edgevec 0.9.0

High-performance embedded vector database for Browser, Node, and Edge
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
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353

//! SIMD-accelerated popcount for Hamming distance computation.
//!
//! This module provides hardware-accelerated XOR + popcount operations
//! for computing Hamming distance between binary vectors of arbitrary length.
//!
//! # Supported Platforms
//!
//! - **x86_64 with AVX2**: Uses `_mm256_xor_si256` + native popcnt extraction
//! - **x86_64 with popcnt**: Uses native `popcnt` instruction on u64 chunks
//! - **aarch64 with NEON**: Uses `veorq_u8` + `vcntq_u8` for parallel popcount
//! - **All platforms**: Scalar fallback using `count_ones()`
//!
//! # Performance
//!
//! | Platform | Expected Speedup vs Scalar |
//! |----------|---------------------------|
//! | AVX2     | 3-5x                      |
//! | popcnt   | 2-3x                      |
//! | NEON     | 2-3x                      |
//! | Scalar   | 1x (baseline)             |
//!
//! # Example
//!
//! ```
//! use edgevec::simd::popcount::{simd_popcount_xor, scalar_popcount_xor};
//!
//! let a = vec![0xAA; 96]; // 10101010...
//! let b = vec![0x55; 96]; // 01010101...
//!
//! let distance = simd_popcount_xor(&a, &b);
//! assert_eq!(distance, 768); // All bits differ
//!
//! // Scalar version for comparison
//! let scalar_distance = scalar_popcount_xor(&a, &b);
//! assert_eq!(distance, scalar_distance);
//! ```

/// Computes popcount of XOR between two byte slices.
///
/// Uses the fastest available implementation:
/// - AVX2 (x86_64 with AVX2): 32-byte parallel processing
/// - popcnt (x86_64 with popcnt): Native instruction per u64
/// - NEON (aarch64): 16-byte parallel processing
/// - Scalar: Portable fallback using `count_ones()`
///
/// # Arguments
///
/// * `a` - First byte slice.
/// * `b` - Second byte slice (must have same length as `a`).
///
/// # Returns
///
/// The total number of bits that differ between `a` and `b`.
///
/// # Panics
///
/// Panics in debug mode if `a.len() != b.len()`.
///
/// # Example
///
/// ```
/// use edgevec::simd::popcount::simd_popcount_xor;
///
/// let a = [0xFF; 16]; // All ones
/// let b = [0x00; 16]; // All zeros
///
/// let distance = simd_popcount_xor(&a, &b);
/// assert_eq!(distance, 128); // 16 * 8 = 128 bits
/// ```
#[inline]
#[must_use]
pub fn simd_popcount_xor(a: &[u8], b: &[u8]) -> u32 {
    debug_assert_eq!(a.len(), b.len(), "slices must have equal length");

    #[cfg(target_arch = "x86_64")]
    {
        if is_x86_feature_detected!("avx2") {
            // SAFETY: We verified AVX2 is available
            return unsafe { avx2_popcount_xor(a, b) };
        }
        if is_x86_feature_detected!("popcnt") {
            return native_popcount_xor(a, b);
        }
    }

    #[cfg(target_arch = "aarch64")]
    {
        if std::arch::is_aarch64_feature_detected!("neon") {
            // SAFETY: We verified NEON is available
            return unsafe { neon_popcount_xor(a, b) };
        }
    }

    scalar_popcount_xor(a, b)
}

/// Scalar (non-SIMD) popcount of XOR.
///
/// This is the baseline implementation used when SIMD is not available.
/// Also useful for testing to verify SIMD correctness.
///
/// # Arguments
///
/// * `a` - First byte slice.
/// * `b` - Second byte slice (must have same length as `a`).
///
/// # Returns
///
/// The total number of bits that differ between `a` and `b`.
#[inline]
#[must_use]
pub fn scalar_popcount_xor(a: &[u8], b: &[u8]) -> u32 {
    debug_assert_eq!(a.len(), b.len());
    a.iter()
        .zip(b.iter())
        .map(|(x, y)| (x ^ y).count_ones())
        .sum()
}

