faiss 0.2.0 → 0.2.4

data/vendor/faiss/faiss/gpu/GpuCloner.cpp

@@ -5,51 +5,52 @@
  * LICENSE file in the root directory of this source tree.
  */
 
-
 #include <faiss/gpu/GpuCloner.h>
+#include <faiss/impl/FaissAssert.h>
 #include <typeinfo>
 
-#include <faiss/gpu/GpuIndex.h>
-#include <faiss/impl/FaissAssert.h>
-#include <faiss/index_io.h>
+#include <faiss/gpu/StandardGpuResources.h>
+
 #include <faiss/IndexFlat.h>
 #include <faiss/IndexIVF.h>
 #include <faiss/IndexIVFFlat.h>
-#include <faiss/IndexScalarQuantizer.h>
 #include <faiss/IndexIVFPQ.h>
-#include <faiss/IndexReplicas.h>
 #include <faiss/IndexPreTransform.h>
+#include <faiss/IndexReplicas.h>
+#include <faiss/IndexScalarQuantizer.h>
 #include <faiss/MetaIndexes.h>
+#include <faiss/gpu/GpuIndex.h>
 #include <faiss/gpu/GpuIndexFlat.h>
 #include <faiss/gpu/GpuIndexIVFFlat.h>
 #include <faiss/gpu/GpuIndexIVFPQ.h>
 #include <faiss/gpu/GpuIndexIVFScalarQuantizer.h>
 #include <faiss/gpu/utils/DeviceUtils.h>
+#include <faiss/impl/FaissAssert.h>
+#include <faiss/index_io.h>
 
-namespace faiss { namespace gpu {
-
+namespace faiss {
+namespace gpu {
 
 /**********************************************************
  * Cloning to CPU
  **********************************************************/
 
-void ToCPUCloner::merge_index(Index *dst, Index *src, bool successive_ids)
-{
-    if (auto ifl = dynamic_cast<IndexFlat *>(dst)) {
-        auto ifl2 = dynamic_cast<const IndexFlat *>(src);
+void ToCPUCloner::merge_index(Index* dst, Index* src, bool successive_ids) {
+    if (auto ifl = dynamic_cast<IndexFlat*>(dst)) {
+        auto ifl2 = dynamic_cast<const IndexFlat*>(src);
         FAISS_ASSERT(ifl2);
         FAISS_ASSERT(successive_ids);
-        ifl->add(ifl2->ntotal, ifl2->xb.data());
-    } else if(auto ifl = dynamic_cast<IndexIVFFlat *>(dst)) {
-        auto ifl2 = dynamic_cast<IndexIVFFlat *>(src);
+        ifl->add(ifl2->ntotal, ifl2->get_xb());
+    } else if (auto ifl = dynamic_cast<IndexIVFFlat*>(dst)) {
+        auto ifl2 = dynamic_cast<IndexIVFFlat*>(src);
         FAISS_ASSERT(ifl2);
         ifl->merge_from(*ifl2, successive_ids ? ifl->ntotal : 0);
-    } else if(auto ifl = dynamic_cast<IndexIVFScalarQuantizer *>(dst)) {
-        auto ifl2 = dynamic_cast<IndexIVFScalarQuantizer *>(src);
+    } else if (auto ifl = dynamic_cast<IndexIVFScalarQuantizer*>(dst)) {
+        auto ifl2 = dynamic_cast<IndexIVFScalarQuantizer*>(src);
         FAISS_ASSERT(ifl2);
         ifl->merge_from(*ifl2, successive_ids ? ifl->ntotal : 0);
-    } else if(auto ifl = dynamic_cast<IndexIVFPQ *>(dst)) {
-        auto ifl2 = dynamic_cast<IndexIVFPQ *>(src);
+    } else if (auto ifl = dynamic_cast<IndexIVFPQ*>(dst)) {
+        auto ifl2 = dynamic_cast<IndexIVFPQ*>(src);
         FAISS_ASSERT(ifl2);
         ifl->merge_from(*ifl2, successive_ids ? ifl->ntotal : 0);
     } else {
@@ -57,24 +58,22 @@ void ToCPUCloner::merge_index(Index *dst, Index *src, bool successive_ids)
     }
 }
 
