faiss 0.2.7 → 0.3.1

data/vendor/faiss/faiss/impl/index_read.cpp

@@ -33,6 +33,7 @@
 #include <faiss/IndexIVFAdditiveQuantizer.h>
 #include <faiss/IndexIVFAdditiveQuantizerFastScan.h>
 #include <faiss/IndexIVFFlat.h>
+#include <faiss/IndexIVFIndependentQuantizer.h>
 #include <faiss/IndexIVFPQ.h>
 #include <faiss/IndexIVFPQFastScan.h>
 #include <faiss/IndexIVFPQR.h>
@@ -291,11 +292,17 @@ static void read_AdditiveQuantizer(AdditiveQuantizer* aq, IOReader* f) {
     aq->set_derived_values();
 }
 
-static void read_ResidualQuantizer(ResidualQuantizer* rq, IOReader* f) {
+static void read_ResidualQuantizer(
+        ResidualQuantizer* rq,
+        IOReader* f,
+        int io_flags) {
     read_AdditiveQuantizer(rq, f);
     READ1(rq->train_type);
     READ1(rq->max_beam_size);
-    if (!(rq->train_type & ResidualQuantizer::Skip_codebook_tables)) {
+    if ((rq->train_type & ResidualQuantizer::Skip_codebook_tables) ||
+        (io_flags & IO_FLAG_SKIP_PRECOMPUTE_TABLE)) {
+        // don't precompute the tables
+    } else {
         rq->compute_codebook_tables();
     }
 }
@@ -324,12 +331,13 @@ static void read_ProductAdditiveQuantizer(
 
 static void read_ProductResidualQuantizer(
         ProductResidualQuantizer* prq,
-        IOReader* f) {
+        IOReader* f,
+        int io_flags) {
     read_ProductAdditiveQuantizer(prq, f);
 
     for (size_t i = 0; i < prq->nsplits; i++) {
         auto rq = new ResidualQuantizer();
-        read_ResidualQuantizer(rq, f);
+        read_ResidualQuantizer(rq, f, io_flags);
         prq->quantizers.push_back(rq);
     }
 }
@@ -390,15 +398,12 @@ static void read_NSG(NSG* nsg, IOReader* f) {
     graph = std::make_shared<nsg::Graph<int>>(N, R);
     std::fill_n(graph->data, N * R, EMPTY_ID);
 
-    int size = 0;
-
     for (int i = 0; i < N; i++) {
         for (int j = 0; j < R + 1; j++) {
             int id;
             READ1(id);
             if (id != EMPTY_ID) {
                 graph->at(i, j) = id;
-                size += 1;
             } else {
                 break;
             }
@@ -428,7 +433,7 @@ ProductQuantizer* read_ProductQuantizer(const char* fname) {
 
 ProductQuantizer* read_ProductQuantizer(IOReader* reader) {
     ProductQuantizer* pq = new ProductQuantizer();
-    ScopeDeleter1<ProductQuantizer> del(pq);
+    std::unique_ptr<ProductQuantizer> del(pq);
 
