faiss 0.1.5 → 0.2.2

data/vendor/faiss/faiss/IndexBinaryHNSW.h

@@ -9,48 +9,49 @@
 
 #pragma once
 
-#include <faiss/impl/HNSW.h>
 #include <faiss/IndexBinaryFlat.h>
+#include <faiss/impl/HNSW.h>
 #include <faiss/utils/utils.h>
 
-
 namespace faiss {
 
-
 /** The HNSW index is a normal random-access index with a HNSW
  * link structure built on top */
 
 struct IndexBinaryHNSW : IndexBinary {
-  typedef HNSW::storage_idx_t storage_idx_t;
+    typedef HNSW::storage_idx_t storage_idx_t;
 
-  // the link strcuture
-  HNSW hnsw;
+    // the link strcuture
+    HNSW hnsw;
 
-  // the sequential storage
-  bool own_fields;
-  IndexBinary *storage;
+    // the sequential storage
+    bool own_fields;
+    IndexBinary* storage;
 
-  explicit IndexBinaryHNSW();
-  explicit IndexBinaryHNSW(int d, int M = 32);
-  explicit IndexBinaryHNSW(IndexBinary *storage, int M = 32);
+    explicit IndexBinaryHNSW();
+    explicit IndexBinaryHNSW(int d, int M = 32);
+    explicit IndexBinaryHNSW(IndexBinary* storage, int M = 32);
 
-  ~IndexBinaryHNSW() override;
+    ~IndexBinaryHNSW() override;
 
-  DistanceComputer *get_distance_computer() const;
+    DistanceComputer* get_distance_computer() const;
 
-  void add(idx_t n, const uint8_t *x) override;
+    void add(idx_t n, const uint8_t* x) override;
 
-  /// Trains the storage if needed
-  void train(idx_t n, const uint8_t* x) override;
+    /// Trains the storage if needed
+    void train(idx_t n, const uint8_t* x) override;
 
-  /// entry point for search
-  void search(idx_t n, const uint8_t *x, idx_t k,
-              int32_t *distances, idx_t *labels) const override;
+    /// entry point for search
+    void search(
+            idx_t n,
+            const uint8_t* x,
+            idx_t k,
+            int32_t* distances,
+            idx_t* labels) const override;
 
-  void reconstruct(idx_t key, uint8_t* recons) const override;
+    void reconstruct(idx_t key, uint8_t* recons) const override;
 
-  void reset() override;
+    void reset() override;
 };
 
-
-}  // namespace faiss
+} // namespace faiss

data/vendor/faiss/faiss/IndexBinaryHash.cpp

@@ -22,44 +22,42 @@
 
 namespace faiss {
 
-void IndexBinaryHash::InvertedList::add (
-        idx_t id, size_t code_size, const uint8_t *code)
-{
+void IndexBinaryHash::InvertedList::add(
+        idx_t id,
+        size_t code_size,
+        const uint8_t* code) {
     ids.push_back(id);
     vecs.insert(vecs.end(), code, code + code_size);
 }
 
-IndexBinaryHash::IndexBinaryHash(int d, int b):
-    IndexBinary(d), b(b), nflip(0)
-{
+IndexBinaryHash::IndexBinaryHash(int d, int b)
+        : IndexBinary(d), b(b), nflip(0) {
     is_trained = true;
 }
 
-IndexBinaryHash::IndexBinaryHash(): b(0), nflip(0)
-{
+IndexBinaryHash::IndexBinaryHash() : b(0), nflip(0) {
     is_trained = true;
 }
 
-void IndexBinaryHash::reset()
-{
+void IndexBinaryHash::reset() {
     invlists.clear();
     ntotal = 0;
 }
 
-
-void IndexBinaryHash::add(idx_t n, const uint8_t *x)
-{
+void IndexBinaryHash::add(idx_t n, const uint8_t* x) {
     add_with_ids(n, x, nullptr);
 }
 
