faiss 0.6.0 → 0.6.1

data/vendor/faiss/faiss/IndexIVFPQ.cpp

@@ -9,7 +9,6 @@
 
 #include <faiss/IndexIVFPQ.h>
 
-#include <cassert>
 #include <cinttypes>
 #include <cmath>
 #include <cstdint>
@@ -17,7 +16,6 @@
 
 #include <algorithm>
 
-#include <faiss/utils/Heap.h>
 #include <faiss/utils/distances_dispatch.h>
 #include <faiss/utils/utils.h>
 
@@ -30,9 +28,15 @@
 #include <faiss/impl/IDSelector.h>
 #include <faiss/impl/ProductQuantizer.h>
 #include <faiss/impl/ResultHandler.h>
-#include <faiss/impl/pq_code_distance/pq_code_distance-inl.h>
+#include <faiss/impl/pq_code_distance/pq_code_distance-generic.h>
 #include <faiss/impl/simd_dispatch.h>
 
+// Scalar (NONE) fallback for dynamic dispatch
+#define THE_SIMD_LEVEL SIMDLevel::NONE
+// NOLINTNEXTLINE(facebook-hte-InlineHeader)
+#include <faiss/impl/pq_code_distance/IVFPQScanner_impl.h>
+#undef THE_SIMD_LEVEL
+
 namespace faiss {
 
 /*****************************************
@@ -40,17 +44,17 @@ namespace faiss {
  ******************************************/
 
 IndexIVFPQ::IndexIVFPQ(
-        Index* quantizer,
-        size_t d,
-        size_t nlist,
+        Index* quantizer_in,
+        size_t d_in,
+        size_t nlist_in,
         size_t M,
         size_t nbits_per_idx,
         MetricType metric,
-        bool own_invlists)
-        : IndexIVF(quantizer, d, nlist, 0, metric, own_invlists),
-          pq(d, M, nbits_per_idx) {
+        bool own_invlists_in)
+        : IndexIVF(quantizer_in, d_in, nlist_in, 0, metric, own_invlists_in),
+          pq(d_in, M, nbits_per_idx) {
     code_size = pq.code_size;
-    if (own_invlists) {
+    if (own_invlists_in) {
         invlists->code_size = code_size;
     }
     is_trained = false;
@@ -66,12 +70,16 @@ IndexIVFPQ::IndexIVFPQ(
 /****************************************************************
  * training                                                     */
 
-void IndexIVFPQ::train_encoder(idx_t n, const float* x, const idx_t* assign) {
+void IndexIVFPQ::train_encoder(
+        idx_t n,
+        const float* x,
+        const idx_t* /*assign*/) {
     pq.train(n, x);
 
     if (do_polysemous_training) {
-        if (verbose)
+        if (verbose) {
             printf("doing polysemous training for PQ\n");
+        }
         PolysemousTraining default_pt;
         PolysemousTraining* pt =
                 polysemous_training ? polysemous_training : &default_pt;
@@ -96,8 +104,9 @@ void IndexIVFPQ::encode(idx_t key, const float* x, uint8_t* code) const {
         std::vector<float> residual_vec(d);
         quantizer->compute_residual(x, residual_vec.data(), key);
         pq.compute_code(residual_vec.data(), code);
-    } else
+    } else {
         pq.compute_code(x, code);
+    }
 }
 
 void IndexIVFPQ::encode_multiple(
@@ -106,8 +115,9 @@ void IndexIVFPQ::encode_multiple(
         const float* x,
         uint8_t* xcodes,
         bool compute_keys) const {
-    if (compute_keys)
+    if (compute_keys) {
         quantizer->assign(n, x, keys);
+    }
 
     encode_vectors(n, x, keys, xcodes);
 }
@@ -123,7 +133,7 @@ void IndexIVFPQ::decode_multiple(
         for (size_t i = 0; i < n; i++) {
             quantizer->reconstruct(keys[i], centroid.data());
             float* xi = x + i * d;
-            for (size_t j = 0; j < d; j++) {
+            for (int j = 0; j < d; j++) {
                 xi[j] += centroid[j];
             }
         }
@@ -149,13 +159,15 @@ static std::unique_ptr<float[]> compute_residuals(
         const idx_t* list_nos) {
     size_t d = quantizer->d;
     std::unique_ptr<float[]> residuals(new float[n * d]);
-    // TODO: parallelize?
-    for (size_t i = 0; i < n; i++) {
-        if (list_nos[i] < 0)
+    // Parallelize with OpenMP (each iteration is independent)
+#pragma omp parallel for if (n > 1000)
+    for (idx_t i = 0; i < n; i++) {
+        if (list_nos[i] < 0) {
             memset(residuals.get() + i * d, 0, sizeof(float) * d);
-        else
+        } else {
             quantizer->compute_residual(
                     x + i * d, residuals.get() + i * d, list_nos[i]);
+        }
     }
     return residuals;
 }
@@ -207,7 +219,7 @@ void IndexIVFPQ::sa_decode(idx_t n, const uint8_t* codes, float* x) const {
             pq.decode(code + coarse_size, xi);
             if (by_residual) {
                 quantizer->reconstruct(list_no, residual.data());
-                for (size_t j = 0; j < d; j++) {
+                for (int j = 0; j < d; j++) {
                     xi[j] += residual[j];
                 }
             }
@@ -282,14 +294,15 @@ void IndexIVFPQ::add_core_o(
     double t2 = getmillisecs();
     // TODO: parallelize?
     size_t n_ignore = 0;
-    for (size_t i = 0; i < n; i++) {
+    for (idx_t i = 0; i < n; i++) {
         idx_t key = idx[i];
         idx_t id = xids ? xids[i] : ntotal + i;
         if (key < 0) {
             direct_map.add_single_id(id, -1, 0);
             n_ignore++;
-            if (residuals_2)
+            if (residuals_2) {
                 memset(residuals_2, 0, sizeof(*residuals_2) * d);
+            }
             continue;
         }
 
