simd_adler32/imp/
wasm.rs

1use super::Adler32Imp;
2
3/// Resolves update implementation if CPU supports simd128 instructions.
4pub fn get_imp() -> Option<Adler32Imp> {
5  get_imp_inner()
6}
7
8#[inline]
9#[cfg(all(
10  target_feature = "simd128",
11  any(
12    target_arch = "wasm32",
13    all(feature = "nightly", target_arch = "wasm64")
14  )
15))]
16fn get_imp_inner() -> Option<Adler32Imp> {
17  Some(imp::update)
18}
19
20#[inline]
21#[cfg(not(all(
22  target_feature = "simd128",
23  any(
24    target_arch = "wasm32",
25    all(feature = "nightly", target_arch = "wasm64")
26  )
27)))]
28fn get_imp_inner() -> Option<Adler32Imp> {
29  None
30}
31
32#[cfg(all(
33  target_feature = "simd128",
34  any(
35    target_arch = "wasm32",
36    all(feature = "nightly", target_arch = "wasm64")
37  )
38))]
39mod imp {
40  const MOD: u32 = 65521;
41  const NMAX: usize = 5552;
42  const BLOCK_SIZE: usize = 32;
43  const CHUNK_SIZE: usize = NMAX / BLOCK_SIZE * BLOCK_SIZE;
44
45  #[cfg(target_arch = "wasm32")]
46  use core::arch::wasm32::*;
47  #[cfg(target_arch = "wasm64")]
48  use core::arch::wasm64::*;
49
50  pub fn update(a: u16, b: u16, data: &[u8]) -> (u16, u16) {
51    update_imp(a, b, data)
52  }
53
54  #[inline]
55  #[target_feature(enable = "simd128")]
56  fn update_imp(a: u16, b: u16, data: &[u8]) -> (u16, u16) {
57    let mut a = a as u32;
58    let mut b = b as u32;
59
60    let chunks = data.chunks_exact(CHUNK_SIZE);
61    let remainder = chunks.remainder();
62    for chunk in chunks {
63      update_chunk_block(&mut a, &mut b, chunk);
64    }
65
66    update_block(&mut a, &mut b, remainder);
67
68    (a as u16, b as u16)
69  }
70
71  fn update_chunk_block(a: &mut u32, b: &mut u32, chunk: &[u8]) {
72    debug_assert_eq!(
73      chunk.len(),
74      CHUNK_SIZE,
75      "Unexpected chunk size (expected {}, got {})",
76      CHUNK_SIZE,
77      chunk.len()
78    );
79
80    reduce_add_blocks(a, b, chunk);
81
82    *a %= MOD;
83    *b %= MOD;
84  }
85
86  fn update_block(a: &mut u32, b: &mut u32, chunk: &[u8]) {
87    debug_assert!(
88      chunk.len() <= CHUNK_SIZE,
89      "Unexpected chunk size (expected <= {}, got {})",
90      CHUNK_SIZE,
91      chunk.len()
92    );
93
94    for byte in reduce_add_blocks(a, b, chunk) {
95      *a += *byte as u32;
96      *b += *a;
97    }
98
99    *a %= MOD;
100    *b %= MOD;
101  }
102
103  #[inline(always)]
104  fn reduce_add_blocks<'a>(a: &mut u32, b: &mut u32, chunk: &'a [u8]) -> &'a [u8] {
105    if chunk.len() < BLOCK_SIZE {
106      return chunk;
107    }
108
109    let blocks = chunk.chunks_exact(BLOCK_SIZE);
110    let blocks_remainder = blocks.remainder();
111
112    let weight_hi_v = get_weight_hi();
113    let weight_lo_v = get_weight_lo();
114
115    let mut p_v = u32x4(*a * blocks.len() as u32, 0, 0, 0);
116    let mut a_v = u32x4(0, 0, 0, 0);
117    let mut b_v = u32x4(*b, 0, 0, 0);
118
119    for block in blocks {
120      let block_ptr = block.as_ptr() as *const v128;
121      let v_lo = unsafe { block_ptr.read_unaligned() };
