hnswlib 0.4.0 → 0.5.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 99f5f1403a51083df75ef842a11b996fe6c159f95c3798f217a076cb5f535254
-  data.tar.gz: f9972ac4e644727a126e937e81494cc11bc051bc7ff0daf57139845aed1e7719
+  metadata.gz: ec8d1a750f703968418f3aca9beffc33c061c9c5e7e6d24050a8aa8cd15e0d35
+  data.tar.gz: e8bb3c9b706dfd6d6d25d5f471970b50a6f191efbe9719437d0ce00c0c7a1420
 SHA512:
-  metadata.gz: c4f96fceab228e3cffd9f2b77e513b86fe71ec3156cdff10c0bc9206ad7951f87c507531088f91f45a940c399102bff6a2dedc5fee8b09b37a5f5739e9a02b6c
-  data.tar.gz: 8297e1e0d1f6753b3e8900b07b80dba271276a14303b102cd8084e480ec491da269baac1fc8be08ab350976fc22cb859583745dc219d73d6f56b33fe7502ac2b
+  metadata.gz: 8f39ec12aac12a0af0a4882f092c61b29417d14c575580b821ccb2154568e9c04e6a4170617cfa0185ca32dc322181cd18dd116bd527fc3f61b52cee3180b648
+  data.tar.gz: e355334afba1ed2f8817e2dda8eedd4bc61670f08ea349e44b7259c01584a195e2bb33f3de9f2f5d4ce59cd6cb80bd17b7f0c3b5e5dedd26dd0d8880c75d25ad

data/CHANGELOG.md

@@ -1,3 +1,7 @@
+## [0.5.0] - 2021-12-12
+
+- Update bundled hnswlib version to 0.6.0.
+
 ## [0.4.0] - 2021-09-12
 
 - Add type declaration file.

data/ext/hnswlib/src/bruteforce.h

@@ -92,13 +92,13 @@ namespace hnswlib {
         searchKnn(const void *query_data, size_t k) const {
             std::priority_queue<std::pair<dist_t, labeltype >> topResults;
             if (cur_element_count == 0) return topResults;
-            for (size_t i = 0; i < k; i++) {
+            for (int i = 0; i < k; i++) {
                 dist_t dist = fstdistfunc_(query_data, data_ + size_per_element_ * i, dist_func_param_);
                 topResults.push(std::pair<dist_t, labeltype>(dist, *((labeltype *) (data_ + size_per_element_ * i +
                                                                                     data_size_))));
             }
             dist_t lastdist = topResults.top().first;
-            for (size_t i = k; i < cur_element_count; i++) {
+            for (int i = k; i < cur_element_count; i++) {
                 dist_t dist = fstdistfunc_(query_data, data_ + size_per_element_ * i, dist_func_param_);
                 if (dist <= lastdist) {
                     topResults.push(std::pair<dist_t, labeltype>(dist, *((labeltype *) (data_ + size_per_element_ * i +

data/ext/hnswlib/src/hnswalg.h

@@ -19,7 +19,6 @@ namespace hnswlib {
         static const tableint max_update_element_locks = 65536;
         HierarchicalNSW() : visited_list_pool_(nullptr), data_level0_memory_(nullptr), linkLists_(nullptr), cur_element_count(0) { }
         HierarchicalNSW(SpaceInterface<dist_t> *s) {
-
         }
 
         HierarchicalNSW(SpaceInterface<dist_t> *s, const std::string &location, bool nmslib = false, size_t max_elements=0) {
@@ -30,7 +29,7 @@ namespace hnswlib {
                 link_list_locks_(max_elements), link_list_update_locks_(max_update_element_locks), element_levels_(max_elements) {
             max_elements_ = max_elements;
 
-            has_deletions_=false;
+            num_deleted_ = 0;
             data_size_ = s->get_data_size();
             fstdistfunc_ = s->get_dist_func();
             dist_func_param_ = s->get_dist_func_param();
@@ -57,8 +56,6 @@ namespace hnswlib {
 
             visited_list_pool_ = new VisitedListPool(1, max_elements);
 
