crc-fast 1.10.0

World's fastest generic CRC16, CRC32, and CRC64 calculator using SIMD. Supplies a C-compatible shared library for use in other languages.
Documentation 
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
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266

//! Provides CRC-32/ISCSI calculations using a fusion of native CLMUL
//! instructions and native CRC calculation instructions on aarch64.
//!
//! https://dougallj.wordpress.com/2022/05/22/faster-crc32-on-the-apple-m1/
//!
//! Converted to Rust from the original C code generated by https://github.com/corsix/fast-crc32/
//! with the help of Claude.ai.
//!
//! Modified as necessary for this Rust implementation.
//!
//! MIT licensed.

#![cfg(target_arch = "aarch64")]

use crate::crc32::fusion::aarch64::{clmul_hi, clmul_lo, clmul_scalar};
use core::arch::aarch64::{
    __crc32cb, __crc32cd, __crc32cw, uint64x2_t, veor3q_u64, veorq_u64, vgetq_lane_u64, vld1q_u64,
    vmov_n_u64, vmull_p8, vreinterpret_p8_u64, vreinterpretq_u64_p16,
};

/// Converted to Rust from the original C code generated by https://github.com/corsix/fast-crc32/
/// with the help of Claude.ai using:
///
/// ./generate -i neon_eor3 -p crc32c -a v9s3x2e_s3
///
/// Modified as necessary for this Rust implementation.
#[inline]
#[target_feature(enable = "crc,aes,sha3")]
pub(crate) unsafe fn crc32_iscsi_eor3_v9s3x2e_s3(
    mut crc0: u32,
    mut buf: *const u8,
    mut len: usize,
) -> u32 {
    // Align to 8-byte boundary
    while len > 0 && (buf as usize & 7) != 0 {
        crc0 = __crc32cb(crc0, *buf);
        buf = buf.add(1);
        len -= 1;
    }

    // Handle 8-byte alignment
    if (buf as usize & 8) != 0 && len >= 8 {
        crc0 = __crc32cd(crc0, *(buf as *const u64));
        buf = buf.add(8);
        len -= 8;
    }

    if len >= 192 {
        let end = buf.add(len);
        let blk = len / 192;
        let klen = blk * 16;
        let buf2 = buf.add(klen * 3);
        let limit = buf.add(klen).sub(32);
        let mut crc1 = 0u32;
        let mut crc2 = 0u32;

        // First vector chunk
        let mut x0 = vld1q_u64(buf2 as *const u64);
        let mut x1 = vld1q_u64(buf2.add(16) as *const u64);
        let mut x2 = vld1q_u64(buf2.add(32) as *const u64);
        let mut x3 = vld1q_u64(buf2.add(48) as *const u64);
        let mut x4 = vld1q_u64(buf2.add(64) as *const u64);
        let mut x5 = vld1q_u64(buf2.add(80) as *const u64);
        let mut x6 = vld1q_u64(buf2.add(96) as *const u64);
        let mut x7 = vld1q_u64(buf2.add(112) as *const u64);
        let mut x8 = vld1q_u64(buf2.add(128) as *const u64);

        let k_vals: [u64; 2] = [0x7e908048, 0xc96cfdc0];
        let mut k = vld1q_u64(k_vals.as_ptr());
        let mut buf2 = buf2.add(144);

        // Main loop
        while buf <= limit {
            let y0 = clmul_lo(x0, k);
            x0 = clmul_hi(x0, k);
            let y1 = clmul_lo(x1, k);
            x1 = clmul_hi(x1, k);
            let y2 = clmul_lo(x2, k);
            x2 = clmul_hi(x2, k);
            let y3 = clmul_lo(x3, k);
            x3 = clmul_hi(x3, k);
            let y4 = clmul_lo(x4, k);
            x4 = clmul_hi(x4, k);
            let y5 = clmul_lo(x5, k);
            x5 = clmul_hi(x5, k);
            let y6 = clmul_lo(x6, k);
            x6 = clmul_hi(x6, k);
            let y7 = clmul_lo(x7, k);
            x7 = clmul_hi(x7, k);
            let y8 = clmul_lo(x8, k);
            x8 = clmul_hi(x8, k);

            x0 = veor3q_u64(x0, y0, vld1q_u64(buf2 as *const u64));
            x1 = veor3q_u64(x1, y1, vld1q_u64(buf2.add(16) as *const u64));
            x2 = veor3q_u64(x2, y2, vld1q_u64(buf2.add(32) as *const u64));
            x3 = veor3q_u64(x3, y3, vld1q_u64(buf2.add(48) as *const u64));
            x4 = veor3q_u64(x4, y4, vld1q_u64(buf2.add(64) as *const u64));
            x5 = veor3q_u64(x5, y5, vld1q_u64(buf2.add(80) as *const u64));
            x6 = veor3q_u64(x6, y6, vld1q_u64(buf2.add(96) as *const u64));
            x7 = veor3q_u64(x7, y7, vld1q_u64(buf2.add(112) as *const u64));
            x8 = veor3q_u64(x8, y8, vld1q_u64(buf2.add(128) as *const u64));

            crc0 = __crc32cd(crc0, *(buf as *const u64));
            crc1 = __crc32cd(crc1, *(buf.add(klen) as *const u64));
            crc2 = __crc32cd(crc2, *(buf.add(klen * 2) as *const u64));
            crc0 = __crc32cd(crc0, *(buf.add(8) as *const u64));
            crc1 = __crc32cd(crc1, *(buf.add(klen + 8) as *const u64));
            crc2 = __crc32cd(crc2, *(buf.add(klen * 2 + 8) as *const u64));

