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/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 ++++++++++++++++++++++++++++++ 8 files changed, 1458 insertions(+) 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 (limited to 'vendor/simd-adler32/src/imp') 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()); + } + } +} -- cgit v1.2.3