| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | /* | ||
| 2 | * Copyright (c) 2026 Tiger Data, Inc. | ||
| 3 | * Licensed under the PostgreSQL License. See LICENSE for details. | ||
| 4 | * | ||
| 5 | * hkmeans.h - Hierarchical k-means tree builder | ||
| 6 | * | ||
| 7 | * Builds a BFS-ordered tree of centroids using hierarchical k-means. | ||
| 8 | * At each node, runs vs_kmeans_f32() to split vectors into fan_out | ||
| 9 | * children, then recurses on each partition. | ||
| 10 | * | ||
| 11 | * The result is self-contained: all data is stored in a single | ||
| 12 | * contiguous allocation using byte offsets instead of pointers. | ||
| 13 | * This allows memcpy into shared memory (DSM) for parallel builds. | ||
| 14 | * | ||
| 15 | * Used by both the PG IAM build (build.c) and the CLI | ||
| 16 | * benchmark (bench_search.c). | ||
| 17 | */ | ||
| 18 | |||
| 19 | #ifndef VS_HKMEANS_H | ||
| 20 | #define VS_HKMEANS_H | ||
| 21 | |||
| 22 | #include <stdint.h> | ||
| 23 | |||
| 24 | #include "algo/kmeans.h" | ||
| 25 | #include "core/types.h" | ||
| 26 | |||
| 27 | /* | ||
| 28 | * HKMeansNode - One node in the BFS-ordered tree | ||
| 29 | * | ||
| 30 | * centroid_offset is a byte offset from the HKMeansResult base, | ||
| 31 | * making the tree self-contained and memcpy-able. | ||
| 32 | */ | ||
| 33 | typedef struct HKMeansNode | ||
| 34 | { | ||
| 35 | uint32_t centroid_offset; /* byte offset from HKMeansResult* */ | ||
| 36 | uint32_t nchildren; /* actual cluster count (<= fan_out) */ | ||
| 37 | uint32_t level; /* tree level (0 = root) */ | ||
| 38 | uint32_t first_child; /* index in nodes[] of first child */ | ||
| 39 | uint32_t first_leaf; /* offset into leaf_centroids (leaf-parent only) */ | ||
| 40 | } HKMeansNode; | ||
| 41 | |||
| 42 | /* Sentinel for leaf nodes with no children */ | ||
| 43 | #define HKMEANS_NO_CHILD UINT32_MAX | ||
| 44 | |||
| 45 | /* | ||
| 46 | * HKMeansResult - Complete hierarchical k-means tree | ||
| 47 | * | ||
| 48 | * Self-contained: all data (nodes, leaf centroids, internal | ||
| 49 | * centroids) is stored in a single contiguous allocation. | ||
| 50 | * The struct can be memcpy'd into shared memory (DSM) and | ||
| 51 | * used directly by other processes without deserialization. | ||
| 52 | * | ||
| 53 | * Layout: | ||
| 54 | * [HKMeansResult header] | ||
| 55 | * [HKMeansNode nodes[nnodes]] — at nodes_offset | ||
| 56 | * [float leaf_centroids[nleaves*dim]] — at leaf_offset | ||
| 57 | * [float internal_centroids[...]] — packed after leaves | ||
| 58 | */ | ||
| 59 | typedef struct HKMeansResult | ||
| 60 | { | ||
| 61 | uint32_t nodes_offset; /* byte offset to nodes[] */ | ||
| 62 | uint32_t leaf_offset; /* byte offset to leaf centroids */ | ||
| 63 | uint32_t total_size; /* total allocation size in bytes */ | ||
| 64 | uint32_t nnodes; /* total internal nodes */ | ||
| 65 | uint32_t nlevels; /* tree depth */ | ||
| 66 | uint32_t nleaves; /* total leaf centroids */ | ||
| 67 | uint32_t fan_out; /* children per node (max) */ | ||
| 68 | Dimension dim; /* vector dimension */ | ||
| 69 | } HKMeansResult; | ||
| 70 | |||
| 71 | /* ---------------------------------------------------------------- | ||
| 72 | * Accessors — resolve byte offsets to typed pointers | ||
| 73 | * ---------------------------------------------------------------- */ | ||
| 74 | |||
| 75 | static inline HKMeansNode * | ||
| 76 | 13601 | hk_nodes(const HKMeansResult *r) | |
| 77 | { | ||
| 78 |
3/4✓ Branch 0 taken 20 times.
✓ Branch 1 taken 187 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 414 times.