-void IndexBinaryHash::add_with_ids(idx_t n, const uint8_t *x, const idx_t *xids)
-{
+void IndexBinaryHash::add_with_ids(
+        idx_t n,
+        const uint8_t* x,
+        const idx_t* xids) {
     uint64_t mask = ((uint64_t)1 << b) - 1;
     // simplistic add function. Cannot really be parallelized.
 
     for (idx_t i = 0; i < n; i++) {
         idx_t id = xids ? xids[i] : ntotal + i;
-        const uint8_t * xi = x + i * code_size;
+        const uint8_t* xi = x + i * code_size;
         idx_t hash = *((uint64_t*)xi) & mask;
         invlists[hash].add(id, code_size, xi);
     }
@@ -68,7 +66,6 @@ void IndexBinaryHash::add_with_ids(idx_t n, const uint8_t *x, const idx_t *xids)
 
 namespace {
 
-
 /** Enumerate all bit vectors of size nbit with up to maxflip 1s
  * test in P127257851 P127258235
  */
@@ -76,7 +73,7 @@ struct FlipEnumerator {
     int nbit, nflip, maxflip;
     uint64_t mask, x;
 
-    FlipEnumerator (int nbit, int maxflip): nbit(nbit), maxflip(maxflip) {
+    FlipEnumerator(int nbit, int maxflip) : nbit(nbit), maxflip(maxflip) {
         nflip = 0;
         mask = 0;
         x = 0;
@@ -108,128 +105,133 @@ struct FlipEnumerator {
         }
         return true;
     }
-
 };
 
 using idx_t = Index::idx_t;
 
-
 struct RangeSearchResults {
     int radius;
-    RangeQueryResult &qres;
+    RangeQueryResult& qres;
 
-    inline void add (float dis, idx_t id) {
+    inline void add(float dis, idx_t id) {
         if (dis < radius) {
-            qres.add (dis, id);
+            qres.add(dis, id);
         }
     }
-
 };
 
 struct KnnSearchResults {
     // heap params
     idx_t k;
-    int32_t * heap_sim;
-    idx_t * heap_ids;
+    int32_t* heap_sim;
+    idx_t* heap_ids;
 
     using C = CMax<int, idx_t>;
 
-    inline void add (float dis, idx_t id) {
+    inline void add(float dis, idx_t id) {
         if (dis < heap_sim[0]) {
-            heap_replace_top<C> (k, heap_sim, heap_ids, dis, id);
+            heap_replace_top<C>(k, heap_sim, heap_ids, dis, id);
         }
     }
-
 };
 
-template<class HammingComputer, class SearchResults>
-void
-search_single_query_template(const IndexBinaryHash & index, const uint8_t *q,
-                    SearchResults &res,
-                    size_t &n0, size_t &nlist, size_t &ndis)
-{
+template <class HammingComputer, class SearchResults>
+void search_single_query_template(
+        const IndexBinaryHash& index,
+        const uint8_t* q,
+        SearchResults& res,
+        size_t& n0,
+        size_t& nlist,
+        size_t& ndis) {
     size_t code_size = index.code_size;
     uint64_t mask = ((uint64_t)1 << index.b) - 1;
     uint64_t qhash = *((uint64_t*)q) & mask;
-    HammingComputer hc (q, code_size);
+    HammingComputer hc(q, code_size);
     FlipEnumerator fe(index.b, index.nflip);
 
     // loop over neighbors that are at most at nflip bits
     do {
         uint64_t hash = qhash ^ fe.x;
-        auto it = index.invlists.find (hash);
+        auto it = index.invlists.find(hash);
 
         if (it == index.invlists.end()) {
             continue;
         }
 
-        const IndexBinaryHash::InvertedList &il = it->second;
+        const IndexBinaryHash::InvertedList& il = it->second;
 
         size_t nv = il.ids.size();
 
         if (nv == 0) {
             n0++;
         } else {
-            const uint8_t *codes = il.vecs.data();
+            const uint8_t* codes = il.vecs.data();
             for (size_t i = 0; i < nv; i++) {
-                int dis = hc.hamming (codes);
+                int dis = hc.hamming(codes);
                 res.add(dis, il.ids[i]);
                 codes += code_size;
             }
             ndis += nv;
             nlist++;
         }
-    } while(fe.next());
+    } while (fe.next());
 }
 
