eigen-db 4.1.0 → 4.3.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (16) hide show
  1. package/CHANGELOG.md +8 -0
  2. package/README.md +79 -27
  3. package/dist/eigen-db.js +317 -195
  4. package/dist/eigen-db.js.map +1 -1
  5. package/dist/eigen-db.umd.cjs +1 -1
  6. package/dist/eigen-db.umd.cjs.map +1 -1
  7. package/package.json +1 -1
  8. package/src/lib/__tests__/result-set.test.ts +19 -19
  9. package/src/lib/__tests__/vector-db.test.ts +429 -16
  10. package/src/lib/memory-manager.ts +8 -0
  11. package/src/lib/result-set.ts +16 -15
  12. package/src/lib/simd-binary.ts +1 -1
  13. package/src/lib/simd-optimized.wat +362 -0
  14. package/src/lib/simd.wat +42 -248
  15. package/src/lib/types.ts +4 -6
  16. package/src/lib/vector-db.ts +241 -9
package/src/lib/result-set.ts CHANGED
@@ -7,27 +7,28 @@
7
7
  * 2. iterableResults — returns a lazy Iterable<ResultItem> where keys are
8
8
  * resolved only as each item is consumed (for pagination / streaming)
9
9
  *
10
- * Distance is defined as `1 - dotProduct`. For normalized vectors (the default),
11
- * this equals cosine distance, ranging from 0 (identical) to 2 (opposite).
10
+ * Similarity is the dot product of query and stored vectors. For normalized
11
+ * vectors (the default), this equals cosine similarity, ranging from 1
12
+ * (identical) to -1 (opposite).
12
13
  */
13
14
 
14
15
  export interface ResultItem {
15
16
  key: string;
16
- distance: number;
17
+ similarity: number;
17
18
  }
18
19
 
19
20
  export type KeyResolver = (index: number) => string;
20
21
 
21
22
  /**
22
- * Sort by ascending distance and return the top K results as a plain array.
23
+ * Sort by descending similarity and return the top K results as a plain array.
23
24
  * All keys are resolved eagerly.
24
- * If maxDistance is provided, results with distance > maxDistance are excluded.
25
+ * If minSimilarity is provided, results with similarity < minSimilarity are excluded.
25
26
  */
26
27
  export function topKResults(
27
28
  scores: Float32Array,
28
29
  resolveKey: KeyResolver,
29
30
  topK: number,
30
- maxDistance?: number,
31
+ minSimilarity?: number,
31
32
  ): ResultItem[] {
32
33
  const n = scores.length;
33
34
  if (n === 0) return [];
@@ -40,19 +41,19 @@ export function topKResults(
40
41
  const results: ResultItem[] = [];
41
42
  for (let i = 0; i < k; i++) {
42
43
  const idx = indices[i];
43
- const distance = 1 - scores[idx];
44
- if (maxDistance !== undefined && distance > maxDistance) break;
45
- results.push({ key: resolveKey(idx), distance });
44
+ const similarity = scores[idx];
45
+ if (minSimilarity !== undefined && similarity < minSimilarity) break;
46
+ results.push({ key: resolveKey(idx), similarity });
46
47
  }
47
48
  return results;
48
49
  }
49
50
 
50
51
  /**
51
- * Sort by ascending distance and return a lazy iterable over the top K results.
52
+ * Sort by descending similarity and return a lazy iterable over the top K results.
52
53
  * Keys are resolved only when each item is consumed, saving allocations
53
54
  * when the caller iterates partially (e.g., pagination).
54
55
  *
55
- * If maxDistance is provided, iteration stops when distance > maxDistance.
56
+ * If minSimilarity is provided, iteration stops when similarity < minSimilarity.
56
57
  *
57
58
  * The returned iterable is re-iterable — each call to [Symbol.iterator]()
58
59
  * produces a fresh cursor over the same pre-sorted data.