|
13601 | return (HKMeansNode *)((char *)r + r->nodes_offset); |
| 79 | } | ||
| 80 | |||
| 81 | static inline float * | ||
| 82 | 2107 | hk_leaf_centroids(const HKMeansResult *r) | |
| 83 | { | ||
| 84 | 2107 | return (float *)((char *)r + r->leaf_offset); | |
| 85 | } | ||
| 86 | |||
| 87 | static inline float * | ||
| 88 | 23779 | hk_node_centroids(const HKMeansResult *r, const HKMeansNode *node) | |
| 89 | { | ||
| 90 | 23779 | return (float *)((char *)r + node->centroid_offset); | |
| 91 | } | ||
| 92 | |||
| 93 | /* ---------------------------------------------------------------- | ||
| 94 | * Public API | ||
| 95 | * ---------------------------------------------------------------- */ | ||
| 96 | |||
| 97 | /* | ||
| 98 | * Build a hierarchical k-means tree. | ||
| 99 | * | ||
| 100 | * indices: optional index array for indirect access (NULL = identity). | ||
| 101 | * When non-NULL, vector i is at vectors[indices[i] * dim]. | ||
| 102 | * This allows subsampling without copying. | ||
| 103 | * | ||
| 104 | * Returns a single contiguous allocation on success, NULL on failure. | ||
| 105 | * Caller must free with vs_free(). | ||
| 106 | */ | ||
| 107 | HKMeansResult *vs_hkmeans_f32( | ||
| 108 | const float *vectors, | ||
| 109 | uint32_t nvecs, | ||
| 110 | const uint32_t *indices, | ||
| 111 | Dimension dim, | ||
| 112 | uint32_t nlist, | ||
| 113 | uint32_t fan_out, | ||
| 114 | DistanceMetric metric, | ||
| 115 | const KMeansOptions *options); | ||
| 116 | |||
| 117 | /* | ||
| 118 | * Route a vector to its nearest leaf centroid by descending the tree. | ||
| 119 | * Returns the leaf index (0..nleaves-1). | ||
| 120 | * | ||
| 121 | * Optionally writes the distance to the nearest leaf centroid into | ||
| 122 | * *out_distance (may be NULL). | ||
| 123 | */ | ||
| 124 | uint32_t vs_hkmeans_assign( | ||
| 125 | const HKMeansResult *tree, | ||
| 126 | const float *vec, | ||
| 127 | DistanceMetric metric, | ||
| 128 | Distance *out_distance); | ||
| 129 | |||
| 130 | /* Upper bound on k / beam_width for vs_hkmeans_assign_topk (keeps the | ||
| 131 | * beam scratch on the stack). */ | ||
| 132 | #define VS_HK_MAX_TOPK 64 | ||
| 133 | |||
| 134 | /* | ||
| 135 | * Beam-search the tree for the k nearest leaf centroids. | ||
| 136 | * | ||
| 137 | * Maintains a beam of the best `beam_width` nodes per level, then keeps | ||
| 138 | * the k nearest leaves at the leaf level. Approximate for k/beam_width | ||
| 139 | * smaller than the tree fan-out, but far cheaper than scanning all | ||
| 140 | * leaves — used for secondary (boundary) cluster assignment during | ||
| 141 | * build. k and beam_width are clamped to VS_HK_MAX_TOPK. | ||
| 142 | * | ||
| 143 | * out_leaves[k] receives leaf indices sorted by ascending distance; | ||
| 144 | * out_dists[k] (optional) the matching distances. Returns the number of | ||
| 145 | * leaves written (<= k). | ||
| 146 | */ | ||
| 147 | uint32_t vs_hkmeans_assign_topk( | ||
| 148 | const HKMeansResult *tree, | ||
| 149 | const float *vec, | ||
| 150 | DistanceMetric metric, | ||
| 151 | uint32_t k, | ||
| 152 | uint32_t beam_width, | ||
| 153 | uint32_t *out_leaves, | ||
| 154 | Distance *out_dists); | ||
| 155 | |||
| 156 | /* | ||
| 157 | * Upper bound (bytes) on the contiguous size of a tree built for `nlist` | ||
| 158 | * leaves with `fan_out`. Used to size the fixed per-subtree DSM slots the | ||
| 159 | * parallel build's participants write their subtrees into. | ||
| 160 | */ | ||
| 161 | size_t | ||
| 162 | vs_hkmeans_max_blob_size(uint32_t nlist, uint32_t fan_out, Dimension dim); | ||
| 163 | |||
| 164 | /* | ||
| 165 | * Same bound with every level's width capped at max_leaves: a node exists | ||
| 166 | * only where a training vector landed, so a subtree clustered from | ||
| 167 | * max_leaves vectors can never exceed it, whatever the nlist target. | ||
| 168 | */ | ||
| 169 | size_t vs_hkmeans_max_blob_size_capped( | ||
| 170 | uint32_t nlist, uint32_t fan_out, Dimension dim, uint64_t max_leaves); | ||
| 171 | |||
| 172 | /* | ||
| 173 | * Tree depth for `nlist` leaves at `fan_out` — the same value the tree build | ||
| 174 | * uses internally. Exposed so the streaming (page-backed) centroid-tree build | ||
| 175 | * can compute the level structure without materializing a tree. | ||
| 176 | */ | ||
| 177 | uint32_t vs_hkmeans_nlevels(uint32_t nlist, uint32_t fan_out); | ||
| 178 | |||
| 179 | /* | ||
| 180 | * Build a one-level (flat) tree directly from pre-computed leaf centroids. | ||
| 181 | * | ||
| 182 | * For a flat clustering (nleaves <= fan_out) the root k-means already produced | ||
| 183 | * every leaf centroid, so the parallel build can assemble the tree straight | ||
| 184 | * from them rather than re-gathering the samples and re-clustering. Produces | ||
| 185 | * the same shape vs_hkmeans_f32 does for a single-level build: one | ||
| 186 | * leaf-parent root with nleaves children. Returns a contiguous allocation; | ||
| 187 | * caller frees with vs_free(). | ||
| 188 | */ | ||
| 189 | HKMeansResult *vs_hkmeans_build_flat( | ||
| 190 | const float *centroids, | ||
| 191 | uint32_t nleaves, | ||
| 192 | uint32_t fan_out, | ||
| 193 | Dimension dim); | ||
| 194 | |||
| 195 | #endif /* VS_HKMEANS_H */ | ||
| 196 |