@@ -301,8 +314,9 @@ void IndexIVFPQ::add_core_o(
             float* res2 = residuals_2 + i * d;
             const float* xi = to_encode + i * d;
             pq.decode(code, res2);
-            for (int j = 0; j < d; j++)
+            for (int j = 0; j < d; j++) {
                 res2[j] = xi[j] - res2[j];
+            }
         }
 
         direct_map.add_single_id(id, key, offset);
@@ -311,8 +325,9 @@ void IndexIVFPQ::add_core_o(
     double t3 = getmillisecs();
     if (verbose) {
         char comment[100] = {0};
-        if (n_ignore > 0)
+        if (n_ignore > 0) {
             snprintf(comment, 100, "(%zd vectors ignored)", n_ignore);
+        }
         printf(" add_core times: %.3f %.3f %.3f %s\n",
                t1 - t0,
                t2 - t1,
@@ -379,6 +394,7 @@ void initialize_IVFPQ_precomputed_table(
         AlignedTable<float>& precomputed_table,
         bool by_residual,
         bool verbose) {
+    FAISS_THROW_IF_NOT_MSG(quantizer, "IVF quantizer must not be null");
     size_t nlist = quantizer->ntotal;
     size_t d = quantizer->d;
     FAISS_THROW_IF_NOT(d == pq.d);
@@ -388,6 +404,9 @@ void initialize_IVFPQ_precomputed_table(
         return;
     }
 
+    const size_t m_ksub =
+            mul_no_overflow(pq.M, pq.ksub, "IVFPQ precomputed_table");
+
     if (use_precomputed_table == 0) { // then choose the type of table
         if (!(quantizer->metric_type == METRIC_L2 && by_residual)) {
             if (verbose) {
@@ -399,10 +418,13 @@ void initialize_IVFPQ_precomputed_table(
         }
         const MultiIndexQuantizer* miq =
                 dynamic_cast<const MultiIndexQuantizer*>(quantizer);
-        if (miq && pq.M % miq->pq.M == 0)
+        if (miq && pq.M % miq->pq.M == 0) {
             use_precomputed_table = 2;
-        else {
-            size_t table_size = pq.M * pq.ksub * nlist * sizeof(float);
+        } else {
+            size_t table_size = mul_no_overflow(
+                    mul_no_overflow(m_ksub, nlist, "IVFPQ precomputed_table"),
+                    sizeof(float),
+                    "IVFPQ precomputed_table");
             if (table_size > precomputed_table_max_bytes) {
                 if (verbose) {
                     printf("IndexIVFPQ::precompute_table: not precomputing table, "
@@ -422,22 +444,25 @@ void initialize_IVFPQ_precomputed_table(
     }
 
     // squared norms of the PQ centroids
-    std::vector<float> r_norms(pq.M * pq.ksub, NAN);
-    for (int m = 0; m < pq.M; m++)
-        for (int j = 0; j < pq.ksub; j++)
+    std::vector<float> r_norms(m_ksub, NAN);
+    for (size_t m = 0; m < pq.M; m++) {
+        for (size_t j = 0; j < pq.ksub; j++) {
             r_norms[m * pq.ksub + j] =
                     fvec_norm_L2sqr_dispatch(pq.get_centroids(m, j), pq.dsub);
+        }
+    }
 
     if (use_precomputed_table == 1) {
-        precomputed_table.resize(nlist * pq.M * pq.ksub);
+        precomputed_table.resize(
+                mul_no_overflow(nlist, m_ksub, "IVFPQ precomputed_table"));
         std::vector<float> centroid(d);
 
         for (size_t i = 0; i < nlist; i++) {
             quantizer->reconstruct(i, centroid.data());
 
-            float* tab = &precomputed_table[i * pq.M * pq.ksub];
+            float* tab = &precomputed_table[i * m_ksub];
             pq.compute_inner_prod_table(centroid.data(), tab);
-            fvec_madd_dispatch(pq.M * pq.ksub, r_norms.data(), 2.0, tab, tab);
+            fvec_madd_dispatch(m_ksub, r_norms.data(), 2.0, tab, tab);
         }
     } else if (use_precomputed_table == 2) {
         const MultiIndexQuantizer* miq =
@@ -446,12 +471,13 @@ void initialize_IVFPQ_precomputed_table(
         const ProductQuantizer& cpq = miq->pq;
         FAISS_THROW_IF_NOT(pq.M % cpq.M == 0);
 