@@ -61,7 +62,7 @@ export function iterableResults(
61
62
  scores: Float32Array,
62
63
  resolveKey: KeyResolver,
63
64
  topK: number,
64
- maxDistance?: number,
65
+ minSimilarity?: number,
65
66
  ): Iterable<ResultItem> {
66
67
  const n = scores.length;
67
68
  if (n === 0) return [];
@@ -79,11 +80,11 @@ export function iterableResults(
79
80
  next(): IteratorResult<ResultItem> {
80
81
  if (i >= k) return { done: true, value: undefined };
81
82
  const idx = indices[i++];
82
- const distance = 1 - scores[idx];
83
- if (maxDistance !== undefined && distance > maxDistance) return { done: true, value: undefined };
83
+ const similarity = scores[idx];
84
+ if (minSimilarity !== undefined && similarity < minSimilarity) return { done: true, value: undefined };
84
85
  return {
85
86
  done: false,
86
- value: { key: resolveKey(idx), distance },
87
+ value: { key: resolveKey(idx), similarity },
87
88
  };
88
89
  },
89
90
  };
package/src/lib/simd-binary.ts CHANGED
@@ -1,5 +1,5 @@
1
1
  // AUTO-GENERATED - Do not edit. Run: npx tsx scripts/compile-wat.ts
2
- const SIMD_WASM_BASE64 = "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";
2
+ const SIMD_WASM_BASE64 = "AGFzbQEAAAABDgJgAn9/AGAFf39/f38AAg8BA2VudgZtZW1vcnkCAAEDAwIAAQcaAglub3JtYWxpemUAAApzZWFyY2hfYWxsAAEKpgQCrQIFAX8BewN9AXsCf/0MAAAAAAAAAAAAAAAAAAAAACEDIAFBfHEhCEEAIQICQANAIAIgCE8NASAAIAJBAnRqIQkgAyAJ/QAEACAJ/QAEAP3mAf3kASEDIAJBBGohAgwACwsgA/0fACAD/R8BkiAD/R8CIAP9HwOSkiEEAkADQCACIAFPDQEgACACQQJ0aiEJIAQgCSoCACAJKgIAlJIhBCACQQFqIQIMAAsLIASRIQUgBUMAAAAAWwRADwtDAACAPyAFlSEGIAb9EyEHQQAhAgJAA0AgAiAITw0BIAAgAkECdGohCSAJIAn9AAQAIAf95gH9CwQAIAJBBGohAgwACwsCQANAIAIgAU8NASAAIAJBAnRqIQkgCSAJKgIAIAaUOAIAIAJBAWohAgwACwsL9AEEAn8BewF9BX8gBEF8cSEKIARBAnQhDUEAIQUCQANAIAUgA08NASABIAUgDWxqIQn9DAAAAAAAAAAAAAAAAAAAAAAhB0EAIQYCQANAIAYgCk8NASAAIAZBAnRqIQsgCSAGQQJ0aiEMIAcgC/0ABAAgDP0ABAD95gH95AEhByAGQQRqIQYMAAsLIAf9HwAgB/0fAZIgB/0fAiAH/R8DkpIhCAJAA0AgBiAETw0BIAAgBkECdGohCyAJIAZBAnRqIQwgCCALKgIAIAwqAgCUkiEIIAZBAWohBgwACwsgAiAFQQJ0aiAIOAIAIAVBAWohBQwACwsL";
3
3
 
4
4
  export function getSimdWasmBinary(): Uint8Array {
5
5
  const binaryString = atob(SIMD_WASM_BASE64);
package/src/lib/simd-optimized.wat ADDED
@@ -0,0 +1,362 @@
1
+ (module
2
+ ;; Import shared memory from JavaScript host
3
+ (import "env" "memory" (memory 1))
4
+
5
+ ;; normalize(ptr: i32, dimensions: i32)
6
+ ;; Normalizes a vector in-place to unit length using SIMD.
7
+ ;; Optimized with 4x loop unrolling (16 floats/iteration) and multiple accumulators.
