_arch = "x86", target_arch = "x86_64"))) +))] +fn get_imp_inner() -> Option { + None +} + +#[cfg(all( + any(target_arch = "x86", target_arch = "x86_64"), + any(feature = "std", target_feature = "ssse3") +))] +mod imp { + const MOD: u32 = 65521; + const NMAX: usize = 5552; + const BLOCK_SIZE: usize = 32; + const CHUNK_SIZE: usize = NMAX / BLOCK_SIZE * BLOCK_SIZE; + + #[cfg(target_arch = "x86")] + use core::arch::x86::*; + #[cfg(target_arch = "x86_64")] + use core::arch::x86_64::*; + + pub fn update(a: u16, b: u16, data: &[u8]) -> (u16, u16) { + unsafe { update_imp(a, b, data) } + } + + #[inline] + #[target_feature(enable = "ssse3")] + unsafe fn update_imp(a: u16, b: u16, data: &[u8]) -> (u16, u16) { + let mut a = a as u32; + let mut b = b as u32; + + let chunks = data.chunks_exact(CHUNK_SIZE); + let remainder = chunks.remainder(); + for chunk in chunks { + update_chunk_block(&mut a, &mut b, chunk); + } + + update_block(&mut a, &mut b, remainder); + + (a as u16, b as u16) + } + + unsafe fn update_chunk_block(a: &mut u32, b: &mut u32, chunk: &[u8]) { + debug_assert_eq!( + chunk.len(), + CHUNK_SIZE, + "Unexpected chunk size (expected {}, got {})", + CHUNK_SIZE, + chunk.len() + ); + + reduce_add_blocks(a, b, chunk); + + *a %= MOD; + *b %= MOD; + } + + unsafe fn update_block(a: &mut u32, b: &mut u32, chunk: &[u8]) { + debug_assert!( + chunk.len() <= CHUNK_SIZE, + "Unexpected chunk size (expected <= {}, got {})", + CHUNK_SIZE, + chunk.len() + ); + + for byte in reduce_add_blocks(a, b, chunk) { + *a += *byte as u32; + *b += *a; + } + + *a %= MOD; + *b %= MOD; + } + + #[inline(always)] + unsafe fn reduce_add_blocks<'a>(a: &mut u32, b: &mut u32, chunk: &'a [u8]) -> &'a [u8] { + if chunk.len() < BLOCK_SIZE { + return chunk; + } + + let blocks = chunk.chunks_exact(BLOCK_SIZE); + let blocks_remainder = blocks.remainder(); + + let one_v = _mm_set1_epi16(1); + let zero_v = _mm_set1_epi16(0); + let weight_hi_v = get_weight_hi(); + let weight_lo_v = get_weight_lo(); + + let mut p_v = _mm_set_epi32(0, 0, 0, (*a * blocks.len() as u32) as _); + let mut a_v = _mm_set_epi32(0, 0, 0, 0); + let mut b_v = _mm_set_epi32(0, 0, 0, *b as _); + + for block in blocks { + let block_ptr = block.as_ptr() as *const _; + let left_v = _mm_loadu_si128(block_ptr); + let right_v = _mm_loadu_si128(block_ptr.add(1)); + + p_v = _mm_add_epi32(p_v, a_v); + + a_v = _mm_add_epi32(a_v, _mm_sad_epu8(left_v, zero_v)); + let mad = _mm_maddubs_epi16(left_v, weight_hi_v); + b_v = _mm_add_epi32(b_v, _mm_madd_epi16(mad, one_v)); + + a_v = _mm_add_epi32(a_v, _mm_sad_epu8(right_v, zero_v)); + let mad = _mm_maddubs_epi16(right_v, weight_lo_v); + b_v = _mm_add_epi32(b_v, _mm_madd_epi16(mad, one_v)); + } + + b_v = _mm_add_epi32(b_v, _mm_slli_epi32(p_v, 5)); + + *a += reduce_add(a_v); + *b = reduce_add(b_v); + + blocks_remainder + } + + #[inline(always)] + unsafe fn reduce_add(v: __m128i) -> u32 { + let hi = _mm_unpackhi_epi64(v, v); + let sum = _mm_add_epi32(hi, v); + let hi = _mm_shuffle_epi32(sum, crate::imp::_MM_SHUFFLE(2, 3, 0, 1)); + let sum = _mm_add_epi32(sum, hi); + + _mm_cvtsi128_si32(sum) as _ + } + + #[inline(always)] + unsafe fn get_weight_lo() -> __m128i { + _mm_set_epi8(1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16) + } + + #[inline(always)] + unsafe fn get_weight_hi() -> __m128i { + _mm_set_epi8( + 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, + ) + } +} + +#[cfg(test)] +mod tests { + use rand::Rng; + + #[test] + fn zeroes() { + assert_sum_eq(&[]); + assert_sum_eq(&[0]); + assert_sum_eq(&[0, 