-        precomputed_table.resize(cpq.ksub * pq.M * pq.ksub);
+        precomputed_table.resize(
+                mul_no_overflow(cpq.ksub, m_ksub, "IVFPQ precomputed_table"));
 
         // reorder PQ centroid table
         std::vector<float> centroids(d * cpq.ksub, NAN);
 
-        for (int m = 0; m < cpq.M; m++) {
+        for (size_t m = 0; m < cpq.M; m++) {
             for (size_t i = 0; i < cpq.ksub; i++) {
                 memcpy(centroids.data() + i * d + m * cpq.dsub,
                        cpq.get_centroids(m, i),
@@ -463,8 +489,8 @@ void initialize_IVFPQ_precomputed_table(
                 cpq.ksub, centroids.data(), precomputed_table.data());
 
         for (size_t i = 0; i < cpq.ksub; i++) {
-            float* tab = &precomputed_table[i * pq.M * pq.ksub];
-            fvec_madd_dispatch(pq.M * pq.ksub, r_norms.data(), 2.0, tab, tab);
+            float* tab = &precomputed_table[i * m_ksub];
+            fvec_madd_dispatch(m_ksub, r_norms.data(), 2.0, tab, tab);
         }
     }
 }
@@ -479,812 +505,13 @@ void IndexIVFPQ::precompute_table() {
             verbose);
 }
 