122      let v_hi = unsafe { block_ptr.add(1).read_unaligned() };
123
124      p_v = u32x4_add(p_v, a_v);
125
126      a_v = u32x4_add(a_v, u32x4_extadd_quarters_u8x16(v_lo));
127      let mad = i32x4_dot_i8x16(v_lo, weight_lo_v);
128      b_v = u32x4_add(b_v, mad);
129
130      a_v = u32x4_add(a_v, u32x4_extadd_quarters_u8x16(v_hi));
131      let mad = i32x4_dot_i8x16(v_hi, weight_hi_v);
132      b_v = u32x4_add(b_v, mad);
133    }
134
135    b_v = u32x4_add(b_v, u32x4_shl(p_v, 5));
136
137    *a += reduce_add(a_v);
138    *b = reduce_add(b_v);
139
140    blocks_remainder
141  }
142
143  #[inline(always)]
144  fn i32x4_dot_i8x16(a: v128, b: v128) -> v128 {
145    let a_lo = u16x8_extend_low_u8x16(a);
146    let a_hi = u16x8_extend_high_u8x16(a);
147
148    let b_lo = u16x8_extend_low_u8x16(b);
149    let b_hi = u16x8_extend_high_u8x16(b);
150
151    let lo = i32x4_dot_i16x8(a_lo, b_lo);
152    let hi = i32x4_dot_i16x8(a_hi, b_hi);
153
154    i32x4_add(lo, hi)
155  }
156
157  #[inline(always)]
158  fn u32x4_extadd_quarters_u8x16(a: v128) -> v128 {
159    u32x4_extadd_pairwise_u16x8(u16x8_extadd_pairwise_u8x16(a))
160  }
161
162  #[inline(always)]
163  fn reduce_add(v: v128) -> u32 {
164    let arr: [u32; 4] = unsafe { core::mem::transmute(v) };
165    let mut sum = 0u32;
166    for val in arr {
167      sum = sum.wrapping_add(val);
168    }
169    sum
170  }
171
172  #[inline(always)]
173  fn get_weight_lo() -> v128 {
174    u8x16(
175      32, 31, 30, 29, 28, 27, 26, 25, 24, 23, 22, 21, 20, 19, 18, 17,
176    )
177  }
178
179  #[inline(always)]
180  fn get_weight_hi() -> v128 {
181    u8x16(16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1)
182  }
183}
184
185#[cfg(test)]
186mod tests {
187  use rand::Rng;
188
189  #[test]
190  fn zeroes() {
191    assert_sum_eq(&[]);
192    assert_sum_eq(&[0]);
193    assert_sum_eq(&[0, 0]);
194    assert_sum_eq(&[0; 100]);
195    assert_sum_eq(&[0; 1024]);
196    assert_sum_eq(&[0; 512 * 1024]);
197  }
198
199  #[test]
200  fn ones() {
201    assert_sum_eq(&[]);
202    assert_sum_eq(&[1]);
203    assert_sum_eq(&[1, 1]);
204    assert_sum_eq(&[1; 100]);
205    assert_sum_eq(&[1; 1024]);
206    assert_sum_eq(&[1; 512 * 1024]);
207  }
208
209  #[test]
210  fn random() {
211    let mut random = [0; 512 * 1024];
212    rand::thread_rng().fill(&mut random[..]);
213
214    assert_sum_eq(&random[..1]);
215    assert_sum_eq(&random[..100]);
216    assert_sum_eq(&random[..1024]);
217    assert_sum_eq(&random[..512 * 1024]);
218  }
219
220  /// Example calculation from https://en.wikipedia.org/wiki/Adler-32.
221  #[test]
222  fn wiki() {
223    assert_sum_eq(b"Wikipedia");
224  }
225
226  fn assert_sum_eq(data: &[u8]) {
227    if let Some(update) = super::get_imp() {
228      let (a, b) = update(1, 0, data);
229      let left = u32::from(b) << 16 | u32::from(a);
230      let right = adler::adler32_slice(data);
231
232      assert_eq!(left, right, "len({})", data.len());
233    }
234  }
235}
simd_adler32/imp/wasm.rs

simd_adler32/imp/
wasm.rs