8
+ (func (export "normalize") (param $ptr i32) (param $dim i32)
9
+ (local $i i32)
10
+ (local $acc0 v128)
11
+ (local $acc1 v128)
12
+ (local $acc2 v128)
13
+ (local $acc3 v128)
14
+ (local $sum f32)
15
+ (local $mag f32)
16
+ (local $inv_mag f32)
17
+ (local $inv_vec v128)
18
+ (local $unroll_end i32)
19
+ (local $simd_end i32)
20
+ (local $offset i32)
21
+
22
+ ;; Phase 1: Sum of squares with 4x unroll and 4 independent accumulators
23
+ (local.set $acc0 (v128.const f32x4 0 0 0 0))
24
+ (local.set $acc1 (v128.const f32x4 0 0 0 0))
25
+ (local.set $acc2 (v128.const f32x4 0 0 0 0))
26
+ (local.set $acc3 (v128.const f32x4 0 0 0 0))
27
+ (local.set $unroll_end (i32.and (local.get $dim) (i32.const -16)))
28
+ (local.set $simd_end (i32.and (local.get $dim) (i32.const -4)))
29
+ (local.set $i (i32.const 0))
30
+
31
+ (block $break_sum_u
32
+ (loop $loop_sum_u
33
+ (br_if $break_sum_u (i32.ge_u (local.get $i) (local.get $unroll_end)))
34
+ (local.set $offset (i32.add (local.get $ptr) (i32.shl (local.get $i) (i32.const 2))))
35
+
36
+ (local.set $acc0 (f32x4.add (local.get $acc0)
37
+ (f32x4.mul (v128.load (local.get $offset)) (v128.load (local.get $offset)))))
38
+ (local.set $acc1 (f32x4.add (local.get $acc1)
39
+ (f32x4.mul (v128.load offset=16 (local.get $offset)) (v128.load offset=16 (local.get $offset)))))
40
+ (local.set $acc2 (f32x4.add (local.get $acc2)
41
+ (f32x4.mul (v128.load offset=32 (local.get $offset)) (v128.load offset=32 (local.get $offset)))))
42
+ (local.set $acc3 (f32x4.add (local.get $acc3)
43
+ (f32x4.mul (v128.load offset=48 (local.get $offset)) (v128.load offset=48 (local.get $offset)))))
44
+
45
+ (local.set $i (i32.add (local.get $i) (i32.const 16)))
46
+ (br $loop_sum_u)
47
+ )
48
+ )
49
+
50
+ ;; Merge 4 accumulators
51
+ (local.set $acc0 (f32x4.add (f32x4.add (local.get $acc0) (local.get $acc1))
52
+ (f32x4.add (local.get $acc2) (local.get $acc3))))
53
+
54
+ ;; Remaining 4-wide chunks
55
+ (block $break_sum4
56
+ (loop $loop_sum4
57
+ (br_if $break_sum4 (i32.ge_u (local.get $i) (local.get $simd_end)))
58
+ (local.set $offset (i32.add (local.get $ptr) (i32.shl (local.get $i) (i32.const 2))))
59
+ (local.set $acc0 (f32x4.add (local.get $acc0)
60
+ (f32x4.mul (v128.load (local.get $offset)) (v128.load (local.get $offset)))))
61
+ (local.set $i (i32.add (local.get $i) (i32.const 4)))
62
+ (br $loop_sum4)
63
+ )
64
+ )
65
+
66
+ ;; Horizontal sum
67
+ (local.set $sum
68
+ (f32.add
69
+ (f32.add (f32x4.extract_lane 0 (local.get $acc0)) (f32x4.extract_lane 1 (local.get $acc0)))
70
+ (f32.add (f32x4.extract_lane 2 (local.get $acc0)) (f32x4.extract_lane 3 (local.get $acc0)))))
71
+
72
+ ;; Scalar remainder
73
+ (block $break_rem_sum
74
+ (loop $loop_rem_sum
75
+ (br_if $break_rem_sum (i32.ge_u (local.get $i) (local.get $dim)))
76
+ (local.set $offset (i32.add (local.get $ptr) (i32.shl (local.get $i) (i32.const 2))))
77
+ (local.set $sum (f32.add (local.get $sum)
78
+ (f32.mul (f32.load (local.get $offset)) (f32.load (local.get $offset)))))
79
+ (local.set $i (i32.add (local.get $i) (i32.const 1)))
80
+ (br $loop_rem_sum)
81
+ )
82
+ )
83
+
84
+ ;; Magnitude check
85