     read_ProductQuantizer(pq, reader);
     del.release();
@@ -587,7 +592,7 @@ Index* read_index(IOReader* f, int io_flags) {
             READ1(idxp->encode_signs);
             READ1(idxp->polysemous_ht);
         }
-        // Old versoins of PQ all had metric_type set to INNER_PRODUCT
+        // Old versions of PQ all had metric_type set to INNER_PRODUCT
         // when they were in fact using L2. Therefore, we force metric type
         // to L2 when the old format is detected
         if (h == fourcc("IxPQ") || h == fourcc("IxPo")) {
@@ -600,7 +605,7 @@ Index* read_index(IOReader* f, int io_flags) {
         if (h == fourcc("IxRQ")) {
             read_ResidualQuantizer_old(&idxr->rq, f);
         } else {
-            read_ResidualQuantizer(&idxr->rq, f);
+            read_ResidualQuantizer(&idxr->rq, f, io_flags);
         }
         READ1(idxr->code_size);
         READVECTOR(idxr->codes);
@@ -615,7 +620,7 @@ Index* read_index(IOReader* f, int io_flags) {
     } else if (h == fourcc("IxPR")) {
         auto idxpr = new IndexProductResidualQuantizer();
         read_index_header(idxpr, f);
-        read_ProductResidualQuantizer(&idxpr->prq, f);
+        read_ProductResidualQuantizer(&idxpr->prq, f, io_flags);
         READ1(idxpr->code_size);
         READVECTOR(idxpr->codes);
         idx = idxpr;
@@ -629,8 +634,13 @@ Index* read_index(IOReader* f, int io_flags) {
     } else if (h == fourcc("ImRQ")) {
         ResidualCoarseQuantizer* idxr = new ResidualCoarseQuantizer();
         read_index_header(idxr, f);
-        read_ResidualQuantizer(&idxr->rq, f);
+        read_ResidualQuantizer(&idxr->rq, f, io_flags);
         READ1(idxr->beam_factor);
+        if (io_flags & IO_FLAG_SKIP_PRECOMPUTE_TABLE) {
+            // then we force the beam factor to -1
+            // which skips the table precomputation.
+            idxr->beam_factor = -1;
+        }
         idxr->set_beam_factor(idxr->beam_factor);
         idx = idxr;
     } else if (
@@ -655,13 +665,14 @@ Index* read_index(IOReader* f, int io_flags) {
         if (is_LSQ) {
             read_LocalSearchQuantizer((LocalSearchQuantizer*)idxaqfs->aq, f);
         } else if (is_RQ) {
-            read_ResidualQuantizer((ResidualQuantizer*)idxaqfs->aq, f);
+            read_ResidualQuantizer(
+                    (ResidualQuantizer*)idxaqfs->aq, f, io_flags);
         } else if (is_PLSQ) {
             read_ProductLocalSearchQuantizer(
                     (ProductLocalSearchQuantizer*)idxaqfs->aq, f);
         } else {
             read_ProductResidualQuantizer(
-                    (ProductResidualQuantizer*)idxaqfs->aq, f);
+                    (ProductResidualQuantizer*)idxaqfs->aq, f, io_flags);
         }
 
         READ1(idxaqfs->implem);
@@ -703,13 +714,13 @@ Index* read_index(IOReader* f, int io_flags) {
         if (is_LSQ) {
             read_LocalSearchQuantizer((LocalSearchQuantizer*)ivaqfs->aq, f);
         } else if (is_RQ) {
-            read_ResidualQuantizer((ResidualQuantizer*)ivaqfs->aq, f);
+            read_ResidualQuantizer((ResidualQuantizer*)ivaqfs->aq, f, io_flags);
         } else if (is_PLSQ) {
             read_ProductLocalSearchQuantizer(
                     (ProductLocalSearchQuantizer*)ivaqfs->aq, f);
         } else {
             read_ProductResidualQuantizer(
-                    (ProductResidualQuantizer*)ivaqfs->aq, f);
+                    (ProductResidualQuantizer*)ivaqfs->aq, f, io_flags);
         }
 
         READ1(ivaqfs->by_residual);
@@ -831,13 +842,13 @@ Index* read_index(IOReader* f, int io_flags) {
         if (is_LSQ) {
             read_LocalSearchQuantizer((LocalSearchQuantizer*)iva->aq, f);
         } else if (is_RQ) {
-            read_ResidualQuantizer((ResidualQuantizer*)iva->aq, f);
+            read_ResidualQuantizer((ResidualQuantizer*)iva->aq, f, io_flags);
         } else if (is_PLSQ) {
             read_ProductLocalSearchQuantizer(
                     (ProductLocalSearchQuantizer*)iva->aq, f);
         } else {
             read_ProductResidualQuantizer(
-                    (ProductResidualQuantizer*)iva->aq, f);
+                    (ProductResidualQuantizer*)iva->aq, f, io_flags);
         }
         READ1(iva->by_residual);
         READ1(iva->use_precomputed_table);
@@ -860,7 +871,22 @@ Index* read_index(IOReader* f, int io_flags) {
             h == fourcc("IvPQ") || h == fourcc("IvQR") || h == fourcc("IwPQ") ||
             h == fourcc("IwQR")) {
         idx = read_ivfpq(f, h, io_flags);
-
+    } else if (h == fourcc("IwIQ")) {
+        auto* indep = new IndexIVFIndependentQuantizer();
+        indep->own_fields = true;
+        read_index_header(indep, f);
+        indep->quantizer = read_index(f, io_flags);
+        bool has_vt;
+        READ1(has_vt);
+        if (has_vt) {
+            indep->vt = read_VectorTransform(f);
+        }
+        indep->index_ivf = dynamic_cast<IndexIVF*>(read_index(f, io_flags));
+        FAISS_THROW_IF_NOT(indep->index_ivf);
+        if (auto index_ivfpq = dynamic_cast<IndexIVFPQ*>(indep->index_ivf)) {
+            READ1(index_ivfpq->use_precomputed_table);
+        }
+        idx = indep;
     } else if (h == fourcc("IxPT")) {
         IndexPreTransform* ixpt = new IndexPreTransform();
         ixpt->own_fields = true;