-
-
             //initializations for special treatment of the first node
             enterpoint_node_ = -1;
             maxlevel_ = -1;
@@ -93,6 +90,7 @@ namespace hnswlib {
         size_t cur_element_count;
         size_t size_data_per_element_;
         size_t size_links_per_element_;
+        size_t num_deleted_;
 
         size_t M_;
         size_t maxM_;
@@ -113,20 +111,15 @@ namespace hnswlib {
         std::vector<std::mutex> link_list_update_locks_;
         tableint enterpoint_node_;
 
-
         size_t size_links_level0_;
         size_t offsetData_, offsetLevel0_;
 
-
         char *data_level0_memory_;
         char **linkLists_;
         std::vector<int> element_levels_;
 
         size_t data_size_;
 
-        bool has_deletions_;
-
-
         size_t label_offset_;
         DISTFUNC<dist_t> fstdistfunc_;
         void *dist_func_param_;
@@ -183,7 +176,7 @@ namespace hnswlib {
 
             while (!candidateSet.empty()) {
                 std::pair<dist_t, tableint> curr_el_pair = candidateSet.top();
-                if ((-curr_el_pair.first) > lowerBound) {
+                if ((-curr_el_pair.first) > lowerBound && top_candidates.size() == ef_construction_) {
                     break;
                 }
                 candidateSet.pop();
@@ -272,7 +265,7 @@ namespace hnswlib {
 
                 std::pair<dist_t, tableint> current_node_pair = candidate_set.top();
 
-                if ((-current_node_pair.first) > lowerBound) {
+                if ((-current_node_pair.first) > lowerBound && (top_candidates.size() == ef || has_deletions == false)) {
                     break;
                 }
                 candidate_set.pop();
@@ -548,7 +541,7 @@ namespace hnswlib {
                 }
             }
 
-            if (has_deletions_) {
+            if (num_deleted_) {
                 std::priority_queue<std::pair<dist_t, tableint  >> top_candidates1=searchBaseLayerST<true>(currObj, query_data,
                                                                                                            ef_);
                 top_candidates.swap(top_candidates1);
@@ -624,8 +617,6 @@ namespace hnswlib {
         }
 
         void loadIndex(const std::string &location, SpaceInterface<dist_t> *s, size_t max_elements_i=0) {
-
-
             std::ifstream input(location, std::ios::binary);
 
             if (!input.is_open())
@@ -640,7 +631,7 @@ namespace hnswlib {
             readBinaryPOD(input, max_elements_);
             readBinaryPOD(input, cur_element_count);
 
-            size_t max_elements=max_elements_i;
+            size_t max_elements = max_elements_i;
             if(max_elements < cur_element_count)
                 max_elements = max_elements_;
             max_elements_ = max_elements;
@@ -689,26 +680,19 @@ namespace hnswlib {
 
             input.seekg(pos,input.beg);
 
-
             data_level0_memory_ = (char *) malloc(max_elements * size_data_per_element_);
             if (data_level0_memory_ == nullptr)
                 throw std::runtime_error("Not enough memory: loadIndex failed to allocate level0");
             input.read(data_level0_memory_, cur_element_count * size_data_per_element_);
 
-
-
-
             size_links_per_element_ = maxM_ * sizeof(tableint) + sizeof(linklistsizeint);
 
-
             size_links_level0_ = maxM0_ * sizeof(tableint) + sizeof(linklistsizeint);
             std::vector<std::mutex>(max_elements).swap(link_list_locks_);
             std::vector<std::mutex>(max_update_element_locks).swap(link_list_update_locks_);
 
-
             visited_list_pool_ = new VisitedListPool(1, max_elements);
 
-
             linkLists_ = (char **) malloc(sizeof(void *) * max_elements);
             if (linkLists_ == nullptr)
                 throw std::runtime_error("Not enough memory: loadIndex failed to allocate linklists");
@@ -732,11 +716,9 @@ namespace hnswlib {
                 }
             }
 