+ (local.set $mag (f32.sqrt (local.get $sum)))
86
+ (if (f32.eq (local.get $mag) (f32.const 0))
87
+ (then (return)))
88
+
89
+ ;; Phase 2: Scale by inverse magnitude (4x unrolled)
90
+ (local.set $inv_mag (f32.div (f32.const 1) (local.get $mag)))
91
+ (local.set $inv_vec (f32x4.splat (local.get $inv_mag)))
92
+ (local.set $i (i32.const 0))
93
+
94
+ (block $break_norm_u
95
+ (loop $loop_norm_u
96
+ (br_if $break_norm_u (i32.ge_u (local.get $i) (local.get $unroll_end)))
97
+ (local.set $offset (i32.add (local.get $ptr) (i32.shl (local.get $i) (i32.const 2))))
98
+
99
+ (v128.store (local.get $offset)
100
+ (f32x4.mul (v128.load (local.get $offset)) (local.get $inv_vec)))
101
+ (v128.store offset=16 (local.get $offset)
102
+ (f32x4.mul (v128.load offset=16 (local.get $offset)) (local.get $inv_vec)))
103
+ (v128.store offset=32 (local.get $offset)
104
+ (f32x4.mul (v128.load offset=32 (local.get $offset)) (local.get $inv_vec)))
105
+ (v128.store offset=48 (local.get $offset)
106
+ (f32x4.mul (v128.load offset=48 (local.get $offset)) (local.get $inv_vec)))
107
+
108
+ (local.set $i (i32.add (local.get $i) (i32.const 16)))
109
+ (br $loop_norm_u)
110
+ )
111
+ )
112
+
113
+ ;; Remaining 4-wide chunks
114
+ (block $break_norm4
115
+ (loop $loop_norm4
116
+ (br_if $break_norm4 (i32.ge_u (local.get $i) (local.get $simd_end)))
117
+ (local.set $offset (i32.add (local.get $ptr) (i32.shl (local.get $i) (i32.const 2))))
118
+ (v128.store (local.get $offset)
119
+ (f32x4.mul (v128.load (local.get $offset)) (local.get $inv_vec)))
120
+ (local.set $i (i32.add (local.get $i) (i32.const 4)))
121
+ (br $loop_norm4)
122
+ )
123
+ )
124
+
125
+ ;; Scalar remainder
126
+ (block $break_rem_norm
127
+ (loop $loop_rem_norm
128
+ (br_if $break_rem_norm (i32.ge_u (local.get $i) (local.get $dim)))
129
+ (local.set $offset (i32.add (local.get $ptr) (i32.shl (local.get $i) (i32.const 2))))
130
+ (f32.store (local.get $offset)
131
+ (f32.mul (f32.load (local.get $offset)) (local.get $inv_mag)))
132
+ (local.set $i (i32.add (local.get $i) (i32.const 1)))
133
+ (br $loop_rem_norm)
134
+ )
135
+ )
136
+ )
137
+
138
+ ;; search_all(query_ptr, db_ptr, scores_ptr, db_size, dimensions)
139
+ ;; Computes dot products of query against every vector in the database.
140
+ ;; Optimized with:
141
+ ;; - 2-vector outer loop unrolling (halves query memory reads)
142
+ ;; - 4x inner loop unrolling (16 floats/iteration, 4 accumulators per vector)
143
+ (func (export "search_all") (param $query_ptr i32) (param $db_ptr i32) (param $scores_ptr i32) (param $db_size i32) (param $dim i32)
144
+ (local $i i32)
145
+ (local $j i32)
146
+ (local $accA0 v128)
147
+ (local $accA1 v128)
148
+ (local $accA2 v128)
149
+ (local $accA3 v128)
150
+ (local $accB0 v128)
151
+ (local $accB1 v128)
152
+ (local $accB2 v128)
153
+ (local $accB3 v128)
154
+ (local $q0 v128)
155
+ (local $q1 v128)
156
+ (local $q2 v128)
157
+ (local $q3 v128)
158
+ (local $dotA f32)
159
+ (local $dotB f32)
160
+ (local $vec_ptrA i32)
161
+ (local $vec_ptrB i32)
162
+ (local $unroll_end i32)
163
+ (local $simd_end i32)
164
+ (local $q_offset i32)
165
+ (local $vA_offset i32)
166
+ (local $vB_offset i32)
167