            buf = buf.add(16);
            buf2 = buf2.add(144);
        }

        // Reduce x0 ... x8 to just x0
        let k_vals: [u64; 2] = [0xf20c0dfe, 0x493c7d27];
        k = vld1q_u64(k_vals.as_ptr());

        let y0 = clmul_lo(x0, k);
        x0 = clmul_hi(x0, k);
        x0 = veor3q_u64(x0, y0, x1);
        x1 = x2;
        x2 = x3;
        x3 = x4;
        x4 = x5;
        x5 = x6;
        x6 = x7;
        x7 = x8;

        let y0 = clmul_lo(x0, k);
        x0 = clmul_hi(x0, k);
        let y2 = clmul_lo(x2, k);
        x2 = clmul_hi(x2, k);
        let y4 = clmul_lo(x4, k);
        x4 = clmul_hi(x4, k);
        let y6 = clmul_lo(x6, k);
        x6 = clmul_hi(x6, k);

        x0 = veor3q_u64(x0, y0, x1);
        x2 = veor3q_u64(x2, y2, x3);
        x4 = veor3q_u64(x4, y4, x5);
        x6 = veor3q_u64(x6, y6, x7);

        let k_vals: [u64; 2] = [0x3da6d0cb, 0xba4fc28e];
        k = vld1q_u64(k_vals.as_ptr());

        let y0 = clmul_lo(x0, k);
        x0 = clmul_hi(x0, k);
        let y4 = clmul_lo(x4, k);
        x4 = clmul_hi(x4, k);

        x0 = veor3q_u64(x0, y0, x2);
        x4 = veor3q_u64(x4, y4, x6);

        let k_vals: [u64; 2] = [0x740eef02, 0x9e4addf8];
        k = vld1q_u64(k_vals.as_ptr());

        let y0 = clmul_lo(x0, k);
        x0 = clmul_hi(x0, k);
        x0 = veor3q_u64(x0, y0, x4);

        // Final scalar chunk
        crc0 = __crc32cd(crc0, *(buf as *const u64));
        crc1 = __crc32cd(crc1, *(buf.add(klen) as *const u64));
        crc2 = __crc32cd(crc2, *(buf.add(klen * 2) as *const u64));
        crc0 = __crc32cd(crc0, *(buf.add(8) as *const u64));
        crc1 = __crc32cd(crc1, *(buf.add(klen + 8) as *const u64));
        crc2 = __crc32cd(crc2, *(buf.add(klen * 2 + 8) as *const u64));

        let vc0 = crc_shift_iscsi(crc0, klen * 2 + blk * 144);
        let vc1 = crc_shift_iscsi(crc1, klen + blk * 144);
        let vc2 = crc_shift_iscsi(crc2, blk * 144);
        let vc = vgetq_lane_u64(veor3q_u64(vc0, vc1, vc2), 0);

        // Reduce 128 bits to 32 bits, and multiply by x^32
        crc0 = __crc32cd(0, vgetq_lane_u64(x0, 0));
        crc0 = __crc32cd(crc0, vc ^ vgetq_lane_u64(x0, 1));

        buf = buf2;
        len = end.offset_from(buf) as usize;
    }

    if len >= 32 {
        let klen = ((len - 8) / 24) * 8;
        let mut crc1 = 0u32;
        let mut crc2 = 0u32;

        // Main loop
        loop {
            crc0 = __crc32cd(crc0, *(buf as *const u64));
            crc1 = __crc32cd(crc1, *(buf.add(klen) as *const u64));
            crc2 = __crc32cd(crc2, *(buf.add(klen * 2) as *const u64));
            buf = buf.add(8);
            len -= 24;
            if len < 32 {
                break;
            }
        }

        let vc0 = crc_shift_iscsi(crc0, klen * 2 + 8);
        let vc1 = crc_shift_iscsi(crc1, klen + 8);
        let vc = vgetq_lane_u64(veorq_u64(vc0, vc1), 0);

        // Final 8 bytes
        buf = buf.add(klen * 2);
        crc0 = crc2;
        crc0 = __crc32cd(crc0, *(buf as *const u64) ^ vc);
        buf = buf.add(8);
        len -= 8;
    }

    while len >= 8 {
        crc0 = __crc32cd(crc0, *(buf as *const u64));
        buf = buf.add(8);
        len -= 8;
    }

    while len > 0 {
        crc0 = __crc32cb(crc0, *buf);
        buf = buf.add(1);
        len -= 1;
    }

    crc0
}

#[inline]
#[target_feature(enable = "aes")]
unsafe fn crc_shift_iscsi(crc: u32, nbytes: usize) -> uint64x2_t {
    clmul_scalar(crc, xnmodp_crc32_iscsi((nbytes * 8 - 33) as u64))
}

// x^n mod P, in log(n) time
#[inline]
#[target_feature(enable = "crc,aes")]
unsafe fn xnmodp_crc32_iscsi(mut n: u64) -> u32 {
    let mut stack = !1u64;
    let mut acc: u32;
    let mut low: u32;

    while n > 191 {
        stack = (stack << 1) + (n & 1);
        n = (n >> 1) - 16;
    }
    stack = !stack;
    acc = 0x80000000u32 >> (n & 31);
    n >>= 5;

    while n > 0 {
        // ARM CRC32 instruction
        acc = __crc32cw(acc, 0);
        n -= 1;
    }

    while {
        low = (stack & 1) as u32;
        stack >>= 1;
        stack != 0
    } {
        // Convert to polynomial type and square it
        let x = vreinterpret_p8_u64(vmov_n_u64(acc as u64));
        let squared = vmull_p8(x, x);
        let y = vgetq_lane_u64(vreinterpretq_u64_p16(squared), 0);
        acc = __crc32cd(0, y << low);
    }
    acc
}