0]); + assert_sum_eq(&[0; 100]); + assert_sum_eq(&[0; 1024]); + assert_sum_eq(&[0; 1024 * 1024]); + } + + #[test] + fn ones() { + assert_sum_eq(&[]); + assert_sum_eq(&[1]); + assert_sum_eq(&[1, 1]); + assert_sum_eq(&[1; 100]); + assert_sum_eq(&[1; 1024]); + assert_sum_eq(&[1; 1024 * 1024]); + } + + #[test] + fn random() { + let mut random = [0; 1024 * 1024]; + rand::thread_rng().fill(&mut random[..]); + + assert_sum_eq(&random[..1]); + assert_sum_eq(&random[..100]); + assert_sum_eq(&random[..1024]); + assert_sum_eq(&random[..1024 * 1024]); + } + + /// Example calculation from https://en.wikipedia.org/wiki/Adler-32. + #[test] + fn wiki() { + assert_sum_eq(b"Wikipedia"); + } + + fn assert_sum_eq(data: &[u8]) { + if let Some(update) = super::get_imp() { + let (a, b) = update(1, 0, data); + let left = u32::from(b) << 16 | u32::from(a); + let right = adler::adler32_slice(data); + + assert_eq!(left, right, "len({})", data.len()); + } + } +} diff --git a/vendor/simd-adler32/src/imp/wasm.rs b/vendor/simd-adler32/src/imp/wasm.rs new file mode 100644 index 0000000..5a9152b --- /dev/null +++ b/vendor/simd-adler32/src/imp/wasm.rs @@ -0,0 +1,217 @@ +use super::Adler32Imp; + +/// Resolves update implementation if CPU supports simd128 instructions. +pub fn get_imp() -> Option { + get_imp_inner() +} + +#[inline] +#[cfg(target_feature = "simd128")] +fn get_imp_inner() -> Option { + Some(imp::update) +} + +#[inline] +#[cfg(not(target_feature = "simd128"))] +fn get_imp_inner() -> Option { + None +} + +#[cfg(target_feature = "simd128")] +mod imp { + const MOD: u32 = 65521; + const NMAX: usize = 5552; + const BLOCK_SIZE: usize = 32; + const CHUNK_SIZE: usize = NMAX / BLOCK_SIZE * BLOCK_SIZE; + + #[cfg(target_arch = "wasm32")] + use core::arch::wasm32::*; + #[cfg(target_arch = "wasm64")] + use core::arch::wasm64::*; + + pub fn update(a: u16, b: u16, data: &[u8]) -> (u16, u16) { + update_imp(a, b, data) + } + + #[inline] + #[target_feature(enable = "simd128")] + fn update_imp(a: u16, b: u16, data: &[u8]) -> (u16, u16) { + let mut a = a as u32; + let mut b = b as u32; + + let chunks = data.chunks_exact(CHUNK_SIZE); + let remainder = chunks.remainder(); + for chunk in chunks { + update_chunk_block(&mut a, &mut b, chunk); + } + + update_block(&mut a, &mut b, remainder); + + (a as u16, b as u16) + } + + fn update_chunk_block(a: &mut u32, b: &mut u32, chunk: &[u8]) { + debug_assert_eq!( + chunk.len(), + CHUNK_SIZE, + "Unexpected chunk size (expected {}, got {})", + CHUNK_SIZE, + chunk.len() + ); + + reduce_add_blocks(a, b, chunk); + + *a %= MOD; + *b %= MOD; + } + + fn update_block(a: &mut u32, b: &mut u32, chunk: &[u8]) { + debug_assert!( + chunk.len() <= CHUNK_SIZE, + "Unexpected chunk size (expected <= {}, got {})", + CHUNK_SIZE, + chunk.len() + ); + + for byte in reduce_add_blocks(a, b, chunk) { + *a += *byte as u32; + *b += *a; + } + + *a %= MOD; + *b %= MOD; + } + + #[inline(always)] + fn reduce_add_blocks<'a>(a: &mut u32, b: &mut u32, chunk: &'a [u8]) -> &'a [u8] { + if chunk.len() < BLOCK_SIZE { + return chunk; + } + + let blocks = chunk.chunks_exact(BLOCK_SIZE); + let blocks_remainder = blocks.remainder(); + + let weight_hi_v = get_weight_hi(); + let weight_lo_v = get_weight_lo(); + + let mut p_v = u32x4(*a * blocks.len() as u32, 0, 0, 0); + let mut a_v = u32x4(0, 0, 0, 0); + let mut b_v = u32x4(*b, 0, 0, 0); + + for block in blocks { + let block_ptr = block.as_ptr() as *const v128; + let v_lo = unsafe { block_ptr.read_unaligned() }; + let