-
-Index *ToCPUCloner::clone_Index(const Index *index)
-{
-    if(auto ifl = dynamic_cast<const GpuIndexFlat *>(index)) {
-        IndexFlat *res = new IndexFlat();
+Index* ToCPUCloner::clone_Index(const Index* index) {
+    if (auto ifl = dynamic_cast<const GpuIndexFlat*>(index)) {
+        IndexFlat* res = new IndexFlat();
         ifl->copyTo(res);
         return res;
-    } else if(auto ifl = dynamic_cast<const GpuIndexIVFFlat *>(index)) {
-        IndexIVFFlat *res = new IndexIVFFlat();
+    } else if (auto ifl = dynamic_cast<const GpuIndexIVFFlat*>(index)) {
+        IndexIVFFlat* res = new IndexIVFFlat();
         ifl->copyTo(res);
         return res;
-    } else if(auto ifl =
-              dynamic_cast<const GpuIndexIVFScalarQuantizer *>(index)) {
-        IndexIVFScalarQuantizer *res = new IndexIVFScalarQuantizer();
+    } else if (
+            auto ifl = dynamic_cast<const GpuIndexIVFScalarQuantizer*>(index)) {
+        IndexIVFScalarQuantizer* res = new IndexIVFScalarQuantizer();
         ifl->copyTo(res);
         return res;
-    } else if(auto ipq = dynamic_cast<const GpuIndexIVFPQ *>(index)) {
-        IndexIVFPQ *res = new IndexIVFPQ();
+    } else if (auto ipq = dynamic_cast<const GpuIndexIVFPQ*>(index)) {
+        IndexIVFPQ* res = new IndexIVFPQ();
         ipq->copyTo(res);
         return res;
 
@@ -82,17 +81,17 @@ Index *ToCPUCloner::clone_Index(const Index *index)
         // objective is to make a single component out of them
         // (inverse op of ToGpuClonerMultiple)
 
-    } else if(auto ish = dynamic_cast<const IndexShards *>(index)) {
+    } else if (auto ish = dynamic_cast<const IndexShards*>(index)) {
         int nshard = ish->count();
         FAISS_ASSERT(nshard > 0);
-        Index *res = clone_Index(ish->at(0));
-        for(int i = 1; i < ish->count(); i++) {
-            Index *res_i = clone_Index(ish->at(i));
+        Index* res = clone_Index(ish->at(0));
+        for (int i = 1; i < ish->count(); i++) {
+            Index* res_i = clone_Index(ish->at(i));
             merge_index(res, res_i, ish->successive_ids);
             delete res_i;
         }
         return res;
-    } else if(auto ipr = dynamic_cast<const IndexReplicas *>(index)) {
+    } else if (auto ipr = dynamic_cast<const IndexReplicas*>(index)) {
         // just clone one of the replicas
         FAISS_ASSERT(ipr->count() > 0);
         return clone_Index(ipr->at(0));
@@ -101,81 +100,97 @@ Index *ToCPUCloner::clone_Index(const Index *index)
     }
 }
 
-faiss::Index * index_gpu_to_cpu(const faiss::Index *gpu_index)
-{
+faiss::Index* index_gpu_to_cpu(const faiss::Index* gpu_index) {
     ToCPUCloner cl;
     return cl.clone_Index(gpu_index);
 }
 
-
-
-
 /**********************************************************
  * Cloning to 1 GPU
  **********************************************************/
 
-ToGpuCloner::ToGpuCloner(GpuResourcesProvider *prov, int device,
-                         const GpuClonerOptions &options):
-    GpuClonerOptions(options), provider(prov), device(device)
-{}
+ToGpuCloner::ToGpuCloner(
+        GpuResourcesProvider* prov,
+        int device,
+        const GpuClonerOptions& options)
+        : GpuClonerOptions(options), provider(prov), device(device) {}
 
