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
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
use crate::Adler32;
use std::ops::{AddAssign, MulAssign, RemAssign};

impl Adler32 {
    pub(crate) fn compute(&mut self, bytes: &[u8]) {
        // The basic algorithm is, for every byte:
        //   a = (a + byte) % MOD
        //   b = (b + a) % MOD
        // where MOD = 65521.
        //
        // For efficiency, we can defer the `% MOD` operations as long as neither a nor b overflows:
        // - Between calls to `write`, we ensure that a and b are always in range 0..MOD.
        // - We use 32-bit arithmetic in this function.
        // - Therefore, a and b must not increase by more than 2^32-MOD without performing a `% MOD`
        //   operation.
        //
        // According to Wikipedia, b is calculated as follows for non-incremental checksumming:
        //   b = n×D1 + (n−1)×D2 + (n−2)×D3 + ... + Dn + n*1 (mod 65521)
        // Where n is the number of bytes and Di is the i-th Byte. We need to change this to account
        // for the previous values of a and b, as well as treat every input Byte as being 255:
        //   b_inc = n×255 + (n-1)×255 + ... + 255 + n*65520
        // Or in other words:
        //   b_inc = n*65520 + n(n+1)/2*255
        // The max chunk size is thus the largest value of n so that b_inc <= 2^32-65521.
        //   2^32-65521 = n*65520 + n(n+1)/2*255
        // Plugging this into an equation solver since I can't math gives n = 5552.18..., so 5552.
        //
        // On top of the optimization outlined above, the algorithm can also be parallelized with a
        // bit more work:
        //
        // Note that b is a linear combination of a vector of input bytes (D1, ..., Dn).
        //
        // If we fix some value k<N and rewrite indices 1, ..., N as
        //
        //   1_1, 1_2, ..., 1_k, 2_1, ..., 2_k, ..., (N/k)_k,
        //
        // then we can express a and b in terms of sums of smaller sequences kb and ka:
        //
        //   ka(j) := D1_j + D2_j + ... + D(N/k)_j where j <= k
        //   kb(j) := (N/k)*D1_j + (N/k-1)*D2_j + ... + D(N/k)_j where j <= k
        //
        //  a = ka(1) + ka(2) + ... + ka(k) + 1
        //  b = k*(kb(1) + kb(2) + ... + kb(k)) - 1*ka(2) - ...  - (k-1)*ka(k) + N
        //
        // We use this insight to unroll the main loop and process k=4 bytes at a time.
        // The resulting code is highly amenable to SIMD acceleration, although the immediate speedups
        // stem from increased pipeline parallelism rather than auto-vectorization.
        //
        // This technique is described in-depth (here:)[https://software.intel.com/content/www/us/\
        // en/develop/articles/fast-computation-of-fletcher-checksums.html]

        const MOD: u32 = 65521;
        const CHUNK_SIZE: usize = 5552 * 4;

        let mut a = u32::from(self.a);
        let mut b = u32::from(self.b);
        let mut a_vec = U32X4([0; 4]);
        let mut b_vec = a_vec;

        let (bytes, remainder) = bytes.split_at(bytes.len() - bytes.len() % 4);

        // iterate over 4 bytes at a time
        let chunk_iter = bytes.chunks_exact(CHUNK_SIZE);
        let remainder_chunk = chunk_iter.remainder();
        for chunk in chunk_iter {
            for byte_vec in chunk.chunks_exact(4) {
                let val = U32X4::from(byte_vec);
                a_vec += val;
                b_vec += a_vec;
            }

            b += CHUNK_SIZE as u32 * a;
            a_vec %= MOD;
            b_vec %= MOD;
            b %= MOD;
        }
        // special-case the final chunk because it may be shorter than the rest
        for byte_vec in remainder_chunk.chunks_exact(4) {
            let val = U32X4::from(byte_vec);
            a_vec += val;
            b_vec += a_vec;
        }
        b += remainder_chunk.len() as u32 * a;
        a_vec %= MOD;
        b_vec %= MOD;
        b %= MOD;

        // combine the sub-sum results into the main sum
        b_vec *= 4;
        b_vec.0[1] += MOD - a_vec.0[1];
        b_vec.0[2] += (MOD - a_vec.0[2]) * 2;
        b_vec.0[3] += (MOD - a_vec.0[3]) * 3;
        for &av in a_vec.0.iter() {
            a += av;
        }
        for &bv in b_vec.0.iter() {
            b += bv;
        }

        // iterate over the remaining few bytes in serial
        for &byte in remainder.iter() {
            a += u32::from(byte);
            b += a;
        }

        self.a = (a % MOD) as u16;
        self.b = (b % MOD) as u16;
    }
}

#[derive(Copy, Clone)]
struct U32X4([u32; 4]);

impl U32X4 {
    #[inline]
    fn from(bytes: &[u8]) -> Self {
        U32X4([
            u32::from(bytes[0]),
            u32::from(bytes[1]),
            u32::from(bytes[2]),
            u32::from(bytes[3]),
        ])
    }
}

impl AddAssign<Self> for U32X4 {
    #[inline]
    fn add_assign(&mut self, other: Self) {
        // Implement this in a primitive manner to help out the compiler a bit.
        self.0[0] += other.0[0];
        self.0[1] += other.0[1];
        self.0[2] += other.0[2];
        self.0[3] += other.0[3];
    }
}

impl RemAssign<u32> for U32X4 {
    #[inline]
    fn rem_assign(&mut self, quotient: u32) {
        self.0[0] %= quotient;
        self.0[1] %= quotient;
        self.0[2] %= quotient;
        self.0[3] %= quotient;
    }
}

impl MulAssign<u32> for U32X4 {
    #[inline]
    fn mul_assign(&mut self, rhs: u32) {
        self.0[0] *= rhs;
        self.0[1] *= rhs;
        self.0[2] *= rhs;
        self.0[3] *= rhs;
    }
}