data/vendor/faiss/faiss/impl/index_write.cpp

@@ -34,6 +34,7 @@
 #include <faiss/IndexIVFAdditiveQuantizer.h>
 #include <faiss/IndexIVFAdditiveQuantizerFastScan.h>
 #include <faiss/IndexIVFFlat.h>
+#include <faiss/IndexIVFIndependentQuantizer.h>
 #include <faiss/IndexIVFPQ.h>
 #include <faiss/IndexIVFPQFastScan.h>
 #include <faiss/IndexIVFPQR.h>
@@ -337,13 +338,11 @@ static void write_NSG(const NSG* nsg, IOWriter* f) {
     FAISS_THROW_IF_NOT(K == nsg->R);
     FAISS_THROW_IF_NOT(true == graph->own_fields);
 
-    int size = 0;
     for (int i = 0; i < N; i++) {
         for (int j = 0; j < K; j++) {
             int id = graph->at(i, j);
             if (id != EMPTY_ID) {
                 WRITE1(id);
-                size += 1;
             } else {
                 break;
             }
@@ -385,6 +384,8 @@ static void write_ivf_header(const IndexIVF* ivf, IOWriter* f) {
     write_index_header(ivf, f);
     WRITE1(ivf->nlist);
     WRITE1(ivf->nprobe);
+    // subclasses write by_residual (some of them support only one setting of
+    // by_residual).
     write_index(ivf->quantizer, f);
     write_direct_map(&ivf->direct_map, f);
 }
@@ -430,13 +431,14 @@ void write_index(const Index* idx, IOWriter* f) {
         WRITE1(idxr->code_size);
         WRITEVECTOR(idxr->codes);
     } else if (
-            auto* idxr = dynamic_cast<const IndexLocalSearchQuantizer*>(idx)) {
+            auto* idxr_2 =
+                    dynamic_cast<const IndexLocalSearchQuantizer*>(idx)) {
         uint32_t h = fourcc("IxLS");
         WRITE1(h);
         write_index_header(idx, f);
-        write_LocalSearchQuantizer(&idxr->lsq, f);
-        WRITE1(idxr->code_size);
-        WRITEVECTOR(idxr->codes);
+        write_LocalSearchQuantizer(&idxr_2->lsq, f);
+        WRITE1(idxr_2->code_size);
+        WRITEVECTOR(idxr_2->codes);
     } else if (
             const IndexProductResidualQuantizer* idxpr =
                     dynamic_cast<const IndexProductResidualQuantizer*>(idx)) {
@@ -569,26 +571,26 @@ void write_index(const Index* idx, IOWriter* f) {
 