-namespace {
-
-#define TIC t0 = get_cycles()
-#define TOC get_cycles() - t0
-
-/** QueryTables manages the various ways of searching an
- * IndexIVFPQ. The code contains a lot of branches, depending on:
- * - metric_type: are we computing L2 or Inner product similarity?
- * - by_residual: do we encode raw vectors or residuals?
- * - use_precomputed_table: are x_R|x_C tables precomputed?
- * - polysemous_ht: are we filtering with polysemous codes?
- */
-struct QueryTables {
-    /*****************************************************
-     * General data from the IVFPQ
-     *****************************************************/
-
-    const IndexIVFPQ& ivfpq;
-    const IVFSearchParameters* params;
-
-    // copied from IndexIVFPQ for easier access
-    int d;
-    const ProductQuantizer& pq;
-    MetricType metric_type;
-    bool by_residual;
-    int use_precomputed_table;
-    int polysemous_ht;
-
-    // pre-allocated data buffers
-    float *sim_table, *sim_table_2;
-    float *residual_vec, *decoded_vec;
-
-    // single data buffer
-    std::vector<float> mem;
-
-    // for table pointers
-    std::vector<const float*> sim_table_ptrs;
-
-    explicit QueryTables(
-            const IndexIVFPQ& ivfpq,
-            const IVFSearchParameters* params)
-            : ivfpq(ivfpq),
-              d(ivfpq.d),
-              pq(ivfpq.pq),
-              metric_type(ivfpq.metric_type),
-              by_residual(ivfpq.by_residual),
-              use_precomputed_table(ivfpq.use_precomputed_table) {
-        mem.resize(pq.ksub * pq.M * 2 + d * 2);
-        sim_table = mem.data();
-        sim_table_2 = sim_table + pq.ksub * pq.M;
-        residual_vec = sim_table_2 + pq.ksub * pq.M;
-        decoded_vec = residual_vec + d;
-
-        // for polysemous
-        polysemous_ht = ivfpq.polysemous_ht;
-        if (auto ivfpq_params =
-                    dynamic_cast<const IVFPQSearchParameters*>(params)) {
-            polysemous_ht = ivfpq_params->polysemous_ht;
-        }
-        if (polysemous_ht != 0) {
-            q_code.resize(pq.code_size);
-        }
-        init_list_cycles = 0;
-        sim_table_ptrs.resize(pq.M);
-    }
-
-    /*****************************************************
-     * What we do when query is known
-     *****************************************************/
-
-    // field specific to query
-    const float* qi;
-
-    // query-specific initialization
-    void init_query(const float* qi) {
-        this->qi = qi;
-        if (metric_type == METRIC_INNER_PRODUCT)
-            init_query_IP();
-        else
-            init_query_L2();
-        if (!by_residual && polysemous_ht != 0)
-            pq.compute_code(qi, q_code.data());
-    }
-
-    void init_query_IP() {
-        // precompute some tables specific to the query qi
-        pq.compute_inner_prod_table(qi, sim_table);
-    }
-
-    void init_query_L2() {
-        if (!by_residual) {
-            pq.compute_distance_table(qi, sim_table);
-        } else if (use_precomputed_table) {
-            pq.compute_inner_prod_table(qi, sim_table_2);
-        }
-    }
-
-    /*****************************************************
-     * When inverted list is known: prepare computations
-     *****************************************************/
-
-    // fields specific to list
-    idx_t key;
-    float coarse_dis;
-    std::vector<uint8_t> q_code;
-
-    uint64_t init_list_cycles;
-
-    /// once we know the query and the centroid, we can prepare the
-    /// sim_table that will be used for accumulation
-    /// and dis0, the initial value
-    float precompute_list_tables() {
-        float dis0 = 0;
-        uint64_t t0;
-        TIC;
-        if (by_residual) {
-            if (metric_type == METRIC_INNER_PRODUCT)
-                dis0 = precompute_list_tables_IP();
-            else
-                dis0 = precompute_list_tables_L2();
-        }
-        init_list_cycles += TOC;
-        return dis0;
-    }
-
-    float precompute_list_table_pointers() {
-        float dis0 = 0;
-        uint64_t t0;
-        TIC;
-        if (by_residual) {
-            if (metric_type == METRIC_INNER_PRODUCT)
-                FAISS_THROW_MSG("not implemented");
-            else
-                dis0 = precompute_list_table_pointers_L2();
-        }
-        init_list_cycles += TOC;
-        return dis0;
-    }
-
-    /*****************************************************
-     * compute tables for inner prod
-     *****************************************************/
-
-    float precompute_list_tables_IP() {
-        // prepare the sim_table that will be used for accumulation
-        // and dis0, the initial value
-        ivfpq.quantizer->reconstruct(key, decoded_vec);
-        // decoded_vec = centroid
-        float dis0 = fvec_inner_product_dispatch(qi, decoded_vec, d);
-
-        if (polysemous_ht) {
-            for (int i = 0; i < d; i++) {
-                residual_vec[i] = qi[i] - decoded_vec[i];
-            }
-            pq.compute_code(residual_vec, q_code.data());
-        }
-        return dis0;
-    }
-
-    /*****************************************************
-     * compute tables for L2 distance
-     *****************************************************/
-
-    float precompute_list_tables_L2() {
-        float dis0 = 0;
-
-        if (use_precomputed_table == 0 || use_precomputed_table == -1) {
-            ivfpq.quantizer->compute_residual(qi, residual_vec, key);
-            pq.compute_distance_table(residual_vec, sim_table);
-
-            if (polysemous_ht != 0) {
-                pq.compute_code(residual_vec, q_code.data());
-            }
-
-        } else if (use_precomputed_table == 1) {
-            dis0 = coarse_dis;
-
-            fvec_madd_dispatch(
-                    pq.M * pq.ksub,
-                    ivfpq.precomputed_table.data() + key * pq.ksub * pq.M,
-                    -2.0,
-                    sim_table_2,
-                    sim_table);
-
-            if (polysemous_ht != 0) {
-                ivfpq.quantizer->compute_residual(qi, residual_vec, key);
-                pq.compute_code(residual_vec, q_code.data());
-            }
-
-        } else if (use_precomputed_table == 2) {
-            dis0 = coarse_dis;
-
-            const MultiIndexQuantizer* miq =
-                    dynamic_cast<const MultiIndexQuantizer*>(ivfpq.quantizer);
-            FAISS_THROW_IF_NOT(miq);
-            const ProductQuantizer& cpq = miq->pq;
-            int Mf = pq.M / cpq.M;
-
-            const float* qtab = sim_table_2; // query-specific table
-            float* ltab = sim_table;         // (output) list-specific table
-
-            long k = key;
-            for (int cm = 0; cm < cpq.M; cm++) {
-                // compute PQ index
-                int ki = k & ((uint64_t(1) << cpq.nbits) - 1);
-                k >>= cpq.nbits;
-
-                // get corresponding table
-                const float* pc = ivfpq.precomputed_table.data() +
-                        (ki * pq.M + cm * Mf) * pq.ksub;
-
-                if (polysemous_ht == 0) {