-            has_deletions_=false;
-
             for (size_t i = 0; i < cur_element_count; i++) {
                 if(isMarkedDeleted(i))
-                    has_deletions_=true;
+                    num_deleted_ += 1;
             }
 
             input.close();
@@ -745,7 +727,7 @@ namespace hnswlib {
         }
 
         template<typename data_t>
-        std::vector<data_t> getDataByLabel(labeltype label)
+        std::vector<data_t> getDataByLabel(labeltype label) const
         {
             tableint label_c;
             auto search = label_lookup_.find(label);
@@ -758,7 +740,7 @@ namespace hnswlib {
             size_t dim = *((size_t *) dist_func_param_);
             std::vector<data_t> data;
             data_t* data_ptr = (data_t*) data_ptrv;
-            for (size_t i = 0; i < dim; i++) {
+            for (int i = 0; i < dim; i++) {
                 data.push_back(*data_ptr);
                 data_ptr += 1;
             }
@@ -766,19 +748,19 @@ namespace hnswlib {
         }
 
         static const unsigned char DELETE_MARK = 0x01;
-//        static const unsigned char REUSE_MARK = 0x10;
+        // static const unsigned char REUSE_MARK = 0x10;
         /**
          * Marks an element with the given label deleted, does NOT really change the current graph.
          * @param label
          */
         void markDelete(labeltype label)
         {
-            has_deletions_=true;
             auto search = label_lookup_.find(label);
             if (search == label_lookup_.end()) {
                 throw std::runtime_error("Label not found");
             }
-            markDeletedInternal(search->second);
+            tableint internalId = search->second;
+            markDeletedInternal(internalId);
         }
 
         /**
@@ -787,8 +769,31 @@ namespace hnswlib {
          * @param internalId
          */
         void markDeletedInternal(tableint internalId) {
-            unsigned char *ll_cur = ((unsigned char *)get_linklist0(internalId))+2;
-            *ll_cur |= DELETE_MARK;
+            assert(internalId < cur_element_count);
+            if (!isMarkedDeleted(internalId))
+            {
+                unsigned char *ll_cur = ((unsigned char *)get_linklist0(internalId))+2;
+                *ll_cur |= DELETE_MARK;
+                num_deleted_ += 1;
+            }
+            else
+            {
+                throw std::runtime_error("The requested to delete element is already deleted");
+            }
+        }
+
+        /**
+         * Remove the deleted mark of the node, does NOT really change the current graph.
+         * @param label
+         */
+        void unmarkDelete(labeltype label)
+        {
+            auto search = label_lookup_.find(label);
+            if (search == label_lookup_.end()) {
+                throw std::runtime_error("Label not found");
+            }
+            tableint internalId = search->second;
+            unmarkDeletedInternal(internalId);
         }
 
         /**
@@ -796,8 +801,17 @@ namespace hnswlib {
          * @param internalId
          */
         void unmarkDeletedInternal(tableint internalId) {
-            unsigned char *ll_cur = ((unsigned char *)get_linklist0(internalId))+2;
-            *ll_cur &= ~DELETE_MARK;
+            assert(internalId < cur_element_count);
+            if (isMarkedDeleted(internalId))
+            {
+                unsigned char *ll_cur = ((unsigned char *)get_linklist0(internalId))+2;
+                *ll_cur &= ~DELETE_MARK;
+                num_deleted_ -= 1;
+            }
+            else
+            {
+                throw std::runtime_error("The requested to undelete element is not deleted");
+            }
         }
 
         /**
@@ -858,8 +872,8 @@ namespace hnswlib {
                 }
 
                 for (auto&& neigh : sNeigh) {
-//                    if (neigh == internalId)
-//                        continue;
+                    // if (neigh == internalId)
+                    //     continue;
 
                     std::priority_queue<std::pair<dist_t, tableint>, std::vector<std::pair<dist_t, tableint>>, CompareByFirst> candidates;
                     size_t size = sCand.find(neigh) == sCand.end() ? sCand.size() : sCand.size() - 1; // sCand guaranteed to have size >= 1
@@ -990,7 +1004,6 @@ namespace hnswlib {
                         unmarkDeletedInternal(existingInternalId);
                     }
                     updatePoint(data_point, existingInternalId, 1.0);
-                    
                     return existingInternalId;
                 }
 