         write_InvertedLists(ivaqfs->invlists, f);
     } else if (
-            const ResidualCoarseQuantizer* idxr =
+            const ResidualCoarseQuantizer* idxr_2 =
                     dynamic_cast<const ResidualCoarseQuantizer*>(idx)) {
         uint32_t h = fourcc("ImRQ");
         WRITE1(h);
         write_index_header(idx, f);
-        write_ResidualQuantizer(&idxr->rq, f);
-        WRITE1(idxr->beam_factor);
+        write_ResidualQuantizer(&idxr_2->rq, f);
+        WRITE1(idxr_2->beam_factor);
     } else if (
-            const Index2Layer* idxp = dynamic_cast<const Index2Layer*>(idx)) {
+            const Index2Layer* idxp_2 = dynamic_cast<const Index2Layer*>(idx)) {
         uint32_t h = fourcc("Ix2L");
         WRITE1(h);
         write_index_header(idx, f);
-        write_index(idxp->q1.quantizer, f);
-        WRITE1(idxp->q1.nlist);
-        WRITE1(idxp->q1.quantizer_trains_alone);
-        write_ProductQuantizer(&idxp->pq, f);
-        WRITE1(idxp->code_size_1);
-        WRITE1(idxp->code_size_2);
-        WRITE1(idxp->code_size);
-        WRITEVECTOR(idxp->codes);
+        write_index(idxp_2->q1.quantizer, f);
+        WRITE1(idxp_2->q1.nlist);
+        WRITE1(idxp_2->q1.quantizer_trains_alone);
+        write_ProductQuantizer(&idxp_2->pq, f);
+        WRITE1(idxp_2->code_size_1);
+        WRITE1(idxp_2->code_size_2);
+        WRITE1(idxp_2->code_size);
+        WRITEVECTOR(idxp_2->codes);
     } else if (
             const IndexScalarQuantizer* idxs =
                     dynamic_cast<const IndexScalarQuantizer*>(idx)) {
@@ -598,15 +600,16 @@ void write_index(const Index* idx, IOWriter* f) {
         write_ScalarQuantizer(&idxs->sq, f);
         WRITEVECTOR(idxs->codes);
     } else if (
-            const IndexLattice* idxl = dynamic_cast<const IndexLattice*>(idx)) {
+            const IndexLattice* idxl_2 =
+                    dynamic_cast<const IndexLattice*>(idx)) {
         uint32_t h = fourcc("IxLa");
         WRITE1(h);
-        WRITE1(idxl->d);
-        WRITE1(idxl->nsq);
-        WRITE1(idxl->scale_nbit);
-        WRITE1(idxl->zn_sphere_codec.r2);
+        WRITE1(idxl_2->d);
+        WRITE1(idxl_2->nsq);
+        WRITE1(idxl_2->scale_nbit);
+        WRITE1(idxl_2->zn_sphere_codec.r2);
         write_index_header(idx, f);
-        WRITEVECTOR(idxl->trained);
+        WRITEVECTOR(idxl_2->trained);
     } else if (
             const IndexIVFFlatDedup* ivfl =
                     dynamic_cast<const IndexIVFFlatDedup*>(idx)) {
@@ -625,11 +628,12 @@ void write_index(const Index* idx, IOWriter* f) {
         }
         write_InvertedLists(ivfl->invlists, f);
     } else if (
-            const IndexIVFFlat* ivfl = dynamic_cast<const IndexIVFFlat*>(idx)) {
+            const IndexIVFFlat* ivfl_2 =
+                    dynamic_cast<const IndexIVFFlat*>(idx)) {
         uint32_t h = fourcc("IwFl");
         WRITE1(h);
-        write_ivf_header(ivfl, f);
-        write_InvertedLists(ivfl->invlists, f);
+        write_ivf_header(ivfl_2, f);
+        write_InvertedLists(ivfl_2->invlists, f);
     } else if (
             const IndexIVFScalarQuantizer* ivsc =
                     dynamic_cast<const IndexIVFScalarQuantizer*>(idx)) {
@@ -700,7 +704,22 @@ void write_index(const Index* idx, IOWriter* f) {
             WRITEVECTOR(ivfpqr->refine_codes);
             WRITE1(ivfpqr->k_factor);
         }
-
+    } else if (
+            auto* indep =
+                    dynamic_cast<const IndexIVFIndependentQuantizer*>(idx)) {
+        uint32_t h = fourcc("IwIQ");
+        WRITE1(h);
+        write_index_header(indep, f);
+        write_index(indep->quantizer, f);
+        bool has_vt = indep->vt != nullptr;
+        WRITE1(has_vt);
+        if (has_vt) {
+            write_VectorTransform(indep->vt, f);
+        }
+        write_index(indep->index_ivf, f);
+        if (auto index_ivfpq = dynamic_cast<IndexIVFPQ*>(indep->index_ivf)) {
+            WRITE1(index_ivfpq->use_precomputed_table);
+        }
     } else if (
             const IndexPreTransform* ixpt =
                     dynamic_cast<const IndexPreTransform*>(idx)) {
@@ -788,19 +807,19 @@ void write_index(const Index* idx, IOWriter* f) {
         WRITE1(idxpqfs->M2);
         WRITEVECTOR(idxpqfs->codes);
     } else if (
-            const IndexIVFPQFastScan* ivpq =
+            const IndexIVFPQFastScan* ivpq_2 =
                     dynamic_cast<const IndexIVFPQFastScan*>(idx)) {
         uint32_t h = fourcc("IwPf");
         WRITE1(h);
-        write_ivf_header(ivpq, f);
-        WRITE1(ivpq->by_residual);
-        WRITE1(ivpq->code_size);
-        WRITE1(ivpq->bbs);
-        WRITE1(ivpq->M2);
-        WRITE1(ivpq->implem);
-        WRITE1(ivpq->qbs2);
-        write_ProductQuantizer(&ivpq->pq, f);
-        write_InvertedLists(ivpq->invlists, f);
+        write_ivf_header(ivpq_2, f);
+        WRITE1(ivpq_2->by_residual);
+        WRITE1(ivpq_2->code_size);
+        WRITE1(ivpq_2->bbs);
+        WRITE1(ivpq_2->M2);
+        WRITE1(ivpq_2->implem);
+        WRITE1(ivpq_2->qbs2);
+        write_ProductQuantizer(&ivpq_2->pq, f);
+        write_InvertedLists(ivpq_2->invlists, f);
     } else if (
             const IndexRowwiseMinMax* imm =
                     dynamic_cast<const IndexRowwiseMinMax*>(idx)) {
@@ -810,13 +829,13 @@ void write_index(const Index* idx, IOWriter* f) {
         write_index_header(imm, f);
         write_index(imm->index, f);
     } else if (
-            const IndexRowwiseMinMaxFP16* imm =
+            const IndexRowwiseMinMaxFP16* imm_2 =
                     dynamic_cast<const IndexRowwiseMinMaxFP16*>(idx)) {
         // IndexRowwiseMinmaxHalf
         uint32_t h = fourcc("IRMh");
         WRITE1(h);
-        write_index_header(imm, f);
-        write_index(imm->index, f);
+        write_index_header(imm_2, f);
+        write_index(imm_2->index, f);
     } else {
         FAISS_THROW_MSG("don't know how to serialize this type of index");
     }