-                    // sum up with query-specific table
-                    fvec_madd_dispatch(Mf * pq.ksub, pc, -2.0, qtab, ltab);
-                    ltab += Mf * pq.ksub;
-                    qtab += Mf * pq.ksub;
-                } else {
-                    for (int m = cm * Mf; m < (cm + 1) * Mf; m++) {
-                        q_code[m] = fvec_madd_and_argmin_dispatch(
-                                pq.ksub, pc, -2, qtab, ltab);
-                        pc += pq.ksub;
-                        ltab += pq.ksub;
-                        qtab += pq.ksub;
-                    }
-                }
-            }
-        }
-
-        return dis0;
-    }
-
-    float precompute_list_table_pointers_L2() {
-        float dis0 = 0;
-
-        if (use_precomputed_table == 1) {
-            dis0 = coarse_dis;
-
-            const float* s =
-                    ivfpq.precomputed_table.data() + key * pq.ksub * pq.M;
-            for (int m = 0; m < pq.M; m++) {
-                sim_table_ptrs[m] = s;
-                s += pq.ksub;
-            }
-        } else if (use_precomputed_table == 2) {
-            dis0 = coarse_dis;
-
-            const MultiIndexQuantizer* miq =
-                    dynamic_cast<const MultiIndexQuantizer*>(ivfpq.quantizer);
-            FAISS_THROW_IF_NOT(miq);
-            const ProductQuantizer& cpq = miq->pq;
-            int Mf = pq.M / cpq.M;
-
-            long k = key;
-            int m0 = 0;
-            for (int cm = 0; cm < cpq.M; cm++) {
-                int ki = k & ((uint64_t(1) << cpq.nbits) - 1);
-                k >>= cpq.nbits;
-
-                const float* pc = ivfpq.precomputed_table.data() +
-                        (ki * pq.M + cm * Mf) * pq.ksub;
-
-                for (int m = m0; m < m0 + Mf; m++) {
-                    sim_table_ptrs[m] = pc;
-                    pc += pq.ksub;
-                }
-                m0 += Mf;
-            }
-        } else {
-            FAISS_THROW_MSG("need precomputed tables");
-        }
-
-        if (polysemous_ht) {
-            FAISS_THROW_MSG("not implemented");
-            // Not clear that it makes sense to implemente this,
-            // because it costs M * ksub, which is what we wanted to
-            // avoid with the tables pointers.
-        }
-
-        return dis0;
-    }
-};
-
-template <class C, bool use_sel>
-struct WrappedSearchResult {
-    ResultHandler& res;
-    size_t nup = 0;
-    idx_t list_no;
-
-    const idx_t* ids;
-    const IDSelector* sel;
-
-    WrappedSearchResult(
-            idx_t list_no,
-            const idx_t* ids,
-            const IDSelector* sel,
-            ResultHandler& res)
-            : res(res), list_no(list_no), ids(ids), sel(sel) {}
-
-    inline bool skip_entry(idx_t j) {
-        return use_sel && !sel->is_member(ids[j]);
-    }
-
-    inline void add(idx_t j, float dis) {
-        if (C::cmp(res.threshold, dis)) {
-            idx_t id = ids ? ids[j] : lo_build(this->list_no, j);
-            res.add_result(dis, id);
-            nup++;
-        }
-    }
-};
-
-/*****************************************************
- * Scaning the codes.
- * The scanning functions call their favorite precompute_*
- * function to precompute the tables they need.
- *****************************************************/
-template <typename IDType, MetricType METRIC_TYPE, class PQCodeDist>
-struct IVFPQScannerT : QueryTables {
-    using PQDecoder = typename PQCodeDist::PQDecoder;
-    const uint8_t* list_codes;
-    const IDType* list_ids;
-    size_t list_size;
-
-    IVFPQScannerT(const IndexIVFPQ& ivfpq, const IVFSearchParameters* params)
-            : QueryTables(ivfpq, params) {
-        assert(METRIC_TYPE == metric_type);
-    }
-
-    float dis0;
-
-    void init_list(idx_t list_no, float coarse_dis, int mode) {
-        this->key = list_no;
-        this->coarse_dis = coarse_dis;
-
-        if (mode == 2) {
-            dis0 = precompute_list_tables();
-        } else if (mode == 1) {
-            dis0 = precompute_list_table_pointers();
-        }
-    }
-
-    /*****************************************************
-     * Scaning the codes: simple PQ scan.
-     *****************************************************/
-
-    // This is the baseline version of scan_list_with_tables().
-    // It demonstrates what this function actually does.
-    //
-    // /// version of the scan where we use precomputed tables.
-    // template <class SearchResultType>
-    // void scan_list_with_table(
-    //         size_t ncode,
-    //         const uint8_t* codes,
-    //         SearchResultType& res) const {
-    //
-    //     for (size_t j = 0; j < ncode; j++, codes += pq.code_size) {
-    //         if (res.skip_entry(j)) {
-    //             continue;
-    //         }
-    //         float dis = dis0 + PQCodeDist::distance_single_code(
-    //             pq, sim_table, codes);
-    //         res.add(j, dis);
-    //     }
-    // }
-
-    // This is the modified version of scan_list_with_tables().
-    // It was observed that doing manual unrolling of the loop that
-    //    utilizes distance_single_code() speeds up the computations.
-
-    /// version of the scan where we use precomputed tables.
-    template <class SearchResultType>
-    void scan_list_with_table(
-            size_t ncode,
-            const uint8_t* codes,
-            SearchResultType& res) const {
-        int counter = 0;
-
-        size_t saved_j[4] = {0, 0, 0, 0};
-        for (size_t j = 0; j < ncode; j++) {
-            if (res.skip_entry(j)) {
-                continue;
-            }
-
-            saved_j[0] = (counter == 0) ? j : saved_j[0];
-            saved_j[1] = (counter == 1) ? j : saved_j[1];
-            saved_j[2] = (counter == 2) ? j : saved_j[2];
-            saved_j[3] = (counter == 3) ? j : saved_j[3];
-
-            counter += 1;
-            if (counter == 4) {
-                float distance_0 = 0;
-                float distance_1 = 0;
-                float distance_2 = 0;
-                float distance_3 = 0;
-                PQCodeDist::distance_four_codes(
-                        pq.M,
-                        pq.nbits,
-                        sim_table,
-                        codes + saved_j[0] * pq.code_size,
-                        codes + saved_j[1] * pq.code_size,
-                        codes + saved_j[2] * pq.code_size,
-                        codes + saved_j[3] * pq.code_size,
-                        distance_0,
-                        distance_1,
-                        distance_2,
-                        distance_3);
-
-                res.add(saved_j[0], dis0 + distance_0);
-                res.add(saved_j[1], dis0 + distance_1);
-                res.add(saved_j[2], dis0 + distance_2);
-                res.add(saved_j[3], dis0 + distance_3);
-                counter = 0;
-            }
-        }
-
-        if (counter >= 1) {
-            float dis = dis0 +
-                    PQCodeDist::distance_single_code(
-                                pq.M,
-                                pq.nbits,