v_hi = unsafe { block_ptr.add(1).read_unaligned() }; + + p_v = u32x4_add(p_v, a_v); + + a_v = u32x4_add(a_v, u32x4_extadd_quarters_u8x16(v_lo)); + let mad = i32x4_dot_i8x16(v_lo, weight_lo_v); + b_v = u32x4_add(b_v, mad); + + a_v = u32x4_add(a_v, u32x4_extadd_quarters_u8x16(v_hi)); + let mad = i32x4_dot_i8x16(v_hi, weight_hi_v); + b_v = u32x4_add(b_v, mad); + } + + b_v = u32x4_add(b_v, u32x4_shl(p_v, 5)); + + *a += reduce_add(a_v); + *b = reduce_add(b_v); + + blocks_remainder + } + + #[inline(always)] + fn i32x4_dot_i8x16(a: v128, b: v128) -> v128 { + let a_lo = u16x8_extend_low_u8x16(a); + let a_hi = u16x8_extend_high_u8x16(a); + + let b_lo = u16x8_extend_low_u8x16(b); + let b_hi = u16x8_extend_high_u8x16(b); + + let lo = i32x4_dot_i16x8(a_lo, b_lo); + let hi = i32x4_dot_i16x8(a_hi, b_hi); + + i32x4_add(lo, hi) + } + + #[inline(always)] + fn u32x4_extadd_quarters_u8x16(a: v128) -> v128 { + u32x4_extadd_pairwise_u16x8(u16x8_extadd_pairwise_u8x16(a)) + } + + #[inline(always)] + fn reduce_add(v: v128) -> u32 { + let arr: [u32; 4] = unsafe { std::mem::transmute(v) }; + let mut sum = 0u32; + for val in arr { + sum = sum.wrapping_add(val); + } + sum + } + + #[inline(always)] + fn get_weight_lo() -> v128 { + u8x16( + 32, 31, 30, 29, 28, 27, 26, 25, 24, 23, 22, 21, 20, 19, 18, 17, + ) + } + + #[inline(always)] + fn get_weight_hi() -> v128 { + u8x16(16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1) + } +} + +#[cfg(test)] +mod tests { + use rand::Rng; + + #[test] + fn zeroes() { + assert_sum_eq(&[]); + assert_sum_eq(&[0]); + assert_sum_eq(&[0, 0]); + assert_sum_eq(&[0; 100]); + assert_sum_eq(&[0; 1024]); + assert_sum_eq(&[0; 512 * 1024]); + } + + #[test] + fn ones() { + assert_sum_eq(&[]); + assert_sum_eq(&[1]); + assert_sum_eq(&[1, 1]); + assert_sum_eq(&[1; 100]); + assert_sum_eq(&[1; 1024]); + assert_sum_eq(&[1; 512 * 1024]); + } + + #[test] + fn random() { + let mut random = [0; 512 * 1024]; + rand::thread_rng().fill(&mut random[..]); + + assert_sum_eq(&random[..1]); + assert_sum_eq(&random[..100]); + assert_sum_eq(&random[..1024]); + assert_sum_eq(&random[..512 * 1024]); + } + + /// Example calculation from https://en.wikipedia.org/wiki/Adler-32. + #[test] + fn wiki() { + assert_sum_eq(b"Wikipedia"); + } + + fn assert_sum_eq(data: &[u8]) { + if let Some(update) = super::get_imp() { + let (a, b) = update(1, 0, data); + let left = u32::from(b) << 16 | u32::from(a); + let right = adler::adler32_slice(data); + + assert_eq!(left, right, "len({})", data.len()); + } + } +} diff --git a/vendor/simd-adler32/src/lib.rs b/vendor/simd-adler32/src/lib.rs new file mode 100644 index 0000000..e61dddd --- /dev/null +++ b/vendor/simd-adler32/src/lib.rs @@ -0,0 +1,310 @@ +//! # simd-adler32 +//! +//! A SIMD-accelerated Adler-32 hash algorithm implementation. +//! +//! ## Features +//! +//! - No dependencies +//! - Support `no_std` (with `default-features = false`) +//! - Runtime CPU feature detection (when `std` enabled) +//! - Blazing fast performance on as many targets as possible (currently only x86 and x86_64) +//! - Default to scalar implementation when simd not available +//! +//! ## Quick start +//! +//! > Cargo.toml +//! +//! ```toml +//! [dependencies] +//! simd-adler32 = "*" +//! ``` +//! +//! > example.rs +//! +//! ```rust +//! use simd_adler32::Adler32; +//! +//! let mut adler = Adler32::new(); +//! adler.write(b"rust is pretty cool, man"); +//! let hash = adler.finish(); +//! +//! println!