data/vendor/faiss/faiss/impl/io.cpp

@@ -196,13 +196,13 @@ size_t BufferedIOWriter::operator()(
     while (size > 0) {
         assert(b0 == bsz);
         // now we need to flush to add more bytes
-        size_t ofs = 0;
+        size_t ofs_2 = 0;
         do {
-            assert(ofs < 10000000);
-            size_t written = (*writer)(buffer.data() + ofs, 1, bsz - ofs);
+            assert(ofs_2 < 10000000);
+            size_t written = (*writer)(buffer.data() + ofs_2, 1, bsz - ofs_2);
             FAISS_THROW_IF_NOT(written > 0);
-            ofs += written;
-        } while (ofs != bsz);
+            ofs_2 += written;
+        } while (ofs_2 != bsz);
 
         // copy src to buffer
         size_t nb1 = std::min(bsz, size);
@@ -217,12 +217,12 @@ size_t BufferedIOWriter::operator()(
 }
 
 BufferedIOWriter::~BufferedIOWriter() {
-    size_t ofs = 0;
-    while (ofs != b0) {
-        // printf("Destructor write %zd \n", b0 - ofs);
-        size_t written = (*writer)(buffer.data() + ofs, 1, b0 - ofs);
+    size_t ofs_2 = 0;
+    while (ofs_2 != b0) {
+        // printf("Destructor write %zd \n", b0 - ofs_2);
+        size_t written = (*writer)(buffer.data() + ofs_2, 1, b0 - ofs_2);
         FAISS_THROW_IF_NOT(written > 0);
-        ofs += written;
+        ofs_2 += written;
     }
 }
 

data/vendor/faiss/faiss/impl/lattice_Zn.cpp

@@ -455,7 +455,7 @@ void ZnSphereCodec::decode(uint64_t code, float* c) const {
     int nnz = 0;
     for (int i = 0; i < dim; i++) {
         if (c[i] != 0) {
-            if (signs & (1UL << nnz)) {
+            if (signs & (uint64_t(1) << nnz)) {
                 c[i] = -c[i];
             }
             nnz++;

data/vendor/faiss/faiss/impl/platform_macros.h

@@ -40,11 +40,13 @@
 
 #include <intrin.h>
 
+#ifndef __clang__
 inline int __builtin_ctzll(uint64_t x) {
     unsigned long ret;
     _BitScanForward64(&ret, x);
     return (int)ret;
 }
+#endif
 
