| 1 |
use super::Adler32Imp; |
| 2 |
|
| 3 |
/// Resolves update implementation if CPU supports avx512f and avx512bw instructions. |
| 4 |
pub fn get_imp() -> Option<Adler32Imp> { |
| 5 |
get_imp_inner() |
| 6 |
} |
| 7 |
|
| 8 |
#[inline] |
| 9 |
#[cfg(all(feature = "std", feature = "nightly", target_arch = "arm"))] |
| 10 |
fn get_imp_inner() -> Option<Adler32Imp> { |
| 11 |
if std::is_arm_feature_detected("neon") { |
| 12 |
Some(imp::update) |
| 13 |
} else { |
| 14 |
None |
| 15 |
} |
| 16 |
} |
| 17 |
|
| 18 |
#[inline] |
| 19 |
#[cfg(all(feature = "std", feature = "nightly", target_arch = "aarch64"))] |
| 20 |
fn get_imp_inner() -> Option<Adler32Imp> { |
| 21 |
if std::is_aarch64_feature_detected("neon") { |
| 22 |
Some(imp::update) |
| 23 |
} else { |
| 24 |
None |
| 25 |
} |
| 26 |
} |
| 27 |
|
| 28 |
#[inline] |
| 29 |
#[cfg(all( |
| 30 |
feature = "nightly", |
| 31 |
target_feature = "neon", |
| 32 |
not(all(feature = "std", any(target_arch = "arm", target_arch = "aarch64"))) |
| 33 |
))] |
| 34 |
fn get_imp_inner() -> Option<Adler32Imp> { |
| 35 |
Some(imp::update) |
| 36 |
} |
| 37 |
|
| 38 |
#[inline] |
| 39 |
#[cfg(all( |
| 40 |
not(target_feature = "neon"), |
| 41 |
not(all( |
| 42 |
feature = "std", |
| 43 |
feature = "nightly", |
| 44 |
any(target_arch = "arm", target_arch = "aarch64") |
| 45 |
)) |
| 46 |
))] |
| 47 |
fn get_imp_inner() -> Option<Adler32Imp> { |
| 48 |
None |
| 49 |
} |
| 50 |
|
| 51 |
#[cfg(all( |
| 52 |
feature = "nightly", |
| 53 |
any(target_arch = "arm", target_arch = "aarch64"), |
| 54 |
any(feature = "std", target_feature = "neon") |
| 55 |
))] |
| 56 |
mod imp { |
| 57 |
const MOD: u32 = 65521; |
| 58 |
const NMAX: usize = 5552; |
| 59 |
const BLOCK_SIZE: usize = 64; |
| 60 |
const CHUNK_SIZE: usize = NMAX / BLOCK_SIZE * BLOCK_SIZE; |
| 61 |
|
| 62 |
#[cfg(target_arch = "aarch64")] |
| 63 |
use core::arch::aarch64::*; |
| 64 |
#[cfg(target_arch = "arm")] |
| 65 |
use core::arch::arm::*; |
| 66 |
|
| 67 |
pub fn update(a: u16, b: u16, data: &[u8]) -> (u16, u16) { |
| 68 |
unsafe { update_imp(a, b, data) } |
| 69 |
} |
| 70 |
|
| 71 |
#[inline] |
| 72 |
#[target_feature(enable = "neon")] |
| 73 |
unsafe fn update_imp(a: u16, b: u16, data: &[u8]) -> (u16, u16) { |
| 74 |
let mut a = a as u32; |
| 75 |
let mut b = b as u32; |
| 76 |
|
| 77 |
let chunks = data.chunks_exact(CHUNK_SIZE); |
| 78 |
let remainder = chunks.remainder(); |
| 79 |
for chunk in chunks { |
| 80 |
update_chunk_block(&mut a, &mut b, chunk); |
| 81 |
} |
| 82 |
|
| 83 |
update_block(&mut a, &mut b, remainder); |
| 84 |
|
| 85 |
(a as u16, b as u16) |
| 86 |
} |
| 87 |
|
| 88 |
#[inline] |
| 89 |
unsafe fn update_chunk_block(a: &mut u32, b: &mut u32, chunk: &[u8]) { |
| 90 |
debug_assert_eq!( |
| 91 |
chunk.len(), |
| 92 |
CHUNK_SIZE, |
| 93 |
"Unexpected chunk size (expected {}, got {})", |
| 94 |
CHUNK_SIZE, |
| 95 |
chunk.len() |
| 96 |
); |
| 97 |
|
| 98 |
reduce_add_blocks(a, b, chunk); |
| 99 |
|
| 100 |
*a %= MOD; |
| 101 |
*b %= MOD; |
| 102 |
} |
| 103 |
|
| 104 |
#[inline] |
| 105 |
unsafe fn update_block(a: &mut u32, b: &mut u32, chunk: &[u8]) { |
| 106 |
debug_assert!( |
| 107 |
chunk.len() <= CHUNK_SIZE, |
| 108 |
"Unexpected chunk size (expected <= {}, got {})", |
| 109 |
CHUNK_SIZE, |
| 110 |
chunk.len() |
| 111 |
); |
| 112 |
|
| 113 |
for byte in reduce_add_blocks(a, b, chunk) { |
| 114 |
*a += *byte as u32; |
| 115 |
*b += *a; |
| 116 |
} |
| 117 |
|
| 118 |
*a %= MOD; |
| 119 |
*b %= MOD; |
| 120 |
} |
| 121 |
|
| 122 |
#[inline(always)] |
| 123 |
unsafe fn reduce_add_blocks<'a>(a: &mut u32, b: &mut u32, chunk: &'a [u8]) -> &'a [u8] { |
| 124 |
if chunk.len() < BLOCK_SIZE { |
| 125 |
return chunk; |
| 126 |
} |
| 127 |
|
| 128 |
let blocks = chunk.chunks_exact(BLOCK_SIZE); |
