Some small improvements.

Signed-off-by: Alexandr Guzhva <alexanderguzhva@gmail.com>

Author: alexanderguzhva

faiss/IndexHNSW.cpp

@@ -275,23 +275,23 @@ void hnsw_search(
         FAISS_THROW_IF_NOT_MSG(params, "params type invalid");
         efSearch = params->efSearch;
     }
-    size_t n1 = 0, n2 = 0, ndis = 0;
+    size_t n1 = 0, n2 = 0, ndis = 0, nhops = 0;
 
     idx_t check_period = InterruptCallback::get_period_hint(
             hnsw.max_level * index->d * efSearch);
 
     for (idx_t i0 = 0; i0 < n; i0 += check_period) {
         idx_t i1 = std::min(i0 + check_period, n);
 
-#pragma omp parallel
+#pragma omp parallel if (i1 - i0 > 1)
         {
             VisitedTable vt(index->ntotal);
             typename BlockResultHandler::SingleResultHandler res(bres);
 
             std::unique_ptr<DistanceComputer> dis(
                     storage_distance_computer(index->storage));
 
-#pragma omp for reduction(+ : n1, n2, ndis) schedule(guided)
+#pragma omp for reduction(+ : n1, n2, ndis, nhops) schedule(guided)
             for (idx_t i = i0; i < i1; i++) {
                 res.begin(i);
                 dis->set_query(x + i * index->d);
@@ -300,13 +300,14 @@ void hnsw_search(
                 n1 += stats.n1;
                 n2 += stats.n2;
                 ndis += stats.ndis;
+                nhops += stats.nhops;
                 res.end();
             }
         }
         InterruptCallback::check();
     }
 
-    hnsw_stats.combine({n1, n2, ndis});
+    hnsw_stats.combine({n1, n2, ndis, nhops});
 }
 
 } // anonymous namespace
@@ -632,6 +633,10 @@ void IndexHNSW::permute_entries(const idx_t* perm) {
     hnsw.permute_entries(perm);
 }
 
+DistanceComputer* IndexHNSW::get_distance_computer() const {
+    return storage->get_distance_computer();
+}
+
 /**************************************************************
  * IndexHNSWFlat implementation
  **************************************************************/
@@ -655,8 +660,13 @@ IndexHNSWFlat::IndexHNSWFlat(int d, int M, MetricType metric)
 
 IndexHNSWPQ::IndexHNSWPQ() = default;
 
-IndexHNSWPQ::IndexHNSWPQ(int d, int pq_m, int M, int pq_nbits)
-        : IndexHNSW(new IndexPQ(d, pq_m, pq_nbits), M) {
+IndexHNSWPQ::IndexHNSWPQ(
+        int d,
+        int pq_m,
+        int M,
+        int pq_nbits,
+        MetricType metric)
+        : IndexHNSW(new IndexPQ(d, pq_m, pq_nbits, metric), M) {
     own_fields = true;
     is_trained = false;
 }
@@ -782,7 +792,7 @@ void IndexHNSW2Level::search(
         IndexHNSW::search(n, x, k, distances, labels);
 
     } else { // "mixed" search
-        size_t n1 = 0, n2 = 0, ndis = 0;
+        size_t n1 = 0, n2 = 0, ndis = 0, nhops = 0;
 
         const IndexIVFPQ* index_ivfpq =
                 dynamic_cast<const IndexIVFPQ*>(storage);
@@ -814,7 +824,7 @@ void IndexHNSW2Level::search(
             int candidates_size = hnsw.upper_beam;
             MinimaxHeap candidates(candidates_size);
 
-#pragma omp for reduction(+ : n1, n2, ndis)
+#pragma omp for reduction(+ : n1, n2, ndis, nhops)
             for (idx_t i = 0; i < n; i++) {
                 idx_t* idxi = labels + i * k;
                 float* simi = distances + i * k;
@@ -860,6 +870,7 @@ void IndexHNSW2Level::search(
                 n1 += search_stats.n1;
                 n2 += search_stats.n2;
                 ndis += search_stats.ndis;
+                nhops += search_stats.nhops;
 
                 vt.advance();
                 vt.advance();
@@ -868,7 +879,7 @@ void IndexHNSW2Level::search(
             }
         }
 
-        hnsw_stats.combine({n1, n2, ndis});
+        hnsw_stats.combine({n1, n2, ndis, nhops});
     }
 }
 

faiss/IndexHNSW.h

@@ -27,7 +27,7 @@ struct IndexHNSW;
 struct IndexHNSW : Index {
     typedef HNSW::storage_idx_t storage_idx_t;
 
-    // the link strcuture
+    // the link structure
     HNSW hnsw;
 