 // cudatoolkit provides __builtin_ctz for NVCC >= 11.0
 #if !defined(__CUDACC__) || __CUDACC_VER_MAJOR__ < 11
@@ -55,13 +57,20 @@ inline int __builtin_ctz(unsigned long x) {
 }
 #endif
 
+#ifndef __clang__
 inline int __builtin_clzll(uint64_t x) {
     return (int)__lzcnt64(x);
 }
+#endif
 
 #define __builtin_popcount __popcnt
 #define __builtin_popcountl __popcnt64
 
+#ifndef __clang__
+#define __m128i_u __m128i
+#define __m256i_u __m256i
+#endif
+
 // MSVC does not define __SSEx__, and _M_IX86_FP is only defined on 32-bit
 // processors cf.
 // https://docs.microsoft.com/en-us/cpp/preprocessor/predefined-macros
@@ -82,6 +91,8 @@ inline int __builtin_clzll(uint64_t x) {
 #define __F16C__ 1
 #endif
 
+#define FAISS_ALWAYS_INLINE __forceinline
+
 #else
 /*******************************************************
  * Linux and OSX
@@ -92,9 +103,15 @@ inline int __builtin_clzll(uint64_t x) {
 
 // aligned should be *in front* of the declaration, for compatibility with
 // windows
+#ifdef SWIG
+#define ALIGNED(x)
+#else
 #define ALIGNED(x) __attribute__((aligned(x)))
+#endif
 
-#endif // _MSC_VER
+#define FAISS_ALWAYS_INLINE __attribute__((always_inline)) inline
+
+#endif
 
 #if defined(__GNUC__) || defined(__clang__)
 #define FAISS_DEPRECATED(msg) __attribute__((deprecated(msg)))

data/vendor/faiss/faiss/impl/pq4_fast_scan.cpp

@@ -54,6 +54,9 @@ void pq4_pack_codes(
     FAISS_THROW_IF_NOT(nb % bbs == 0);
     FAISS_THROW_IF_NOT(nsq % 2 == 0);
 
+    if (nb == 0) {
+        return;
+    }
     memset(blocks, 0, nb * nsq / 2);
     const uint8_t perm0[16] = {
             0, 8, 1, 9, 2, 10, 3, 11, 4, 12, 5, 13, 6, 14, 7, 15};

data/vendor/faiss/faiss/impl/pq4_fast_scan.h

@@ -24,6 +24,9 @@
 
 namespace faiss {
 
+struct NormTableScaler;
+struct SIMDResultHandler;
+
 /** Pack codes for consumption by the SIMD kernels.
  *  The unused bytes are set to 0.
  *
@@ -117,7 +120,6 @@ void pq4_pack_LUT(int nq, int nsq, const uint8_t* src, uint8_t* dest);
  * @param LUT     packed look-up table
  * @param scaler  scaler to scale the encoded norm
  */
-template <class ResultHandler, class Scaler>
 void pq4_accumulate_loop(
         int nq,
         size_t nb,
@@ -125,8 +127,8 @@ void pq4_accumulate_loop(
         int nsq,
         const uint8_t* codes,
         const uint8_t* LUT,
-        ResultHandler& res,
-        const Scaler& scaler);
+        SIMDResultHandler& res,
+        const NormTableScaler* scaler);
 
 /* qbs versions, supported only for bbs=32.
  *
@@ -178,14 +180,13 @@ int pq4_pack_LUT_qbs_q_map(
  * @param res     call-back for the resutls
  * @param scaler  scaler to scale the encoded norm
  */
-template <class ResultHandler, class Scaler>
 void pq4_accumulate_loop_qbs(
         int qbs,
         size_t nb,
         int nsq,
         const uint8_t* codes,
         const uint8_t* LUT,
-        ResultHandler& res,
-        const Scaler& scaler);
+        SIMDResultHandler& res,
+        const NormTableScaler* scaler = nullptr);
 