-template<class SearchResults>
-void
-search_single_query(const IndexBinaryHash & index, const uint8_t *q,
-                    SearchResults &res,
-                    size_t &n0, size_t &nlist, size_t &ndis)
-{
-#define HC(name) search_single_query_template<name>(index, q, res, n0, nlist, ndis);
-    switch(index.code_size) {
-    case 4: HC(HammingComputer4); break;
-    case 8: HC(HammingComputer8); break;
-    case 16: HC(HammingComputer16); break;
-    case 20: HC(HammingComputer20); break;
-    case 32: HC(HammingComputer32); break;
-    default:
-        if (index.code_size % 8 == 0) {
-            HC(HammingComputerM8);
-        } else {
+template <class SearchResults>
+void search_single_query(
+        const IndexBinaryHash& index,
+        const uint8_t* q,
+        SearchResults& res,
+        size_t& n0,
+        size_t& nlist,
+        size_t& ndis) {
+#define HC(name) \
+    search_single_query_template<name>(index, q, res, n0, nlist, ndis);
+    switch (index.code_size) {
+        case 4:
+            HC(HammingComputer4);
+            break;
+        case 8:
+            HC(HammingComputer8);
+            break;
+        case 16:
+            HC(HammingComputer16);
+            break;
+        case 20:
+            HC(HammingComputer20);
+            break;
+        case 32:
+            HC(HammingComputer32);
+            break;
+        default:
             HC(HammingComputerDefault);
-        }
+            break;
     }
 #undef HC
 }
 
-
 } // anonymous namespace
 
-
-
-void IndexBinaryHash::range_search(idx_t n, const uint8_t *x, int radius,
-                                   RangeSearchResult *result) const
-{
-
+void IndexBinaryHash::range_search(
+        idx_t n,
+        const uint8_t* x,
+        int radius,
+        RangeSearchResult* result) const {
     size_t nlist = 0, ndis = 0, n0 = 0;
 
-#pragma omp parallel if(n > 100) reduction(+: ndis, n0, nlist)
+#pragma omp parallel if (n > 100) reduction(+ : ndis, n0, nlist)
     {
-        RangeSearchPartialResult pres (result);
+        RangeSearchPartialResult pres(result);
 
 #pragma omp for
         for (idx_t i = 0; i < n; i++) { // loop queries
-            RangeQueryResult & qres = pres.new_result (i);
+            RangeQueryResult& qres = pres.new_result(i);
             RangeSearchResults res = {radius, qres};
-            const uint8_t *q = x + i * code_size;
-
-            search_single_query (*this, q, res, n0, nlist, ndis);
+            const uint8_t* q = x + i * code_size;
 
+            search_single_query(*this, q, res, n0, nlist, ndis);
         }
-        pres.finalize ();
+        pres.finalize();
     }
     indexBinaryHash_stats.nq += n;
     indexBinaryHash_stats.n0 += n0;
@@ -237,24 +239,29 @@ void IndexBinaryHash::range_search(idx_t n, const uint8_t *x, int radius,
     indexBinaryHash_stats.ndis += ndis;
 }
 
-void IndexBinaryHash::search(idx_t n, const uint8_t *x, idx_t k,
-                             int32_t *distances, idx_t *labels) const
-{
+void IndexBinaryHash::search(
+        idx_t n,
+        const uint8_t* x,
+        idx_t k,
+        int32_t* distances,
+        idx_t* labels) const {
+    FAISS_THROW_IF_NOT(k > 0);
 
     using HeapForL2 = CMax<int32_t, idx_t>;
     size_t nlist = 0, ndis = 0, n0 = 0;
 
-#pragma omp parallel for if(n > 100) reduction(+: nlist, ndis, n0)
+#pragma omp parallel for if (n > 100) reduction(+ : nlist, ndis, n0)
     for (idx_t i = 0; i < n; i++) {
-        int32_t * simi = distances + k * i;
-        idx_t * idxi = labels + k * i;
+        int32_t* simi = distances + k * i;
+        idx_t* idxi = labels + k * i;
 