     // the sequential storage
@@ -111,6 +111,8 @@ struct IndexHNSW : Index {
     void link_singletons();
 
     void permute_entries(const idx_t* perm);
+
+    DistanceComputer* get_distance_computer() const override;
 };
 
 /** Flat index topped with with a HNSW structure to access elements
@@ -127,7 +129,12 @@ struct IndexHNSWFlat : IndexHNSW {
  */
 struct IndexHNSWPQ : IndexHNSW {
     IndexHNSWPQ();
-    IndexHNSWPQ(int d, int pq_m, int M, int pq_nbits = 8);
+    IndexHNSWPQ(
+            int d,
+            int pq_m,
+            int M,
+            int pq_nbits = 8,
+            MetricType metric = METRIC_L2);
     void train(idx_t n, const float* x) override;
 };
 

faiss/IndexRefine.cpp

@@ -68,12 +68,12 @@ template <class C>
 static void reorder_2_heaps(
         idx_t n,
         idx_t k,
-        idx_t* labels,
-        float* distances,
+        idx_t* __restrict labels,
+        float* __restrict distances,
         idx_t k_base,
-        const idx_t* base_labels,
-        const float* base_distances) {
-#pragma omp parallel for
+        const idx_t* __restrict base_labels,
+        const float* __restrict base_distances) {
+#pragma omp parallel for if (n > 1)
     for (idx_t i = 0; i < n; i++) {
         idx_t* idxo = labels + i * k;
         float* diso = distances + i * k;

faiss/impl/HNSW.cpp

@@ -409,18 +409,22 @@ void search_neighbors_to_add(
  **************************************************************/
 
 /// greedily update a nearest vector at a given level
-void greedy_update_nearest(
+HNSWStats greedy_update_nearest(
         const HNSW& hnsw,
         DistanceComputer& qdis,
         int level,
         storage_idx_t& nearest,
         float& d_nearest) {
+    HNSWStats stats;
+
     for (;;) {
         storage_idx_t prev_nearest = nearest;
 
         size_t begin, end;
         hnsw.neighbor_range(nearest, level, &begin, &end);
-        for (size_t i = begin; i < end; i++) {
+
+        size_t ndis = 0;
+        for (size_t i = begin; i < end; i++, ndis++) {
             storage_idx_t v = hnsw.neighbors[i];
             if (v < 0)
                 break;
@@ -430,8 +434,13 @@ void greedy_update_nearest(
                 d_nearest = dis;
             }
         }
+
+        // update stats
+        stats.ndis += ndis;
+        stats.nhops += 1;
+
         if (nearest == prev_nearest) {
-            return;
+            return stats;
         }
     }
 }
@@ -641,6 +650,7 @@ int search_from_candidates(
                 if (dis < threshold) {
                     if (res.add_result(dis, idx)) {
                         threshold = res.threshold;
+                        nres += 1;
                     }
                 }
             }
@@ -692,6 +702,7 @@ int search_from_candidates(
             stats.n2++;
         }
         stats.ndis += ndis;
+        stats.nhops += nstep;
     }
 
     return nres;
@@ -814,6 +825,8 @@ std::priority_queue<HNSW::Node> search_from_candidate_unbounded(
             float dis = qdis(saved_j[icnt]);
             add_to_heap(saved_j[icnt], dis);
         }
+
+        stats.nhops += 1;
     }
 
     ++stats.n1;
@@ -853,7 +866,9 @@ HNSWStats HNSW::search(
         float d_nearest = qdis(nearest);
 
         for (int level = max_level; level >= 1; level--) {
-            greedy_update_nearest(*this, qdis, level, nearest, d_nearest);
+            HNSWStats local_stats = greedy_update_nearest(
+                    *this, qdis, level, nearest, d_nearest);
+            stats.combine(local_stats);
         }
 
         int ef = std::max(params ? params->efSearch : efSearch, k);
@@ -916,11 +931,23 @@ HNSWStats HNSW::search(
             if (level == 0) {
                 nres = search_from_candidates(
                         *this, qdis, res, candidates, vt, stats, 0);
+                nres = std::min(nres, candidates_size);
             } else {
+                const auto nres_prev = nres;
+
                 resh.begin(0);
                 nres = search_from_candidates(
                         *this, qdis, resh, candidates, vt, stats, level);
+                nres = std::min(nres, candidates_size);
                 resh.end();
+
+                // if the search on a particular level produces no improvements,
+                //   then we need to repopulate candidates.
+                // search_from_candidates() will always damage candidates
+                //   by doing 1 pop_min().
+                if (nres == 0) {
+                    nres = nres_prev;
+                }
             }
             vt.advance();
         }
@@ -970,6 +997,7 @@ void HNSW::search_level_0(
                     0,
                     nres,
                     params);
+            nres = std::min(nres, candidates_size);
         }
     } else if (search_type == 2) {
         int candidates_size = std::max(efSearch, int(k));
@@ -1051,7 +1079,99 @@ void HNSW::MinimaxHeap::clear() {
     nvalid = k = 0;
 }
 
