vyre 0.4.0

GPU compute intermediate representation with a standard operation library
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
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306

//! XXH64 and XXH3-64 CPU reference implementations with seed 0.

// accumulate.rs
pub fn accumulate(acc: &mut [u64; 8], stripe: &[u8], secret_offset: usize) {
    for i in 0..8 {
        let lane = read64(stripe, i * 8);
        let value = lane ^ read_secret64(secret_offset + i * 8);
        acc[i ^ 1] = acc[i ^ 1].wrapping_add(lane);
        acc[i] = acc[i].wrapping_add((value & 0xffff_ffff).wrapping_mul(value >> 32));
    }
}

// accumulate_block.rs
pub fn accumulate_block(acc: &mut [u64; 8], block: &[u8]) {
    for (stripe, bytes) in block.chunks_exact(64).enumerate() {
        accumulate(acc, bytes, stripe * 8);
    }
}

// avalanche.rs
pub fn avalanche(mut x: u64) -> u64 {
    x ^= x >> 37;
    x = x.wrapping_mul(PMX1);
    x ^ (x >> 32)
}

// avalanche_xxh64.rs
pub fn avalanche_xxh64(mut x: u64) -> u64 {
    x ^= x >> 33;
    x = x.wrapping_mul(P64_2);
    x ^= x >> 29;
    x = x.wrapping_mul(P64_3);
    x ^ (x >> 32)
}

// consume_tail.rs
pub fn consume_tail(input: &[u8], mut index: usize, mut acc: u64) -> u64 {
    while index + 8 <= input.len() {
        acc ^= round(0, read64(input, index));
        acc = acc.rotate_left(27).wrapping_mul(P64_1).wrapping_add(P64_4);
        index += 8;
    }
    if index + 4 <= input.len() {
        acc ^= u64::from(read32(input, index)).wrapping_mul(P64_1);
        acc = acc.rotate_left(23).wrapping_mul(P64_2).wrapping_add(P64_3);
        index += 4;
    }
    while index < input.len() {
        acc ^= u64::from(input[index]).wrapping_mul(P64_5);
        acc = acc.rotate_left(11).wrapping_mul(P64_1);
        index += 1;
    }
    avalanche_xxh64(acc)
}

// final_merge.rs
pub fn final_merge(acc: &[u64; 8], init: u64, secret_offset: usize) -> u64 {
    let mut result = init;
    for i in 0..4 {
        let product = ((acc[i * 2] ^ read_secret64(secret_offset + i * 16)) as u128)
            .wrapping_mul((acc[i * 2 + 1] ^ read_secret64(secret_offset + i * 16 + 8)) as u128);
        result = result.wrapping_add((product as u64) ^ (product >> 64) as u64);
    }
    avalanche(result)
}

// merge_acc.rs
pub fn merge_acc(acc: u64, lane: u64) -> u64 {
    (acc ^ round(0, lane))
        .wrapping_mul(P64_1)
        .wrapping_add(P64_4)
}

// mix16.rs
pub fn mix16(input: &[u8], offset: usize, secret_offset: usize) -> u64 {
    let lhs = read64(input, offset) ^ read_secret64(secret_offset);
    let rhs = read64(input, offset + 8) ^ read_secret64(secret_offset + 8);
    let product = (lhs as u128).wrapping_mul(rhs as u128);
    (product as u64) ^ (product >> 64) as u64
}

// p32_1.rs
pub const P32_1: u64 = 0x9e37_79b1;

// p32_2.rs
pub const P32_2: u64 = 0x85eb_ca77;

// p32_3.rs
pub const P32_3: u64 = 0xc2b2_ae3d;

// p64_1.rs
pub const P64_1: u64 = 0x9e37_79b1_85eb_ca87;

// p64_2.rs
pub const P64_2: u64 = 0xc2b2_ae3d_27d4_eb4f;

// p64_3.rs
pub const P64_3: u64 = 0x1656_67b1_9e37_79f9;

// p64_4.rs
pub const P64_4: u64 = 0x85eb_ca77_c2b2_ae63;