-                                sim_table,
-                                codes + saved_j[0] * pq.code_size);
-            res.add(saved_j[0], dis);
-        }
-        if (counter >= 2) {
-            float dis = dis0 +
-                    PQCodeDist::distance_single_code(
-                                pq.M,
-                                pq.nbits,
-                                sim_table,
-                                codes + saved_j[1] * pq.code_size);
-            res.add(saved_j[1], dis);
-        }
-        if (counter >= 3) {
-            float dis = dis0 +
-                    PQCodeDist::distance_single_code(
-                                pq.M,
-                                pq.nbits,
-                                sim_table,
-                                codes + saved_j[2] * pq.code_size);
-            res.add(saved_j[2], dis);
-        }
-    }
-
-    /// tables are not precomputed, but pointers are provided to the
-    /// relevant X_c|x_r tables
-    template <class SearchResultType>
-    void scan_list_with_pointer(
-            size_t ncode,
-            const uint8_t* codes,
-            SearchResultType& res) const {
-        for (size_t j = 0; j < ncode; j++, codes += pq.code_size) {
-            if (res.skip_entry(j)) {
-                continue;
-            }
-            PQDecoder decoder(codes, pq.nbits);
-            float dis = dis0;
-            const float* tab = sim_table_2;
-
-            for (size_t m = 0; m < pq.M; m++) {
-                int ci = decoder.decode();
-                dis += sim_table_ptrs[m][ci] - 2 * tab[ci];
-                tab += pq.ksub;
-            }
-            res.add(j, dis);
-        }
-    }
-
-    /// nothing is precomputed: access residuals on-the-fly
-    template <class SearchResultType>
-    void scan_on_the_fly_dist(
-            size_t ncode,
-            const uint8_t* codes,
-            SearchResultType& res) const {
-        const float* dvec;
-        float dis0 = 0;
-        if (by_residual) {
-            if (METRIC_TYPE == METRIC_INNER_PRODUCT) {
-                ivfpq.quantizer->reconstruct(key, residual_vec);
-                dis0 = fvec_inner_product_dispatch(residual_vec, qi, d);
-            } else {
-                ivfpq.quantizer->compute_residual(qi, residual_vec, key);
-            }
-            dvec = residual_vec;
-        } else {
-            dvec = qi;
-            dis0 = 0;
-        }
-
-        for (size_t j = 0; j < ncode; j++, codes += pq.code_size) {
-            if (res.skip_entry(j)) {
-                continue;
-            }
-            pq.decode(codes, decoded_vec);
-
-            float dis;
-            if (METRIC_TYPE == METRIC_INNER_PRODUCT) {
-                dis = dis0 + fvec_inner_product_dispatch(decoded_vec, qi, d);
-            } else {
-                dis = fvec_L2sqr_dispatch(decoded_vec, dvec, d);
-            }
-            res.add(j, dis);
-        }
-    }
-
-    /*****************************************************
-     * Scanning codes with polysemous filtering
-     *****************************************************/
-
-    // This is the baseline version of scan_list_polysemous_hc().
-    // It demonstrates what this function actually does.
-
-    //     template <class HammingComputer, class SearchResultType>
-    //     void scan_list_polysemous_hc(
-    //             size_t ncode,
-    //             const uint8_t* codes,
-    //             SearchResultType& res) const {
-    //         int ht = ivfpq.polysemous_ht;
-    //         size_t n_hamming_pass = 0, nup = 0;
-    //
-    //         int code_size = pq.code_size;
-    //
-    //         HammingComputer hc(q_code.data(), code_size);
-    //
-    //         for (size_t j = 0; j < ncode; j++, codes += code_size) {
-    //             if (res.skip_entry(j)) {
-    //                 continue;
-    //             }
-    //             const uint8_t* b_code = codes;
-    //             int hd = hc.hamming(b_code);
-    //             if (hd < ht) {
-    //                 n_hamming_pass++;
-    //
-    //                 float dis =
-    //                         dis0 +
-    //                         PQCodeDist::distance_single_code(
-    //                             pq, sim_table, codes);
-    //
-    //                 res.add(j, dis);
-    //             }
-    //         }
-    // #pragma omp critical
-    //         { indexIVFPQ_stats.n_hamming_pass += n_hamming_pass; }
-    //     }
-
-    // This is the modified version of scan_list_with_tables().
-    // It was observed that doing manual unrolling of the loop that
-    //    utilizes distance_single_code() speeds up the computations.
-
-    template <class HammingComputer, class SearchResultType>
-    void scan_list_polysemous_hc(
-            size_t ncode,
-            const uint8_t* codes,
-            SearchResultType& res) const {
-        int ht = ivfpq.polysemous_ht;
-        size_t n_hamming_pass = 0;
-
-        int code_size = pq.code_size;
-
-        size_t saved_j[8];
-        int counter = 0;
-
-        HammingComputer hc(q_code.data(), code_size);
-
-        for (size_t j = 0; j < (ncode / 4) * 4; j += 4) {
-            const uint8_t* b_code = codes + j * code_size;
-
-            // Unrolling is a key. Basically, doing multiple popcount
-            // operations one after another speeds things up.
-
-            // 9999999 is just an arbitrary large number
-            int hd0 = (res.skip_entry(j + 0))
-                    ? 99999999
-                    : hc.hamming(b_code + 0 * code_size);
-            int hd1 = (res.skip_entry(j + 1))
-                    ? 99999999
-                    : hc.hamming(b_code + 1 * code_size);
-            int hd2 = (res.skip_entry(j + 2))
-                    ? 99999999
-                    : hc.hamming(b_code + 2 * code_size);
-            int hd3 = (res.skip_entry(j + 3))
-                    ? 99999999
-                    : hc.hamming(b_code + 3 * code_size);
-
-            saved_j[counter] = j + 0;
-            counter = (hd0 < ht) ? (counter + 1) : counter;
-            saved_j[counter] = j + 1;
-            counter = (hd1 < ht) ? (counter + 1) : counter;
-            saved_j[counter] = j + 2;
-            counter = (hd2 < ht) ? (counter + 1) : counter;
-            saved_j[counter] = j + 3;
-            counter = (hd3 < ht) ? (counter + 1) : counter;
-
-            if (counter >= 4) {
-                // process four codes at the same time
-                n_hamming_pass += 4;
-
-                float distance_0 = dis0;
-                float distance_1 = dis0;
-                float distance_2 = dis0;
-                float distance_3 = dis0;
-                PQCodeDist::distance_four_codes(
-                        pq.M,
-                        pq.nbits,
-                        sim_table,
-                        codes + saved_j[0] * pq.code_size,
-                        codes + saved_j[1] * pq.code_size,
-                        codes + saved_j[2] * pq.code_size,
-                        codes + saved_j[3] * pq.code_size,