/// x86_64 AVX2 implementation using native popcnt instruction.
///
/// Processes 32 bytes at a time using AVX2 XOR, then extracts 4×u64 values
/// and uses native popcnt via count_ones(). This is faster than the PSHUFB
/// lookup table method on modern CPUs (Haswell+) that have hardware popcnt.
#[cfg(target_arch = "x86_64")]
#[target_feature(enable = "avx2")]
#[allow(
    clippy::cast_possible_truncation,
    clippy::cast_sign_loss,
    clippy::cast_ptr_alignment
)]
unsafe fn avx2_popcount_xor(a: &[u8], b: &[u8]) -> u32 {
    use std::arch::x86_64::{__m256i, _mm256_extract_epi64, _mm256_loadu_si256, _mm256_xor_si256};

    let mut total = 0u32;
    let len = a.len();

    // Process 32 bytes at a time
    let chunks = len / 32;
    for i in 0..chunks {
        let offset = i * 32;
        let va = _mm256_loadu_si256(a.as_ptr().add(offset).cast::<__m256i>());
        let vb = _mm256_loadu_si256(b.as_ptr().add(offset).cast::<__m256i>());

        // XOR the vectors
        let xor = _mm256_xor_si256(va, vb);

        // Extract 4×u64 and use native popcnt instruction.
        // count_ones() compiles to popcnt on x86_64 with hardware support.
        let v0 = _mm256_extract_epi64(xor, 0) as u64;
        let v1 = _mm256_extract_epi64(xor, 1) as u64;
        let v2 = _mm256_extract_epi64(xor, 2) as u64;
        let v3 = _mm256_extract_epi64(xor, 3) as u64;

        total += v0.count_ones() + v1.count_ones() + v2.count_ones() + v3.count_ones();
    }

    // Handle remainder with scalar
    let remainder_start = chunks * 32;
    for i in remainder_start..len {
        total += (a[i] ^ b[i]).count_ones();
    }

    total
}

/// x86_64 native popcnt implementation.
///
/// Uses the native `popcnt` instruction on u64 chunks.
///
/// # Complexity
///
/// O(n/8) where n is the byte length of the input slices.
#[cfg(target_arch = "x86_64")]
fn native_popcount_xor(a: &[u8], b: &[u8]) -> u32 {
    let mut count = 0u32;

    // Process 8 bytes at a time using u64.
    // chunks_exact guarantees each chunk is exactly 8 bytes.
    for (chunk_a, chunk_b) in a.chunks_exact(8).zip(b.chunks_exact(8)) {
        // SAFETY: chunks_exact(8) guarantees exactly 8 bytes.
        // Using array indexing is panic-free since chunk length is guaranteed.
        let va = u64::from_le_bytes([
            chunk_a[0], chunk_a[1], chunk_a[2], chunk_a[3], chunk_a[4], chunk_a[5], chunk_a[6],
            chunk_a[7],
        ]);
        let vb = u64::from_le_bytes([
            chunk_b[0], chunk_b[1], chunk_b[2], chunk_b[3], chunk_b[4], chunk_b[5], chunk_b[6],
            chunk_b[7],
        ]);
        count += (va ^ vb).count_ones();
    }

    // Handle remainder (0-7 bytes)
    let remainder_start = (a.len() / 8) * 8;
    for i in remainder_start..a.len() {
        count += (a[i] ^ b[i]).count_ones();
    }

    count
}

/// ARM64 NEON implementation.
///
/// Uses `veorq_u8` for XOR and `vcntq_u8` for parallel popcount.
#[cfg(target_arch = "aarch64")]
#[target_feature(enable = "neon")]
unsafe fn neon_popcount_xor(a: &[u8], b: &[u8]) -> u32 {
    use std::arch::aarch64::*;

    let mut total = 0u32;
    let len = a.len();

    // Process 16 bytes at a time
    let chunks = len / 16;
    for i in 0..chunks {
        let offset = i * 16;
        let va = vld1q_u8(a.as_ptr().add(offset));
        let vb = vld1q_u8(b.as_ptr().add(offset));

        // XOR and popcount
        let xor = veorq_u8(va, vb);
        let cnt = vcntq_u8(xor);

        // Horizontal sum: u8x16 -> u16x8 -> u32x4 -> u64x2
        let sum16 = vpaddlq_u8(cnt);
        let sum32 = vpaddlq_u16(sum16);
        let sum64 = vpaddlq_u32(sum32);