+ (local $bytes_per_vec i32)
168
+ (local $pair_end i32)
169
+
170
+ (local.set $unroll_end (i32.and (local.get $dim) (i32.const -16)))
171
+ (local.set $simd_end (i32.and (local.get $dim) (i32.const -4)))
172
+ (local.set $bytes_per_vec (i32.shl (local.get $dim) (i32.const 2)))
173
+ (local.set $pair_end (i32.and (local.get $db_size) (i32.const -2)))
174
+ (local.set $i (i32.const 0))
175
+
176
+ ;; Main loop: process 2 database vectors per iteration
177
+ (block $break_outer
178
+ (loop $loop_outer
179
+ (br_if $break_outer (i32.ge_u (local.get $i) (local.get $pair_end)))
180
+
181
+ (local.set $vec_ptrA
182
+ (i32.add (local.get $db_ptr) (i32.mul (local.get $i) (local.get $bytes_per_vec))))
183
+ (local.set $vec_ptrB
184
+ (i32.add (local.get $vec_ptrA) (local.get $bytes_per_vec)))
185
+
186
+ (local.set $accA0 (v128.const f32x4 0 0 0 0))
187
+ (local.set $accA1 (v128.const f32x4 0 0 0 0))
188
+ (local.set $accA2 (v128.const f32x4 0 0 0 0))
189
+ (local.set $accA3 (v128.const f32x4 0 0 0 0))
190
+ (local.set $accB0 (v128.const f32x4 0 0 0 0))
191
+ (local.set $accB1 (v128.const f32x4 0 0 0 0))
192
+ (local.set $accB2 (v128.const f32x4 0 0 0 0))
193
+ (local.set $accB3 (v128.const f32x4 0 0 0 0))
194
+ (local.set $j (i32.const 0))
195
+
196
+ ;; Inner loop: load query once, dot product against both vectors
197
+ (block $break_inner
198
+ (loop $loop_inner
199
+ (br_if $break_inner (i32.ge_u (local.get $j) (local.get $unroll_end)))
200
+ (local.set $q_offset (i32.add (local.get $query_ptr) (i32.shl (local.get $j) (i32.const 2))))
201
+ (local.set $vA_offset (i32.add (local.get $vec_ptrA) (i32.shl (local.get $j) (i32.const 2))))
202
+ (local.set $vB_offset (i32.add (local.get $vec_ptrB) (i32.shl (local.get $j) (i32.const 2))))
203
+
204
+ ;; Load query chunk (shared between both vectors)
205
+ (local.set $q0 (v128.load (local.get $q_offset)))
206
+ (local.set $q1 (v128.load offset=16 (local.get $q_offset)))
207
+ (local.set $q2 (v128.load offset=32 (local.get $q_offset)))
208
+ (local.set $q3 (v128.load offset=48 (local.get $q_offset)))
209
+
210
+ ;; Vector A
211
+ (local.set $accA0 (f32x4.add (local.get $accA0)
212
+ (f32x4.mul (local.get $q0) (v128.load (local.get $vA_offset)))))
213
+ (local.set $accA1 (f32x4.add (local.get $accA1)
214
+ (f32x4.mul (local.get $q1) (v128.load offset=16 (local.get $vA_offset)))))
215
+ (local.set $accA2 (f32x4.add (local.get $accA2)
216
+ (f32x4.mul (local.get $q2) (v128.load offset=32 (local.get $vA_offset)))))
217
+ (local.set $accA3 (f32x4.add (local.get $accA3)
218
+ (f32x4.mul (local.get $q3) (v128.load offset=48 (local.get $vA_offset)))))
219
+
220
+ ;; Vector B (reuses query loads)
221
+ (local.set $accB0 (f32x4.add (local.get $accB0)
222
+ (f32x4.mul (local.get $q0) (v128.load (local.get $vB_offset)))))
223
+ (local.set $accB1 (f32x4.add (local.get $accB1)
224
+ (f32x4.mul (local.get $q1) (v128.load offset=16 (local.get $vB_offset)))))
225
+ (local.set $accB2 (f32x4.add (local.get $accB2)
226
+ (f32x4.mul (local.get $q2) (v128.load offset=32 (local.get $vB_offset)))))
227
+ (local.set $accB3 (f32x4.add (local.get $accB3)
228
+ (f32x4.mul (local.get $q3) (v128.load offset=48 (local.get $vB_offset)))))