-                        distance_0,
-                        distance_1,
-                        distance_2,
-                        distance_3);
-
-                res.add(saved_j[0], dis0 + distance_0);
-                res.add(saved_j[1], dis0 + distance_1);
-                res.add(saved_j[2], dis0 + distance_2);
-                res.add(saved_j[3], dis0 + distance_3);
-
-                //
-                counter -= 4;
-                saved_j[0] = saved_j[4];
-                saved_j[1] = saved_j[5];
-                saved_j[2] = saved_j[6];
-                saved_j[3] = saved_j[7];
-            }
-        }
-
-        for (size_t kk = 0; kk < counter; kk++) {
-            n_hamming_pass++;
-
-            float dis = dis0 +
-                    PQCodeDist::distance_single_code(
-                                pq.M,
-                                pq.nbits,
-                                sim_table,
-                                codes + saved_j[kk] * pq.code_size);
-
-            res.add(saved_j[kk], dis);
-        }
-
-        // process leftovers
-        for (size_t j = (ncode / 4) * 4; j < ncode; j++) {
-            if (res.skip_entry(j)) {
-                continue;
-            }
-            const uint8_t* b_code = codes + j * code_size;
-            int hd = hc.hamming(b_code);
-            if (hd < ht) {
-                n_hamming_pass++;
-
-                float dis = dis0 +
-                        PQCodeDist::distance_single_code(
-                                    pq.M,
-                                    pq.nbits,
-                                    sim_table,
-                                    codes + j * code_size);
-
-                res.add(j, dis);
-            }
-        }
-
-#pragma omp critical
-        {
-            indexIVFPQ_stats.n_hamming_pass += n_hamming_pass;
-        }
-    }
-
-    template <class SearchResultType>
-    struct Run_scan_list_polysemous_hc {
-        using T = void;
-        template <class HammingComputer, class... Types>
-        void f(const IVFPQScannerT* scanner, Types... args) {
-            scanner->scan_list_polysemous_hc<HammingComputer, SearchResultType>(
-                    args...);
-        }
-    };
-
-    template <class SearchResultType>
-    void scan_list_polysemous(
-            size_t ncode,
-            const uint8_t* codes,
-            SearchResultType& res) const {
-        Run_scan_list_polysemous_hc<SearchResultType> r;
-        dispatch_HammingComputer(pq.code_size, r, this, ncode, codes, res);
-    }
-};
-
-/* We put as many parameters as possible in template. Hopefully the
- * gain in runtime is worth the code bloat.
- *
- * C is the comparator < or >, it is directly related to METRIC_TYPE.
- *
- * precompute_mode is how much we precompute (2 = precompute distance tables,
- * 1 = precompute pointers to distances, 0 = compute distances one by one).
- * Currently only 2 is supported
- *
- * use_sel: store or ignore the IDSelector
- */
-template <MetricType METRIC_TYPE, class C, class PQCodeDist, bool use_sel>
-struct IVFPQScanner : IVFPQScannerT<idx_t, METRIC_TYPE, PQCodeDist>,
-                      InvertedListScanner {
-    int precompute_mode;
-    const IDSelector* sel;
-
-    IVFPQScanner(
-            const IndexIVFPQ& ivfpq,
-            bool store_pairs,
-            int precompute_mode,
-            const IDSelector* sel)
-            : IVFPQScannerT<idx_t, METRIC_TYPE, PQCodeDist>(ivfpq, nullptr),
-              precompute_mode(precompute_mode),
-              sel(sel) {
-        this->store_pairs = store_pairs;
-        this->keep_max = is_similarity_metric(METRIC_TYPE);
-        this->code_size = this->pq.code_size;
-    }
-
-    void set_query(const float* query) override {
-        this->init_query(query);
-    }
-
-    void set_list(idx_t list_no, float coarse_dis) override {
-        this->list_no = list_no;
-        this->init_list(list_no, coarse_dis, precompute_mode);
-    }
-
-    float distance_to_code(const uint8_t* code) const override {
-        assert(precompute_mode == 2);
-        float dis = this->dis0 +
-                PQCodeDist::distance_single_code(
-                            this->pq.M, this->pq.nbits, this->sim_table, code);
-        return dis;
-    }
-
-    size_t scan_codes(
-            size_t ncode,
-            const uint8_t* codes,
-            const idx_t* ids,
-            ResultHandler& handler) const override {
-        WrappedSearchResult<C, use_sel> res(
-                this->key,
-                this->store_pairs ? nullptr : ids,
-                this->sel,
-                handler);
-
-        if (this->polysemous_ht > 0) {
-            assert(precompute_mode == 2);
-            this->scan_list_polysemous(ncode, codes, res);
-        } else if (precompute_mode == 2) {
-            this->scan_list_with_table(ncode, codes, res);
-        } else if (precompute_mode == 1) {
-            this->scan_list_with_pointer(ncode, codes, res);
-        } else if (precompute_mode == 0) {
-            this->scan_on_the_fly_dist(ncode, codes, res);
-        } else {
-            FAISS_THROW_MSG("bad precomp mode");
-        }
-        return res.nup;
-    }
-};
-
-} // anonymous namespace
-
 InvertedListScanner* IndexIVFPQ::get_InvertedListScanner(
         bool store_pairs,
         const IDSelector* sel,
         const IVFSearchParameters*) const {
     return with_simd_level([&]<SIMDLevel SL>() -> InvertedListScanner* {
-        auto make =
-                [&]<class PQCodeDist, bool use_sel>() -> InvertedListScanner* {
-            if (metric_type == METRIC_INNER_PRODUCT) {
-                return new IVFPQScanner<
-                        METRIC_INNER_PRODUCT,
-                        CMin<float, idx_t>,
-                        PQCodeDist,
-                        use_sel>(*this, store_pairs, 2, sel);
-            } else if (metric_type == METRIC_L2) {
-                return new IVFPQScanner<
-                        METRIC_L2,
-                        CMax<float, idx_t>,
-                        PQCodeDist,
-                        use_sel>(*this, store_pairs, 2, sel);
-            } else {
-                FAISS_THROW_MSG("unsupported metric type");
-            }
-        };
-
-        auto with_decoder = [&]<bool use_sel>() -> InvertedListScanner* {
-            if (pq.nbits == 8) {
-                return make.template
-                operator()<PQCodeDistance<PQDecoder8, SL>, use_sel>();
-            } else if (pq.nbits == 16) {
-                return make.template
-                operator()<PQCodeDistance<PQDecoder16, SL>, use_sel>();
-            } else {
-                return make.template
-                operator()<PQCodeDistance<PQDecoderGeneric, SL>, use_sel>();
-            }
-        };
-
-        if (sel) {
-            return with_decoder.template operator()<true>();
-        } else {
-            return with_decoder.template operator()<false>();
-        }
+        return pq_code_distance::make_IVFPQInvertedListScanner<SL>(
+                *this, store_pairs, sel);
     });
 }
 