| 129 |
let blocks_remainder = blocks.remainder(); |
| 130 |
|
| 131 |
let one_v = _mm512_set1_epi16(1); |
| 132 |
let zero_v = _mm512_setzero_si512(); |
| 133 |
let weights = get_weights(); |
| 134 |
|
| 135 |
let p_v = (*a * blocks.len() as u32) as _; |
| 136 |
let mut p_v = _mm512_set_epi32(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, p_v); |
| 137 |
let mut a_v = _mm512_setzero_si512(); |
| 138 |
let mut b_v = _mm512_set_epi32(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, *b as _); |
| 139 |
|
| 140 |
for block in blocks { |
| 141 |
let block_ptr = block.as_ptr() as *const _; |
| 142 |
let block = _mm512_loadu_si512(block_ptr); |
| 143 |
|
| 144 |
p_v = _mm512_add_epi32(p_v, a_v); |
| 145 |
|
| 146 |
a_v = _mm512_add_epi32(a_v, _mm512_sad_epu8(block, zero_v)); |
| 147 |
let mad = _mm512_maddubs_epi16(block, weights); |
| 148 |
b_v = _mm512_add_epi32(b_v, _mm512_madd_epi16(mad, one_v)); |
| 149 |
} |
| 150 |
|
| 151 |
b_v = _mm512_add_epi32(b_v, _mm512_slli_epi32(p_v, 6)); |
| 152 |
|
| 153 |
*a += reduce_add(a_v); |
| 154 |
*b = reduce_add(b_v); |
| 155 |
|
| 156 |
blocks_remainder |
| 157 |
} |
| 158 |
|
| 159 |
#[inline(always)] |
| 160 |
unsafe fn reduce_add(v: __m512i) -> u32 { |
| 161 |
let v: [__m256i; 2] = core::mem::transmute(v); |
| 162 |
|
| 163 |
reduce_add_256(v[0]) + reduce_add_256(v[1]) |
| 164 |
} |
| 165 |
|
| 166 |
#[inline(always)] |
| 167 |
unsafe fn reduce_add_256(v: __m256i) -> u32 { |
| 168 |
let v: [__m128i; 2] = core::mem::transmute(v); |
| 169 |
let sum = _mm_add_epi32(v[0], v[1]); |
| 170 |
let hi = _mm_unpackhi_epi64(sum, sum); |
| 171 |
|
| 172 |
let sum = _mm_add_epi32(hi, sum); |
| 173 |
let hi = _mm_shuffle_epi32(sum, crate::imp::_MM_SHUFFLE(2, 3, 0, 1)); |
| 174 |
|
| 175 |
let sum = _mm_add_epi32(sum, hi); |
| 176 |
let sum = _mm_cvtsi128_si32(sum) as _; |
| 177 |
|
| 178 |
sum |
| 179 |
} |
| 180 |
|
| 181 |
#[inline(always)] |
| 182 |
unsafe fn get_weights() -> __m512i { |
| 183 |
_mm512_set_epi8( |
| 184 |
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, |
| 185 |
24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, |
| 186 |
45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, |
| 187 |
) |
| 188 |
} |
| 189 |
} |
| 190 |
|
| 191 |
#[cfg(test)] |
| 192 |
mod tests { |
| 193 |
use rand::Rng; |
| 194 |
|
| 195 |
#[test] |
| 196 |
fn zeroes() { |
| 197 |
assert_sum_eq(&[]); |
| 198 |
assert_sum_eq(&[0]); |
| 199 |
assert_sum_eq(&[0, 0]); |
| 200 |
assert_sum_eq(&[0; 100]); |
| 201 |
assert_sum_eq(&[0; 1024]); |
| 202 |
assert_sum_eq(&[0; 1024 * 1024]); |
| 203 |
} |
| 204 |
|
| 205 |
#[test] |
| 206 |
fn ones() { |
| 207 |
assert_sum_eq(&[]); |
| 208 |
assert_sum_eq(&[1]); |
| 209 |
assert_sum_eq(&[1, 1]); |
| 210 |
assert_sum_eq(&[1; 100]); |
| 211 |
assert_sum_eq(&[1; 1024]); |
| 212 |
assert_sum_eq(&[1; 1024 * 1024]); |
| 213 |
} |
| 214 |
|
| 215 |
#[test] |
| 216 |
fn random() { |
| 217 |
let mut random = [0; 1024 * 1024]; |
| 218 |
rand::thread_rng().fill(&mut random[..]); |
| 219 |
|
| 220 |
assert_sum_eq(&random[..1]); |
| 221 |
assert_sum_eq(&random[..100]); |
| 222 |
assert_sum_eq(&random[..1024]); |
| 223 |
assert_sum_eq(&random[..1024 * 1024]); |
| 224 |
} |
| 225 |
|
| 226 |
/// Example calculation from https://en.wikipedia.org/wiki/Adler-32. |
| 227 |
#[test] |
| 228 |
fn wiki() { |
| 229 |
assert_sum_eq(b"Wikipedia"); |
| 230 |
} |
| 231 |
|
| 232 |
fn assert_sum_eq(data: &[u8]) { |
| 233 |
if let Some(update) = super::get_imp() { |
| 234 |
let (a, b) = update(1, 0, data); |
| 235 |
let left = u32::from(b) << 16 | u32::from(a); |
| 236 |
let right = adler::adler32_slice(data); |
| 237 |
|
| 238 |
assert_eq!(left, right, "len({})", data.len()); |
| 239 |
} |
| 240 |
} |
| 241 |
} |