// p64_5.rs
pub const P64_5: u64 = 0x27d4_eb2f_1656_67c5;

// pmx1.rs
pub const PMX1: u64 = 0x1656_6791_9e37_79f9;

// pmx2.rs
pub const PMX2: u64 = 0x9fb2_1c65_1e98_df25;

// read32.rs
pub fn read32(bytes: &[u8], offset: usize) -> u32 {
    u32::from_le_bytes([
        bytes[offset],
        bytes[offset + 1],
        bytes[offset + 2],
        bytes[offset + 3],
    ])
}

// read64.rs
pub fn read64(bytes: &[u8], offset: usize) -> u64 {
    u64::from_le_bytes([
        bytes[offset],
        bytes[offset + 1],
        bytes[offset + 2],
        bytes[offset + 3],
        bytes[offset + 4],
        bytes[offset + 5],
        bytes[offset + 6],
        bytes[offset + 7],
    ])
}

// read_secret32.rs
pub fn read_secret32(offset: usize) -> u32 {
    read32(&SECRET, offset)
}

// read_secret64.rs
pub fn read_secret64(offset: usize) -> u64 {
    read64(&SECRET, offset)
}

// round.rs
pub fn round(acc: u64, lane: u64) -> u64 {
    acc.wrapping_add(lane.wrapping_mul(P64_2))
        .rotate_left(31)
        .wrapping_mul(P64_1)
}

// scramble.rs
pub fn scramble(acc: &mut [u64; 8]) {
    for i in 0..8 {
        acc[i] ^= acc[i] >> 47;
        acc[i] ^= read_secret64(SECRET.len() - 64 + i * 8);
        acc[i] = acc[i].wrapping_mul(P32_1);
    }
}

// secret.rs
pub const SECRET: [u8; 192] = [
    0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, 0xf7, 0x21, 0xad, 0x1c,
    0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f,
    0xcb, 0x79, 0xe6, 0x4e, 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21,
    0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, 0x81, 0x3a, 0x26, 0x4c,
    0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3,
    0x71, 0x64, 0x48, 0x97, 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8,
    0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, 0xc7, 0x0b, 0x4f, 0x1d,
    0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64,
    0xea, 0xc5, 0xac, 0x83, 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, 0xfa, 0x13, 0x63, 0xeb,
    0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, 0xd3, 0x16, 0x55, 0x26, 0x29, 0xd4, 0x68, 0x9e,
    0x2b, 0x16, 0xbe, 0x58, 0x7d, 0x47, 0xa1, 0xfc, 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce,
    0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e,
];

// xxh3_large.rs
pub fn xxh3_large(input: &[u8]) -> u64 {
    let mut acc = [P32_3, P64_1, P64_2, P64_3, P64_4, P32_2, P64_5, P32_1];
    let stripes_per_block = (SECRET.len() - 64) / 8;
    let block_size = stripes_per_block * 64;
    let mut offset = 0usize;
    while offset + block_size < input.len() {
        accumulate_block(&mut acc, &input[offset..offset + block_size]);
        scramble(&mut acc);
        offset += block_size;
    }
    let last = &input[offset..];
    let full_stripes = (last.len() - 1) / 64;
    for stripe in 0..full_stripes {
        accumulate(&mut acc, &last[stripe * 64..stripe * 64 + 64], stripe * 8);
    }
    accumulate(&mut acc, &input[input.len() - 64..], SECRET.len() - 71);
    final_merge(&acc, (input.len() as u64).wrapping_mul(P64_1), 11)
}

// xxh3_len_129_to_240.rs
pub fn xxh3_len_129_to_240(input: &[u8]) -> u64 {
    let mut acc = (input.len() as u64).wrapping_mul(P64_1);
    let chunks = input.len() >> 4;
    for i in 0..8 {
        acc = acc.wrapping_add(mix16(input, i * 16, i * 16));
    }
    acc = avalanche(acc);
    for i in 8..chunks {
        acc = acc.wrapping_add(mix16(input, i * 16, (i - 8) * 16 + 3));
    }
    avalanche(acc.wrapping_add(mix16(input, input.len() - 16, 119)))
}