("{}", hash); +//! // 1921255656 +//! ``` +//! +//! ## Feature flags +//! +//! * `std` - Enabled by default +//! +//! Enables std support, see [CPU Feature Detection](#cpu-feature-detection) for runtime +//! detection support. +//! * `nightly` +//! +//! Enables nightly features required for avx512 support. +//! +//! * `const-generics` - Enabled by default +//! +//! Enables const-generics support allowing for user-defined array hashing by value. See +//! [`Adler32Hash`] for details. +//! +//! ## Support +//! +//! **CPU Features** +//! +//! | impl | arch | feature | +//! | ---- | ---------------- | ------- | +//! | βœ… | `x86`, `x86_64` | avx512 | +//! | βœ… | `x86`, `x86_64` | avx2 | +//! | βœ… | `x86`, `x86_64` | ssse3 | +//! | βœ… | `x86`, `x86_64` | sse2 | +//! | 🚧 | `arm`, `aarch64` | neon | +//! | | `wasm32` | simd128 | +//! +//! **MSRV** `1.36.0`\*\* +//! +//! Minimum supported rust version is tested before a new version is published. [**] Feature +//! `const-generics` needs to disabled to build on rustc versions `<1.51` which can be done +//! by updating your dependency definition to the following. +//! +//! ## CPU Feature Detection +//! simd-adler32 supports both runtime and compile time CPU feature detection using the +//! `std::is_x86_feature_detected` macro when the `Adler32` struct is instantiated with +//! the `new` fn. +//! +//! Without `std` feature enabled simd-adler32 falls back to compile time feature detection +//! using `target-feature` or `target-cpu` flags supplied to rustc. See [https://rust-lang.github.io/packed_simd/perf-guide/target-feature/rustflags.html](https://rust-lang.github.io/packed_simd/perf-guide/target-feature/rustflags.html) +//! for more information. +//! +//! Feature detection tries to use the fastest supported feature first. +#![cfg_attr(not(feature = "std"), no_std)] +#![cfg_attr(feature = "nightly", feature(stdsimd, avx512_target_feature))] + +#[doc(hidden)] +pub mod hash; +#[doc(hidden)] +pub mod imp; + +pub use hash::*; +use imp::{get_imp, Adler32Imp}; + +/// An adler32 hash generator type. +#[derive(Clone)] +pub struct Adler32 { + a: u16, + b: u16, + update: Adler32Imp, +} + +impl Adler32 { + /// Constructs a new `Adler32`. + /// + /// Potential overhead here due to runtime feature detection although in testing on 100k + /// and 10k random byte arrays it was not really noticeable. + /// + /// # Examples + /// ```rust + /// use simd_adler32::Adler32; + /// + /// let mut adler = Adler32::new(); + /// ``` + pub fn new() -> Self { + Default::default() + } + + /// Constructs a new `Adler32` using existing checksum. + /// + /// Potential overhead here due to runtime feature detection although in testing on 100k + /// and 10k random byte arrays it was not really noticeable. + /// + /// # Examples + /// ```rust + /// use simd_adler32::Adler32; + /// + /// let mut adler = Adler32::from_checksum(0xdeadbeaf); + /// ``` + pub fn from_checksum(checksum: u32) -> Self { + Self { + a: checksum as u16, + b: (checksum >> 16) as u16, + update: get_imp(), + } + } + + /// Computes hash for supplied data and stores results in internal state. + pub fn write(&mut self, data: &[u8]) { + let (a, b) = (self.update)(self.a, self.b, data); + + self.a = a; + self.b = b; + } + + /// Returns the hash value for the values written so far. + /// + /// Despite its name, the method does not reset the hasher’s internal state. Additional + /// writes will continue from the current value. If you need to start a fresh hash + /// value, you will have to use `reset`. + pub fn finish(&self) -> u32 { + (u32::from(self.b) << 16) | u32::from(self.a) + } + + /// Resets the internal state. + pub fn reset(&mut self) { + self.a = 1; + self.b = 0; + } +} + +/// Compute Adler-32 hash on `Adler32Hash` type. +/// +/// # Arguments +/// * `hash` - A Adler-32 hash-able type. +/// +/// # Examples +/// ```rust +/// use simd_adler32::adler32; +/// +/// let hash = adler32(b"Adler-32"); +/// println!