 } // namespace faiss

data/vendor/faiss/faiss/impl/pq4_fast_scan_search_1.cpp

@@ -125,7 +125,7 @@ void accumulate_fixed_blocks(
         ResultHandler& res,
         const Scaler& scaler) {
     constexpr int bbs = 32 * BB;
-    for (int64_t j0 = 0; j0 < nb; j0 += bbs) {
+    for (size_t j0 = 0; j0 < nb; j0 += bbs) {
         FixedStorageHandler<NQ, 2 * BB> res2;
         kernel_accumulate_block<NQ, BB>(nsq, codes, LUT, res2, scaler);
         res.set_block_origin(0, j0);
@@ -134,10 +134,8 @@ void accumulate_fixed_blocks(
     }
 }
 
-} // anonymous namespace
-
 template <class ResultHandler, class Scaler>
-void pq4_accumulate_loop(
+void pq4_accumulate_loop_fixed_scaler(
         int nq,
         size_t nb,
         int bbs,
@@ -172,39 +170,55 @@ void pq4_accumulate_loop(
 #undef DISPATCH
 }
 
-// explicit template instantiations
-
-#define INSTANTIATE_ACCUMULATE(TH, C, with_id_map, S)         \
-    template void pq4_accumulate_loop<TH<C, with_id_map>, S>( \
-            int,                                              \
-            size_t,                                           \
-            int,                                              \
-            int,                                              \
-            const uint8_t*,                                   \
-            const uint8_t*,                                   \
-            TH<C, with_id_map>&,                              \
-            const S&);
-
-using DS = DummyScaler;
-using NS = NormTableScaler;
-
-#define INSTANTIATE_3(C, with_id_map)                               \
-    INSTANTIATE_ACCUMULATE(SingleResultHandler, C, with_id_map, DS) \
-    INSTANTIATE_ACCUMULATE(HeapHandler, C, with_id_map, DS)         \
-    INSTANTIATE_ACCUMULATE(ReservoirHandler, C, with_id_map, DS)    \
-                                                                    \
-    INSTANTIATE_ACCUMULATE(SingleResultHandler, C, with_id_map, NS) \
-    INSTANTIATE_ACCUMULATE(HeapHandler, C, with_id_map, NS)         \
-    INSTANTIATE_ACCUMULATE(ReservoirHandler, C, with_id_map, NS)
-
-using Csi = CMax<uint16_t, int>;
-INSTANTIATE_3(Csi, false);
-using CsiMin = CMin<uint16_t, int>;
-INSTANTIATE_3(CsiMin, false);
-
-using Csl = CMax<uint16_t, int64_t>;
-INSTANTIATE_3(Csl, true);
-using CslMin = CMin<uint16_t, int64_t>;
-INSTANTIATE_3(CslMin, true);
+template <class ResultHandler>
+void pq4_accumulate_loop_fixed_handler(
+        int nq,
+        size_t nb,
+        int bbs,
+        int nsq,
+        const uint8_t* codes,
+        const uint8_t* LUT,
+        ResultHandler& res,
+        const NormTableScaler* scaler) {
+    if (scaler) {
+        pq4_accumulate_loop_fixed_scaler(
+                nq, nb, bbs, nsq, codes, LUT, res, *scaler);
+    } else {
+        DummyScaler dscaler;
+        pq4_accumulate_loop_fixed_scaler(
+                nq, nb, bbs, nsq, codes, LUT, res, dscaler);
+    }
+}
+
+struct Run_pq4_accumulate_loop {
+    template <class ResultHandler>
+    void f(ResultHandler& res,
+           int nq,
+           size_t nb,
+           int bbs,
+           int nsq,
+           const uint8_t* codes,
+           const uint8_t* LUT,
+           const NormTableScaler* scaler) {
+        pq4_accumulate_loop_fixed_handler(
+                nq, nb, bbs, nsq, codes, LUT, res, scaler);
+    }
+};
+
+} // anonymous namespace
+
+void pq4_accumulate_loop(
+        int nq,
+        size_t nb,
+        int bbs,
+        int nsq,
+        const uint8_t* codes,
+        const uint8_t* LUT,
+        SIMDResultHandler& res,
+        const NormTableScaler* scaler) {
+    Run_pq4_accumulate_loop consumer;
+    dispatch_SIMDResultHanlder(
+            res, consumer, nq, nb, bbs, nsq, codes, LUT, scaler);
+}
 
 } // namespace faiss