-        heap_heapify<HeapForL2> (k, simi, idxi);
+        heap_heapify<HeapForL2>(k, simi, idxi);
         KnnSearchResults res = {k, simi, idxi};
-        const uint8_t *q = x + i * code_size;
+        const uint8_t* q = x + i * code_size;
 
-        search_single_query (*this, q, res, n0, nlist, ndis);
+        search_single_query(*this, q, res, n0, nlist, ndis);
 
+        heap_reorder<HeapForL2>(k, simi, idxi);
     }
     indexBinaryHash_stats.nq += n;
     indexBinaryHash_stats.n0 += n0;
@@ -262,29 +269,23 @@ void IndexBinaryHash::search(idx_t n, const uint8_t *x, idx_t k,
     indexBinaryHash_stats.ndis += ndis;
 }
 
-size_t IndexBinaryHash::hashtable_size() const
-{
+size_t IndexBinaryHash::hashtable_size() const {
     return invlists.size();
 }
 
-
-void IndexBinaryHash::display() const
-{
+void IndexBinaryHash::display() const {
     for (auto it = invlists.begin(); it != invlists.end(); ++it) {
         printf("%" PRId64 ": [", it->first);
-        const std::vector<idx_t> & v = it->second.ids;
-        for (auto x: v) {
+        const std::vector<idx_t>& v = it->second.ids;
+        for (auto x : v) {
             printf("%" PRId64 " ", x);
         }
         printf("]\n");
-
     }
 }
 
-
-void IndexBinaryHashStats::reset()
-{
-    memset ((void*)this, 0, sizeof (*this));
+void IndexBinaryHashStats::reset() {
+    memset((void*)this, 0, sizeof(*this));
 }
 
 IndexBinaryHashStats indexBinaryHash_stats;
@@ -293,47 +294,43 @@ IndexBinaryHashStats indexBinaryHash_stats;
  * IndexBinaryMultiHash implementation
  ******************************************************/
 
-
-IndexBinaryMultiHash::IndexBinaryMultiHash(int d, int nhash, int b):
-    IndexBinary(d),
-    storage(new IndexBinaryFlat(d)), own_fields(true),
-    maps(nhash), nhash(nhash), b(b), nflip(0)
-{
+IndexBinaryMultiHash::IndexBinaryMultiHash(int d, int nhash, int b)
+        : IndexBinary(d),
+          storage(new IndexBinaryFlat(d)),
+          own_fields(true),
+          maps(nhash),
+          nhash(nhash),
+          b(b),
+          nflip(0) {
     FAISS_THROW_IF_NOT(nhash * b <= d);
 }
 
-IndexBinaryMultiHash::IndexBinaryMultiHash():
-    storage(nullptr), own_fields(true),
-    nhash(0), b(0), nflip(0)
-{}
+IndexBinaryMultiHash::IndexBinaryMultiHash()
+        : storage(nullptr), own_fields(true), nhash(0), b(0), nflip(0) {}
 
-IndexBinaryMultiHash::~IndexBinaryMultiHash()
-{
+IndexBinaryMultiHash::~IndexBinaryMultiHash() {
     if (own_fields) {
         delete storage;
     }
 }
 
-
-void IndexBinaryMultiHash::reset()
-{
+void IndexBinaryMultiHash::reset() {
     storage->reset();
     ntotal = 0;
-    for(auto map: maps) {
+    for (auto map : maps) {
         map.clear();
     }
 }
 
-void IndexBinaryMultiHash::add(idx_t n, const uint8_t *x)
-{
+void IndexBinaryMultiHash::add(idx_t n, const uint8_t* x) {
     storage->add(n, x);
     // populate maps
     uint64_t mask = ((uint64_t)1 << b) - 1;
 