        // Extract and accumulate
        total += (vgetq_lane_u64(sum64, 0) + vgetq_lane_u64(sum64, 1)) as u32;
    }

    // Handle remainder with scalar
    let remainder_start = chunks * 16;
    for i in remainder_start..len {
        total += (a[i] ^ b[i]).count_ones();
    }

    total
}

#[cfg(test)]
#[allow(clippy::cast_possible_truncation, clippy::cast_sign_loss)]
mod tests {
    use super::*;

    #[test]
    fn test_scalar_popcount_identical() {
        let a = vec![0xAA; 96];
        let b = vec![0xAA; 96];
        assert_eq!(scalar_popcount_xor(&a, &b), 0);
    }

    #[test]
    fn test_scalar_popcount_opposite() {
        let a = vec![0x00; 96];
        let b = vec![0xFF; 96];
        assert_eq!(scalar_popcount_xor(&a, &b), 768);
    }

    #[test]
    fn test_scalar_popcount_half() {
        let a = vec![0xF0; 96]; // 11110000
        let b = vec![0x0F; 96]; // 00001111
        assert_eq!(scalar_popcount_xor(&a, &b), 768); // All bits differ
    }

    #[test]
    fn test_simd_matches_scalar_96_bytes() {
        let a: Vec<u8> = (0..96).map(|i| i as u8).collect();
        let b: Vec<u8> = (0..96).map(|i| (i * 2) as u8).collect();

        let simd_result = simd_popcount_xor(&a, &b);
        let scalar_result = scalar_popcount_xor(&a, &b);

        assert_eq!(simd_result, scalar_result);
    }

    #[test]
    fn test_simd_matches_scalar_16_bytes() {
        let a = vec![0xAA; 16];
        let b = vec![0x55; 16];

        let simd_result = simd_popcount_xor(&a, &b);
        let scalar_result = scalar_popcount_xor(&a, &b);

        assert_eq!(simd_result, scalar_result);
        assert_eq!(simd_result, 128); // All bits differ
    }

    #[test]
    fn test_simd_matches_scalar_128_bytes() {
        let a: Vec<u8> = (0..128).map(|i| (i * 3) as u8).collect();
        let b: Vec<u8> = (0..128).map(|i| (i * 7) as u8).collect();

        let simd_result = simd_popcount_xor(&a, &b);
        let scalar_result = scalar_popcount_xor(&a, &b);

        assert_eq!(simd_result, scalar_result);
    }

    #[test]
    fn test_simd_matches_scalar_192_bytes() {
        // 1536D vectors = 192 bytes
        let a: Vec<u8> = (0..192).map(|i| (i * 5) as u8).collect();
        let b: Vec<u8> = (0..192).map(|i| (i * 11) as u8).collect();

        let simd_result = simd_popcount_xor(&a, &b);
        let scalar_result = scalar_popcount_xor(&a, &b);

        assert_eq!(simd_result, scalar_result);
    }

    #[test]
    fn test_simd_odd_length() {
        // Test with non-power-of-2 length (7 bytes)
        let a = vec![0xFF; 7];
        let b = vec![0x00; 7];

        let simd_result = simd_popcount_xor(&a, &b);
        let scalar_result = scalar_popcount_xor(&a, &b);

        assert_eq!(simd_result, scalar_result);
        assert_eq!(simd_result, 56); // 7 * 8 = 56 bits
    }

    #[test]
    fn test_simd_empty_slices() {
        let a: Vec<u8> = vec![];
        let b: Vec<u8> = vec![];
        assert_eq!(simd_popcount_xor(&a, &b), 0);
    }

    #[test]
    fn test_simd_single_byte() {
        let a = vec![0b1010_1010];
        let b = vec![0b0101_0101];
        assert_eq!(simd_popcount_xor(&a, &b), 8);
    }

    #[test]
    fn test_simd_large_vectors() {
        // Test with 4096 bytes (very large)
        let a: Vec<u8> = (0..4096).map(|i| (i % 256) as u8).collect();
        let b: Vec<u8> = (0..4096).map(|i| ((i + 128) % 256) as u8).collect();

        let simd_result = simd_popcount_xor(&a, &b);
        let scalar_result = scalar_popcount_xor(&a, &b);

        assert_eq!(simd_result, scalar_result);
    }
}