-Index *ToGpuCloner::clone_Index(const Index *index)
-{
-    if(auto ifl = dynamic_cast<const IndexFlat *>(index)) {
+Index* ToGpuCloner::clone_Index(const Index* index) {
+    using idx_t = Index::idx_t;
+    if (auto ifl = dynamic_cast<const IndexFlat*>(index)) {
         GpuIndexFlatConfig config;
         config.device = device;
         config.useFloat16 = useFloat16;
         config.storeTransposed = storeTransposed;
-
         return new GpuIndexFlat(provider, ifl, config);
-    } else if(auto ifl = dynamic_cast<const faiss::IndexIVFFlat *>(index)) {
+    } else if (
+            dynamic_cast<const IndexScalarQuantizer*>(index) &&
+            static_cast<const IndexScalarQuantizer*>(index)->sq.qtype ==
+                    ScalarQuantizer::QT_fp16) {
+        GpuIndexFlatConfig config;
+        config.device = device;
+        config.useFloat16 = true;
+        GpuIndexFlat* gif = new GpuIndexFlat(
+                provider, index->d, index->metric_type, config);
+        // transfer data by blocks
+        idx_t bs = 1024 * 1024;
+        for (idx_t i0 = 0; i0 < index->ntotal; i0 += bs) {
+            idx_t i1 = std::min(i0 + bs, index->ntotal);
+            std::vector<float> buffer((i1 - i0) * index->d);
+            index->reconstruct_n(i0, i1 - i0, buffer.data());
+            gif->add(i1 - i0, buffer.data());
+        }
+        assert(gif->getNumVecs() == index->ntotal);
+        return gif;
+    } else if (auto ifl = dynamic_cast<const faiss::IndexIVFFlat*>(index)) {
         GpuIndexIVFFlatConfig config;
         config.device = device;
         config.indicesOptions = indicesOptions;
         config.flatConfig.useFloat16 = useFloat16CoarseQuantizer;
         config.flatConfig.storeTransposed = storeTransposed;
 
-        GpuIndexIVFFlat *res =
-            new GpuIndexIVFFlat(provider,
-                                ifl->d,
-                                ifl->nlist,
-                                ifl->metric_type,
-                                config);
-        if(reserveVecs > 0 && ifl->ntotal == 0) {
+        GpuIndexIVFFlat* res = new GpuIndexIVFFlat(
+                provider, ifl->d, ifl->nlist, ifl->metric_type, config);
+        if (reserveVecs > 0 && ifl->ntotal == 0) {
             res->reserveMemory(reserveVecs);
         }
 
         res->copyFrom(ifl);
         return res;
-    } else if(auto ifl =
-              dynamic_cast<const faiss::IndexIVFScalarQuantizer *>(index)) {
+    } else if (
+            auto ifl = dynamic_cast<const faiss::IndexIVFScalarQuantizer*>(
+                    index)) {
         GpuIndexIVFScalarQuantizerConfig config;
         config.device = device;
         config.indicesOptions = indicesOptions;
         config.flatConfig.useFloat16 = useFloat16CoarseQuantizer;
         config.flatConfig.storeTransposed = storeTransposed;
 