@@ -1320,25 +547,26 @@ size_t IndexIVFPQ::find_duplicates(idx_t* dup_ids, size_t* lims) const {
     for (size_t list_no = 0; list_no < nlist; list_no++) {
         size_t n = invlists->list_size(list_no);
         std::vector<int> ord(n);
-        for (int i = 0; i < n; i++)
-            ord[i] = i;
+        for (size_t i = 0; i < n; i++) {
+            ord[i] = static_cast<int>(i);
+        }
         InvertedLists::ScopedCodes codes(invlists, list_no);
         CodeCmp cs = {codes.get(), code_size};
         std::sort(ord.begin(), ord.end(), cs);
 
         InvertedLists::ScopedIds list_ids(invlists, list_no);
         int prev = -1; // all elements from prev to i-1 are equal
-        for (int i = 0; i < n; i++) {
+        for (size_t i = 0; i < n; i++) {
             if (prev >= 0 && cs.cmp(ord[prev], ord[i]) == 0) {
                 // same as previous => remember
-                if (prev + 1 == i) { // start new group
+                if (static_cast<size_t>(prev + 1) == i) { // start new group
                     ngroup++;
                     lims[ngroup] = lims[ngroup - 1];
                     dup_ids[lims[ngroup]++] = list_ids[ord[prev]];
                 }
                 dup_ids[lims[ngroup]++] = list_ids[ord[i]];
             } else { // not same as previous.
-                prev = i;
+                prev = static_cast<int>(i);
             }
         }
     }