("{}", hash); // 800813569 +/// ``` +pub fn adler32(hash: &H) -> u32 { + hash.hash() +} + +/// A Adler-32 hash-able type. +pub trait Adler32Hash { + /// Feeds this value into `Adler32`. + fn hash(&self) -> u32; +} + +impl Default for Adler32 { + fn default() -> Self { + Self { + a: 1, + b: 0, + update: get_imp(), + } + } +} + +#[cfg(feature = "std")] +pub mod read { + //! Reader-based hashing. + //! + //! # Example + //! ```rust + //! use std::io::Cursor; + //! use simd_adler32::read::adler32; + //! + //! let mut reader = Cursor::new(b"Hello there"); + //! let hash = adler32(&mut reader).unwrap(); + //! + //! println!("{}", hash) // 800813569 + //! ``` + use crate::Adler32; + use std::io::{Read, Result}; + + /// Compute Adler-32 hash on reader until EOF. + /// + /// # Example + /// ```rust + /// use std::io::Cursor; + /// use simd_adler32::read::adler32; + /// + /// let mut reader = Cursor::new(b"Hello there"); + /// let hash = adler32(&mut reader).unwrap(); + /// + /// println!("{}", hash) // 800813569 + /// ``` + pub fn adler32(reader: &mut R) -> Result { + let mut hash = Adler32::new(); + let mut buf = [0; 4096]; + + loop { + match reader.read(&mut buf) { + Ok(0) => return Ok(hash.finish()), + Ok(n) => { + hash.write(&buf[..n]); + } + Err(err) => return Err(err), + } + } + } +} + +#[cfg(feature = "std")] +pub mod bufread { + //! BufRead-based hashing. + //! + //! Separate `BufRead` trait implemented to allow for custom buffer size optimization. + //! + //! # Example + //! ```rust + //! use std::io::{Cursor, BufReader}; + //! use simd_adler32::bufread::adler32; + //! + //! let mut reader = Cursor::new(b"Hello there"); + //! let mut reader = BufReader::new(reader); + //! let hash = adler32(&mut reader).unwrap(); + //! + //! println!("{}", hash) // 800813569 + //! ``` + use crate::Adler32; + use std::io::{BufRead, ErrorKind, Result}; + + /// Compute Adler-32 hash on buf reader until EOF. + /// + /// # Example + /// ```rust + /// use std::io::{Cursor, BufReader}; + /// use simd_adler32::bufread::adler32; + /// + /// let mut reader = Cursor::new(b"Hello there"); + /// let mut reader = BufReader::new(reader); + /// let hash = adler32(&mut reader).unwrap(); + /// + /// println!("{}", hash) // 800813569 + /// ``` + pub fn adler32(reader: &mut R) -> Result { + let mut hash = Adler32::new(); + + loop { + let consumed = match reader.fill_buf() { + Ok(buf) => { + if buf.is_empty() { + return Ok(hash.finish()); + } + + hash.write(buf); + buf.len() + } + Err(err) => match err.kind() { + ErrorKind::Interrupted => continue, + ErrorKind::UnexpectedEof => return Ok(hash.finish()), + _ => return Err(err), + }, + }; + + reader.consume(consumed); + } + } +} + +#[cfg(test)] +mod tests { + #[test] + fn test_from_checksum() { + let buf = b"rust is pretty cool man"; + let sum = 0xdeadbeaf; + + let mut simd = super::Adler32::from_checksum(sum); + let mut adler = adler::Adler32::from_checksum(sum); + + simd.write(buf); + adler.write_slice(buf); + + let simd = simd.finish(); + let scalar = adler.checksum(); + + assert_eq!(simd, scalar); + } +} -- cgit v1.2.3