-    for(idx_t i = 0; i < n; i++) {
-        const uint8_t *xi = x + i * code_size;
+    for (idx_t i = 0; i < n; i++) {
+        const uint8_t* xi = x + i * code_size;
         int ho = 0;
-        for(int h = 0; h < nhash; h++) {
+        for (int h = 0; h < nhash; h++) {
             uint64_t hash = *(uint64_t*)(xi + (ho >> 3)) >> (ho & 7);
             hash &= mask;
             maps[h][hash].push_back(i + ntotal);
@@ -343,62 +340,60 @@ void IndexBinaryMultiHash::add(idx_t n, const uint8_t *x)
     ntotal += n;
 }
 
-
 namespace {
 
 template <class HammingComputer, class SearchResults>
-static
-void verify_shortlist(
-        const IndexBinaryFlat & index,
-        const uint8_t * q,
-        const std::unordered_set<Index::idx_t> & shortlist,
-        SearchResults &res)
-{
+static void verify_shortlist(
+        const IndexBinaryFlat& index,
+        const uint8_t* q,
+        const std::unordered_set<Index::idx_t>& shortlist,
+        SearchResults& res) {
     size_t code_size = index.code_size;
     size_t nlist = 0, ndis = 0, n0 = 0;
 
-    HammingComputer hc (q, code_size);
-    const uint8_t *codes = index.xb.data();
+    HammingComputer hc(q, code_size);
+    const uint8_t* codes = index.xb.data();
 
-    for (auto i: shortlist) {
-        int dis = hc.hamming (codes + i * code_size);
+    for (auto i : shortlist) {
+        int dis = hc.hamming(codes + i * code_size);
         res.add(dis, i);
     }
 }
 
-template<class SearchResults>
-void
-search_1_query_multihash(const IndexBinaryMultiHash & index, const uint8_t *xi,
-                         SearchResults &res,
-                         size_t &n0, size_t &nlist, size_t &ndis)
-{
-
+template <class SearchResults>
+void search_1_query_multihash(
+        const IndexBinaryMultiHash& index,
+        const uint8_t* xi,
+        SearchResults& res,
+        size_t& n0,
+        size_t& nlist,
+        size_t& ndis) {
     std::unordered_set<idx_t> shortlist;
     int b = index.b;
     uint64_t mask = ((uint64_t)1 << b) - 1;
 
     int ho = 0;
-    for(int h = 0; h < index.nhash; h++) {
+    for (int h = 0; h < index.nhash; h++) {
         uint64_t qhash = *(uint64_t*)(xi + (ho >> 3)) >> (ho & 7);
         qhash &= mask;
-        const IndexBinaryMultiHash::Map & map = index.maps[h];
+        const IndexBinaryMultiHash::Map& map = index.maps[h];
 
         FlipEnumerator fe(index.b, index.nflip);
         // loop over neighbors that are at most at nflip bits
         do {
             uint64_t hash = qhash ^ fe.x;
-            auto it = map.find (hash);
+            auto it = map.find(hash);
 
             if (it != map.end()) {
-                const std::vector<idx_t> & v = it->second;
-                for (auto i: v) {
+                const std::vector<idx_t>& v = it->second;
+                for (auto i : v) {
                     shortlist.insert(i);
                 }
                 nlist++;
             } else {
                 n0++;
             }
-        } while(fe.next());
+        } while (fe.next());
 
         ho += b;
     }
@@ -406,45 +401,52 @@ search_1_query_multihash(const IndexBinaryMultiHash & index, const uint8_t *xi,
 
     // verify shortlist
 
-#define HC(name) verify_shortlist<name> (*index.storage, xi, shortlist, res)
-    switch(index.code_size) {
-    case 4: HC(HammingComputer4); break;
-    case 8: HC(HammingComputer8); break;
-    case 16: HC(HammingComputer16); break;
-    case 20: HC(HammingComputer20); break;
-    case 32: HC(HammingComputer32); break;
-    default:
-        if (index.code_size % 8 == 0) {
-            HC(HammingComputerM8);
-        } else {
+#define HC(name) verify_shortlist<name>(*index.storage, xi, shortlist, res)
+    switch (index.code_size) {
+        case 4:
+            HC(HammingComputer4);
+            break;
+        case 8:
+            HC(HammingComputer8);
+            break;
+        case 16:
+            HC(HammingComputer16);
+            break;
+        case 20:
+            HC(HammingComputer20);
+            break;
+        case 32:
+            HC(HammingComputer32);
+            break;
+        default:
             HC(HammingComputerDefault);
-        }
+            break;
     }
 #undef HC
 }
 