@@ -1134,7 +1147,7 @@ namespace hnswlib {
             }
 
             std::priority_queue<std::pair<dist_t, tableint>, std::vector<std::pair<dist_t, tableint>>, CompareByFirst> top_candidates;
-            if (has_deletions_) {
+            if (num_deleted_) {
                 top_candidates=searchBaseLayerST<true,true>(
                         currObj, query_data, std::max(ef_, k));
             }

data/ext/hnswlib/src/hnswlib.h

@@ -4,6 +4,9 @@
 #define USE_SSE
 #ifdef __AVX__
 #define USE_AVX
+#ifdef __AVX512F__
+#define USE_AVX512
+#endif
 #endif
 #endif
 #endif
@@ -16,10 +19,16 @@
 #include <x86intrin.h>
 #endif
 
+#if defined(USE_AVX512)
+#include <immintrin.h>
+#endif
+
 #if defined(__GNUC__)
 #define PORTABLE_ALIGN32 __attribute__((aligned(32)))
+#define PORTABLE_ALIGN64 __attribute__((aligned(64)))
 #else
 #define PORTABLE_ALIGN32 __declspec(align(32))
+#define PORTABLE_ALIGN64 __declspec(align(64))
 #endif
 #endif
 

data/ext/hnswlib/src/space_ip.h

@@ -124,7 +124,40 @@ namespace hnswlib {
 
 #endif
 
-#if defined(USE_AVX)
+
+#if defined(USE_AVX512)
+
+    static float
+    InnerProductSIMD16Ext(const void *pVect1v, const void *pVect2v, const void *qty_ptr) {
+        float PORTABLE_ALIGN64 TmpRes[16];
+        float *pVect1 = (float *) pVect1v;
+        float *pVect2 = (float *) pVect2v;
+        size_t qty = *((size_t *) qty_ptr);
+
+        size_t qty16 = qty / 16;
+
+
+        const float *pEnd1 = pVect1 + 16 * qty16;
+
+        __m512 sum512 = _mm512_set1_ps(0);
+
+        while (pVect1 < pEnd1) {
+            //_mm_prefetch((char*)(pVect2 + 16), _MM_HINT_T0);
+
+            __m512 v1 = _mm512_loadu_ps(pVect1);
+            pVect1 += 16;
+            __m512 v2 = _mm512_loadu_ps(pVect2);
+            pVect2 += 16;
+            sum512 = _mm512_add_ps(sum512, _mm512_mul_ps(v1, v2));
+        }
+
+        _mm512_store_ps(TmpRes, sum512);
+        float sum = TmpRes[0] + TmpRes[1] + TmpRes[2] + TmpRes[3] + TmpRes[4] + TmpRes[5] + TmpRes[6] + TmpRes[7] + TmpRes[8] + TmpRes[9] + TmpRes[10] + TmpRes[11] + TmpRes[12] + TmpRes[13] + TmpRes[14] + TmpRes[15];
+
+        return 1.0f - sum;
+    }
+
+#elif defined(USE_AVX)
 
     static float
     InnerProductSIMD16Ext(const void *pVect1v, const void *pVect2v, const void *qty_ptr) {
@@ -211,7 +244,7 @@ namespace hnswlib {
 
 #endif
 
-#if defined(USE_SSE) || defined(USE_AVX)
+#if defined(USE_SSE) || defined(USE_AVX) || defined(USE_AVX512)
     static float
     InnerProductSIMD16ExtResiduals(const void *pVect1v, const void *pVect2v, const void *qty_ptr) {
         size_t qty = *((size_t *) qty_ptr);
@@ -250,7 +283,7 @@ namespace hnswlib {
         InnerProductSpace() : data_size_(0), dim_(0) { }
         InnerProductSpace(size_t dim) {
             fstdistfunc_ = InnerProduct;
-    #if defined(USE_AVX) || defined(USE_SSE)
+    #if defined(USE_AVX) || defined(USE_SSE) || defined(USE_AVX512)
             if (dim % 16 == 0)
                 fstdistfunc_ = InnerProductSIMD16Ext;
             else if (dim % 4 == 0)