229
+
230
+ (local.set $j (i32.add (local.get $j) (i32.const 16)))
231
+ (br $loop_inner)
232
+ )
233
+ )
234
+
235
+ ;; Merge accumulators
236
+ (local.set $accA0 (f32x4.add (f32x4.add (local.get $accA0) (local.get $accA1))
237
+ (f32x4.add (local.get $accA2) (local.get $accA3))))
238
+ (local.set $accB0 (f32x4.add (f32x4.add (local.get $accB0) (local.get $accB1))
239
+ (f32x4.add (local.get $accB2) (local.get $accB3))))
240
+
241
+ ;; 4-wide cleanup (both vectors)
242
+ (block $break_inner4
243
+ (loop $loop_inner4
244
+ (br_if $break_inner4 (i32.ge_u (local.get $j) (local.get $simd_end)))
245
+ (local.set $q_offset (i32.add (local.get $query_ptr) (i32.shl (local.get $j) (i32.const 2))))
246
+ (local.set $vA_offset (i32.add (local.get $vec_ptrA) (i32.shl (local.get $j) (i32.const 2))))
247
+ (local.set $vB_offset (i32.add (local.get $vec_ptrB) (i32.shl (local.get $j) (i32.const 2))))
248
+ (local.set $q0 (v128.load (local.get $q_offset)))
249
+ (local.set $accA0 (f32x4.add (local.get $accA0) (f32x4.mul (local.get $q0) (v128.load (local.get $vA_offset)))))
250
+ (local.set $accB0 (f32x4.add (local.get $accB0) (f32x4.mul (local.get $q0) (v128.load (local.get $vB_offset)))))
251
+ (local.set $j (i32.add (local.get $j) (i32.const 4)))
252
+ (br $loop_inner4)
253
+ )
254
+ )
255
+
256
+ ;; Horizontal sums
257
+ (local.set $dotA
258
+ (f32.add
259
+ (f32.add (f32x4.extract_lane 0 (local.get $accA0)) (f32x4.extract_lane 1 (local.get $accA0)))
260
+ (f32.add (f32x4.extract_lane 2 (local.get $accA0)) (f32x4.extract_lane 3 (local.get $accA0)))))
261
+ (local.set $dotB
262
+ (f32.add
263
+ (f32.add (f32x4.extract_lane 0 (local.get $accB0)) (f32x4.extract_lane 1 (local.get $accB0)))
264
+ (f32.add (f32x4.extract_lane 2 (local.get $accB0)) (f32x4.extract_lane 3 (local.get $accB0)))))
265
+
266
+ ;; Scalar remainder (both vectors)
267
+ (block $break_rem
268
+ (loop $loop_rem
269
+ (br_if $break_rem (i32.ge_u (local.get $j) (local.get $dim)))
270
+ (local.set $q_offset (i32.add (local.get $query_ptr) (i32.shl (local.get $j) (i32.const 2))))
271
+ (local.set $vA_offset (i32.add (local.get $vec_ptrA) (i32.shl (local.get $j) (i32.const 2))))
272
+ (local.set $vB_offset (i32.add (local.get $vec_ptrB) (i32.shl (local.get $j) (i32.const 2))))
273
+ (local.set $dotA (f32.add (local.get $dotA)
274
+ (f32.mul (f32.load (local.get $q_offset)) (f32.load (local.get $vA_offset)))))
275
+ (local.set $dotB (f32.add (local.get $dotB)
276
+ (f32.mul (f32.load (local.get $q_offset)) (f32.load (local.get $vB_offset)))))
277
+ (local.set $j (i32.add (local.get $j) (i32.const 1)))
278
+ (br $loop_rem)
279
+ )
280
+ )
281
+
282
+ ;; Store scores for vectors i and i+1
283
+ (f32.store
284
+ (i32.add (local.get $scores_ptr) (i32.shl (local.get $i) (i32.const 2)))
285
+ (local.get $dotA))
286
+ (f32.store
287
+ (i32.add (local.get $scores_ptr) (i32.shl (i32.add (local.get $i) (i32.const 1)) (i32.const 2)))
288
+ (local.get $dotB))
289
+
290
+ (local.set $i (i32.add (local.get $i) (i32.const 2)))
291
+ (br $loop_outer)
292
+ )
293
+ )
294
+
295
+ ;; Handle last vector if db_size is odd
296