// xxh3_len_17_to_128.rs
pub fn xxh3_len_17_to_128(input: &[u8]) -> u64 {
    let mut acc = (input.len() as u64).wrapping_mul(P64_1);
    let rounds = ((input.len() - 1) >> 5) + 1;
    for i in (0..rounds).rev() {
        acc = acc.wrapping_add(mix16(input, i * 16, i * 32));
        acc = acc.wrapping_add(mix16(input, input.len() - i * 16 - 16, i * 32 + 16));
    }
    avalanche(acc)
}

// xxh3_len_1_to_3.rs
pub fn xxh3_len_1_to_3(input: &[u8]) -> u64 {
    let combined = u32::from(input[input.len() - 1])
        | ((input.len() as u32) << 8)
        | (u32::from(input[0]) << 16)
        | (u32::from(input[input.len() >> 1]) << 24);
    let value = u64::from(read_secret32(0) ^ read_secret32(4)) ^ u64::from(combined);
    avalanche_xxh64(value)
}

// xxh3_len_4_to_8.rs
pub fn xxh3_len_4_to_8(input: &[u8]) -> u64 {
    let first = read32(input, 0);
    let last = read32(input, input.len() - 4);
    let combined = u64::from(last) | (u64::from(first) << 32);
    let mut value = (read_secret64(8) ^ read_secret64(16)) ^ combined;
    value ^= value.rotate_left(49) ^ value.rotate_left(24);
    value = value.wrapping_mul(PMX2);
    value ^= (value >> 35).wrapping_add(input.len() as u64);
    value = value.wrapping_mul(PMX2);
    value ^ (value >> 28)
}

// xxh3_len_9_to_16.rs
pub fn xxh3_len_9_to_16(input: &[u8]) -> u64 {
    let low = (read_secret64(24) ^ read_secret64(32)) ^ read64(input, 0);
    let high = (read_secret64(40) ^ read_secret64(48)) ^ read64(input, input.len() - 8);
    let product = (low as u128).wrapping_mul(high as u128);
    let folded = (product as u64) ^ (product >> 64) as u64;
    avalanche(
        low.swap_bytes()
            .wrapping_add(high)
            .wrapping_add(folded)
            .wrapping_add(input.len() as u64),
    )
}

// xxhash3_64.rs
/// Compute XXH3-64 with seed 0 and the default secret.
#[must_use]
pub(crate) fn xxhash3_64(input: &[u8]) -> u64 {
    match input.len() {
        0 => avalanche_xxh64(read_secret64(56) ^ read_secret64(64)),
        1..=3 => xxh3_len_1_to_3(input),
        4..=8 => xxh3_len_4_to_8(input),
        9..=16 => xxh3_len_9_to_16(input),
        17..=128 => xxh3_len_17_to_128(input),
        129..=240 => xxh3_len_129_to_240(input),
        _ => xxh3_large(input),
    }
}

// xxhash64.rs
/// Compute XXH64 with seed 0.
#[must_use]
pub(crate) fn xxhash64(input: &[u8]) -> u64 {
    let mut index = 0usize;
    let mut acc = if input.len() >= 32 {
        let mut v1 = P64_1.wrapping_add(P64_2);
        let mut v2 = P64_2;
        let mut v3 = 0;
        let mut v4 = 0u64.wrapping_sub(P64_1);
        while index + 32 <= input.len() {
            v1 = round(v1, read64(input, index));
            v2 = round(v2, read64(input, index + 8));
            v3 = round(v3, read64(input, index + 16));
            v4 = round(v4, read64(input, index + 24));
            index += 32;
        }
        let mut merged = v1
            .rotate_left(1)
            .wrapping_add(v2.rotate_left(7))
            .wrapping_add(v3.rotate_left(12))
            .wrapping_add(v4.rotate_left(18));
        for v in [v1, v2, v3, v4] {
            merged = merge_acc(merged, v);
        }
        merged
    } else {
        P64_5
    };
    acc = acc.wrapping_add(input.len() as u64);
    consume_tail(input, index, acc)
}