-#ifdef __AVX2__
+#ifdef __AVX512F__
+
+int HNSW::MinimaxHeap::pop_min(float* vmin_out) {
+    assert(k > 0);
+    static_assert(
+            std::is_same<storage_idx_t, int32_t>::value,
+            "This code expects storage_idx_t to be int32_t");
+
+    int32_t min_idx = -1;
+    float min_dis = std::numeric_limits<float>::infinity();
+
+    __m512i min_indices = _mm512_set1_epi32(-1);
+    __m512 min_distances =
+            _mm512_set1_ps(std::numeric_limits<float>::infinity());
+    __m512i current_indices = _mm512_setr_epi32(
+            0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
+    __m512i offset = _mm512_set1_epi32(16);
+
+    // The following loop tracks the rightmost index with the min distance.
+    // -1 index values are ignored.
+    const int k16 = (k / 16) * 16;
+    for (size_t iii = 0; iii < k16; iii += 16) {
+        __m512i indices =
+                _mm512_loadu_si512((const __m512i*)(ids.data() + iii));
+        __m512 distances = _mm512_loadu_ps(dis.data() + iii);
+
+        // This mask filters out -1 values among indices.
+        __mmask16 m1mask =
+                _mm512_cmpgt_epi32_mask(_mm512_setzero_si512(), indices);
+
+        __mmask16 dmask =
+                _mm512_cmp_ps_mask(min_distances, distances, _CMP_LT_OS);
+        __mmask16 finalmask = m1mask | dmask;
+
+        const __m512i min_indices_new = _mm512_mask_blend_epi32(
+                finalmask, current_indices, min_indices);
+        const __m512 min_distances_new =
+                _mm512_mask_blend_ps(finalmask, distances, min_distances);
+
+        min_indices = min_indices_new;
+        min_distances = min_distances_new;
+
+        current_indices = _mm512_add_epi32(current_indices, offset);
+    }
+
+    // leftovers
+    if (k16 != k) {
+        const __mmask16 kmask = (1 << (k - k16)) - 1;
+
+        __m512i indices = _mm512_mask_loadu_epi32(
+                _mm512_set1_epi32(-1), kmask, ids.data() + k16);
+        __m512 distances = _mm512_maskz_loadu_ps(kmask, dis.data() + k16);
+
+        // This mask filters out -1 values among indices.
+        __mmask16 m1mask =
+                _mm512_cmpgt_epi32_mask(_mm512_setzero_si512(), indices);
+
+        __mmask16 dmask =
+                _mm512_cmp_ps_mask(min_distances, distances, _CMP_LT_OS);
+        __mmask16 finalmask = m1mask | dmask;
+
+        const __m512i min_indices_new = _mm512_mask_blend_epi32(
+                finalmask, current_indices, min_indices);
+        const __m512 min_distances_new =
+                _mm512_mask_blend_ps(finalmask, distances, min_distances);
+
+        min_indices = min_indices_new;
+        min_distances = min_distances_new;
+    }
+
+    // grab min distance
+    min_dis = _mm512_reduce_min_ps(min_distances);
+    // blend
+    __mmask16 mindmask =
+            _mm512_cmpeq_ps_mask(min_distances, _mm512_set1_ps(min_dis));
+    // pick the max one
+    min_idx = _mm512_mask_reduce_max_epi32(mindmask, min_indices);
+
+    if (min_idx == -1) {
+        return -1;
+    }
+
+    if (vmin_out) {
+        *vmin_out = min_dis;
+    }
+    int ret = ids[min_idx];
+    ids[min_idx] = -1;
+    --nvalid;
+    return ret;
+}
+
+#elif __AVX2__
+
 int HNSW::MinimaxHeap::pop_min(float* vmin_out) {
     assert(k > 0);
     static_assert(

faiss/impl/HNSW.h

@@ -234,20 +234,23 @@ struct HNSW {
 };
 
 struct HNSWStats {
-    size_t n1 = 0; /// numbner of vectors searched
+    size_t n1 = 0; /// number of vectors searched
     size_t n2 =
-            0; /// number of queries for which the candidate list is exhasted
-    size_t ndis = 0; /// number of distances computed
+            0; /// number of queries for which the candidate list is exhausted
+    size_t ndis = 0;  /// number of distances computed
+    size_t nhops = 0; /// number of hops aka number of edges traversed
 
     void reset() {
         n1 = n2 = 0;
         ndis = 0;
+        nhops = 0;
     }
 
     void combine(const HNSWStats& other) {
         n1 += other.n1;
         n2 += other.n2;
         ndis += other.ndis;
+        nhops += other.nhops;
     }
 };