+ (if (i32.lt_u (local.get $i) (local.get $db_size))
297
+ (then
298
+ (local.set $vec_ptrA
299
+ (i32.add (local.get $db_ptr) (i32.mul (local.get $i) (local.get $bytes_per_vec))))
300
+ (local.set $accA0 (v128.const f32x4 0 0 0 0))
301
+ (local.set $accA1 (v128.const f32x4 0 0 0 0))
302
+ (local.set $accA2 (v128.const f32x4 0 0 0 0))
303
+ (local.set $accA3 (v128.const f32x4 0 0 0 0))
304
+ (local.set $j (i32.const 0))
305
+
306
+ (block $break_last
307
+ (loop $loop_last
308
+ (br_if $break_last (i32.ge_u (local.get $j) (local.get $unroll_end)))
309
+ (local.set $q_offset (i32.add (local.get $query_ptr) (i32.shl (local.get $j) (i32.const 2))))
310
+ (local.set $vA_offset (i32.add (local.get $vec_ptrA) (i32.shl (local.get $j) (i32.const 2))))
311
+ (local.set $accA0 (f32x4.add (local.get $accA0)
312
+ (f32x4.mul (v128.load (local.get $q_offset)) (v128.load (local.get $vA_offset)))))
313
+ (local.set $accA1 (f32x4.add (local.get $accA1)
314
+ (f32x4.mul (v128.load offset=16 (local.get $q_offset)) (v128.load offset=16 (local.get $vA_offset)))))
315
+ (local.set $accA2 (f32x4.add (local.get $accA2)
316
+ (f32x4.mul (v128.load offset=32 (local.get $q_offset)) (v128.load offset=32 (local.get $vA_offset)))))
317
+ (local.set $accA3 (f32x4.add (local.get $accA3)
318
+ (f32x4.mul (v128.load offset=48 (local.get $q_offset)) (v128.load offset=48 (local.get $vA_offset)))))
319
+ (local.set $j (i32.add (local.get $j) (i32.const 16)))
320
+ (br $loop_last)
321
+ )
322
+ )
323
+
324
+ (local.set $accA0 (f32x4.add (f32x4.add (local.get $accA0) (local.get $accA1))
325
+ (f32x4.add (local.get $accA2) (local.get $accA3))))
326
+
327
+ (block $break_last4
328
+ (loop $loop_last4
329
+ (br_if $break_last4 (i32.ge_u (local.get $j) (local.get $simd_end)))
330
+ (local.set $q_offset (i32.add (local.get $query_ptr) (i32.shl (local.get $j) (i32.const 2))))
331
+ (local.set $vA_offset (i32.add (local.get $vec_ptrA) (i32.shl (local.get $j) (i32.const 2))))
332
+ (local.set $accA0 (f32x4.add (local.get $accA0)
333
+ (f32x4.mul (v128.load (local.get $q_offset)) (v128.load (local.get $vA_offset)))))
334
+ (local.set $j (i32.add (local.get $j) (i32.const 4)))
335
+ (br $loop_last4)
336
+ )
337
+ )
338
+
339
+ (local.set $dotA
340
+ (f32.add
341
+ (f32.add (f32x4.extract_lane 0 (local.get $accA0)) (f32x4.extract_lane 1 (local.get $accA0)))
342
+ (f32.add (f32x4.extract_lane 2 (local.get $accA0)) (f32x4.extract_lane 3 (local.get $accA0)))))
343
+
344
+ (block $break_last_rem
345
+ (loop $loop_last_rem
346
+ (br_if $break_last_rem (i32.ge_u (local.get $j) (local.get $dim)))
347
+ (local.set $q_offset (i32.add (local.get $query_ptr) (i32.shl (local.get $j) (i32.const 2))))
348
+ (local.set $vA_offset (i32.add (local.get $vec_ptrA) (i32.shl (local.get $j) (i32.const 2))))
349
+ (local.set $dotA (f32.add (local.get $dotA)
350
+ (f32.mul (f32.load (local.get $q_offset)) (f32.load (local.get $vA_offset)))))
351
+ (local.set $j (i32.add (local.get $j) (i32.const 1)))
352
+ (br $loop_last_rem)
353
+ )
354
+ )
355
+
356
+ (f32.store
357
+ (i32.add (local.get $scores_ptr) (i32.shl (local.get $i) (i32.const 2)))
358
+ (local.get $dotA))
359
+ )
360
+ )
361
+ )
362
+ )