 } // anonymous namespace
 
-void IndexBinaryMultiHash::range_search(idx_t n, const uint8_t *x, int radius,
-                                   RangeSearchResult *result) const
-{
-
+void IndexBinaryMultiHash::range_search(
+        idx_t n,
+        const uint8_t* x,
+        int radius,
+        RangeSearchResult* result) const {
     size_t nlist = 0, ndis = 0, n0 = 0;
 
-#pragma omp parallel if(n > 100) reduction(+: ndis, n0, nlist)
+#pragma omp parallel if (n > 100) reduction(+ : ndis, n0, nlist)
     {
-        RangeSearchPartialResult pres (result);
+        RangeSearchPartialResult pres(result);
 
 #pragma omp for
         for (idx_t i = 0; i < n; i++) { // loop queries
-            RangeQueryResult & qres = pres.new_result (i);
+            RangeQueryResult& qres = pres.new_result(i);
             RangeSearchResults res = {radius, qres};
-            const uint8_t *q = x + i * code_size;
-
-            search_1_query_multihash (*this, q, res, n0, nlist, ndis);
+            const uint8_t* q = x + i * code_size;
 
+            search_1_query_multihash(*this, q, res, n0, nlist, ndis);
         }
-        pres.finalize ();
+        pres.finalize();
     }
     indexBinaryHash_stats.nq += n;
     indexBinaryHash_stats.n0 += n0;
@@ -452,24 +454,29 @@ void IndexBinaryMultiHash::range_search(idx_t n, const uint8_t *x, int radius,
     indexBinaryHash_stats.ndis += ndis;
 }
 
-void IndexBinaryMultiHash::search(idx_t n, const uint8_t *x, idx_t k,
-                             int32_t *distances, idx_t *labels) const
-{
+void IndexBinaryMultiHash::search(
+        idx_t n,
+        const uint8_t* x,
+        idx_t k,
+        int32_t* distances,
+        idx_t* labels) const {
+    FAISS_THROW_IF_NOT(k > 0);
 
     using HeapForL2 = CMax<int32_t, idx_t>;
     size_t nlist = 0, ndis = 0, n0 = 0;
 
-#pragma omp parallel for if(n > 100) reduction(+: nlist, ndis, n0)
+#pragma omp parallel for if (n > 100) reduction(+ : nlist, ndis, n0)
     for (idx_t i = 0; i < n; i++) {
-        int32_t * simi = distances + k * i;
-        idx_t * idxi = labels + k * i;
+        int32_t* simi = distances + k * i;
+        idx_t* idxi = labels + k * i;
 
-        heap_heapify<HeapForL2> (k, simi, idxi);
+        heap_heapify<HeapForL2>(k, simi, idxi);
         KnnSearchResults res = {k, simi, idxi};
-        const uint8_t *q = x + i * code_size;
+        const uint8_t* q = x + i * code_size;
 
-        search_1_query_multihash (*this, q, res, n0, nlist, ndis);
+        search_1_query_multihash(*this, q, res, n0, nlist, ndis);
 
+        heap_reorder<HeapForL2>(k, simi, idxi);
     }
     indexBinaryHash_stats.nq += n;
     indexBinaryHash_stats.n0 += n0;
@@ -477,15 +484,13 @@ void IndexBinaryMultiHash::search(idx_t n, const uint8_t *x, idx_t k,
     indexBinaryHash_stats.ndis += ndis;
 }
 
-size_t IndexBinaryMultiHash::hashtable_size() const
-{
+size_t IndexBinaryMultiHash::hashtable_size() const {
     size_t tot = 0;
-    for (auto map: maps) {
+    for (auto map : maps) {
         tot += map.size();
     }
 
     return tot;
 }
 
-
-}
+} // namespace faiss