data/ext/hnswlib/src/space_l2.h

@@ -19,7 +19,41 @@ namespace hnswlib {
         return (res);
     }
 
-#if defined(USE_AVX)
+#if defined(USE_AVX512)
+
+    // Favor using AVX512 if available.
+    static float
+    L2SqrSIMD16Ext(const void *pVect1v, const void *pVect2v, const void *qty_ptr) {
+        float *pVect1 = (float *) pVect1v;
+        float *pVect2 = (float *) pVect2v;
+        size_t qty = *((size_t *) qty_ptr);
+        float PORTABLE_ALIGN64 TmpRes[16];
+        size_t qty16 = qty >> 4;
+
+        const float *pEnd1 = pVect1 + (qty16 << 4);
+
+        __m512 diff, v1, v2;
+        __m512 sum = _mm512_set1_ps(0);
+
+        while (pVect1 < pEnd1) {
+            v1 = _mm512_loadu_ps(pVect1);
+            pVect1 += 16;
+            v2 = _mm512_loadu_ps(pVect2);
+            pVect2 += 16;
+            diff = _mm512_sub_ps(v1, v2);
+            // sum = _mm512_fmadd_ps(diff, diff, sum);
+            sum = _mm512_add_ps(sum, _mm512_mul_ps(diff, diff));
+        }
+
+        _mm512_store_ps(TmpRes, sum);
+        float res = TmpRes[0] + TmpRes[1] + TmpRes[2] + TmpRes[3] + TmpRes[4] + TmpRes[5] + TmpRes[6] +
+                TmpRes[7] + TmpRes[8] + TmpRes[9] + TmpRes[10] + TmpRes[11] + TmpRes[12] +
+                TmpRes[13] + TmpRes[14] + TmpRes[15];
+
+        return (res);
+}
+
+#elif defined(USE_AVX)
 
     // Favor using AVX if available.
     static float
@@ -106,7 +140,7 @@ namespace hnswlib {
     }
 #endif
 
-#if defined(USE_SSE) || defined(USE_AVX)
+#if defined(USE_SSE) || defined(USE_AVX) || defined(USE_AVX512)
     static float
     L2SqrSIMD16ExtResiduals(const void *pVect1v, const void *pVect2v, const void *qty_ptr) {
         size_t qty = *((size_t *) qty_ptr);
@@ -175,7 +209,7 @@ namespace hnswlib {
         L2Space() : data_size_(0), dim_(0) { }
         L2Space(size_t dim) {
             fstdistfunc_ = L2Sqr;
-        #if defined(USE_SSE) || defined(USE_AVX)
+        #if defined(USE_SSE) || defined(USE_AVX) || defined(USE_AVX512)
             if (dim % 16 == 0)
                 fstdistfunc_ = L2SqrSIMD16Ext;
             else if (dim % 4 == 0)
@@ -279,4 +313,4 @@ namespace hnswlib {
     };
 
 
-}
+}

data/ext/hnswlib/src/visited_list_pool.h

@@ -2,6 +2,7 @@
 
 #include <mutex>
 #include <string.h>
+#include <deque>
 
 namespace hnswlib {
     typedef unsigned short int vl_type;

data/lib/hnswlib/version.rb

@@ -3,8 +3,8 @@
 # Hnswlib.rb provides Ruby bindings for the Hnswlib.
 module Hnswlib
   # The version of Hnswlib.rb you install.
-  VERSION = '0.4.0'
+  VERSION = '0.5.0'
 
   # The version of Hnswlib included with gem.
-  HSWLIB_VERSION = '0.5.2'
+  HSWLIB_VERSION = '0.6.0'
 end

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: hnswlib
 version: !ruby/object:Gem::Version
-  version: 0.4.0
+  version: 0.5.0
 platform: ruby
 authors:
 - yoshoku
 autorequire:
 bindir: exe
 cert_chain: []
-date: 2021-09-11 00:00:00.000000000 Z
+date: 2021-12-11 00:00:00.000000000 Z
 dependencies: []
 description: Hnswlib.rb provides Ruby bindings for the Hnswlib.
 email:
@@ -64,7 +64,7 @@ required_rubygems_version: !ruby/object:Gem::Requirement
     - !ruby/object:Gem::Version
       version: '0'
 requirements: []
-rubygems_version: 3.2.22
+rubygems_version: 3.2.32
 signing_key:
 specification_version: 4
 summary: Ruby bindings for the Hnswlib.