-        GpuIndexIVFScalarQuantizer *res =
-            new GpuIndexIVFScalarQuantizer(provider,
-                                           ifl->d,
-                                           ifl->nlist,
-                                           ifl->sq.qtype,
-                                           ifl->metric_type,
-                                           ifl->by_residual,
-                                           config);
-        if(reserveVecs > 0 && ifl->ntotal == 0) {
+        GpuIndexIVFScalarQuantizer* res = new GpuIndexIVFScalarQuantizer(
+                provider,
+                ifl->d,
+                ifl->nlist,
+                ifl->sq.qtype,
+                ifl->metric_type,
+                ifl->by_residual,
+                config);
+        if (reserveVecs > 0 && ifl->ntotal == 0) {
             res->reserveMemory(reserveVecs);
         }
 
         res->copyFrom(ifl);
         return res;
-    } else if(auto ipq = dynamic_cast<const faiss::IndexIVFPQ *>(index)) {
-        if(verbose)
+    } else if (auto ipq = dynamic_cast<const faiss::IndexIVFPQ*>(index)) {
+        if (verbose) {
             printf("  IndexIVFPQ size %ld -> GpuIndexIVFPQ "
                    "indicesOptions=%d "
                    "usePrecomputed=%d useFloat16=%d reserveVecs=%ld\n",
-                   ipq->ntotal, indicesOptions, usePrecomputed,
-                   useFloat16, reserveVecs);
+                   ipq->ntotal,
+                   indicesOptions,
+                   usePrecomputed,
+                   useFloat16,
+                   reserveVecs);
+        }
         GpuIndexIVFPQConfig config;
         config.device = device;
         config.indicesOptions = indicesOptions;
@@ -184,154 +199,144 @@ Index *ToGpuCloner::clone_Index(const Index *index)
         config.useFloat16LookupTables = useFloat16;
         config.usePrecomputedTables = usePrecomputed;
 
-        GpuIndexIVFPQ *res = new GpuIndexIVFPQ(provider, ipq, config);
+        GpuIndexIVFPQ* res = new GpuIndexIVFPQ(provider, ipq, config);
 
-        if(reserveVecs > 0 && ipq->ntotal == 0) {
+        if (reserveVecs > 0 && ipq->ntotal == 0) {
             res->reserveMemory(reserveVecs);
         }
 
         return res;
     } else {
+        // default: use CPU cloner
         return Cloner::clone_Index(index);
     }
 }
 
-
-faiss::Index * index_cpu_to_gpu(
-       GpuResourcesProvider* provider, int device,
-       const faiss::Index *index,
-       const GpuClonerOptions *options)
-{
+faiss::Index* index_cpu_to_gpu(
+        GpuResourcesProvider* provider,
+        int device,
+        const faiss::Index* index,
+        const GpuClonerOptions* options) {
     GpuClonerOptions defaults;
     ToGpuCloner cl(provider, device, options ? *options : defaults);
     return cl.clone_Index(index);
 }
 
-
 /**********************************************************
  * Cloning to multiple GPUs
  **********************************************************/
 
 ToGpuClonerMultiple::ToGpuClonerMultiple(
-                        std::vector<GpuResourcesProvider *> & provider,
-                        std::vector<int>& devices,
-                        const GpuMultipleClonerOptions &options):
-    GpuMultipleClonerOptions(options)
-{
+        std::vector<GpuResourcesProvider*>& provider,
+        std::vector<int>& devices,
+        const GpuMultipleClonerOptions& options)
+        : GpuMultipleClonerOptions(options) {
     FAISS_ASSERT(provider.size() == devices.size());
-    for(int i = 0; i < provider.size(); i++) {
+    for (int i = 0; i < provider.size(); i++) {
         sub_cloners.push_back(ToGpuCloner(provider[i], devices[i], options));
     }
 }
 
-
 ToGpuClonerMultiple::ToGpuClonerMultiple(
-                        const std::vector<ToGpuCloner> & sub_cloners,
-                        const GpuMultipleClonerOptions &options):
-    GpuMultipleClonerOptions(options),
-    sub_cloners(sub_cloners)
-{}
-
-
-void ToGpuClonerMultiple::copy_ivf_shard (
-                         const IndexIVF *index_ivf, IndexIVF *idx2,
-                         long n, long i)
-{
+        const std::vector<ToGpuCloner>& sub_cloners,
+        const GpuMultipleClonerOptions& options)
+        : GpuMultipleClonerOptions(options), sub_cloners(sub_cloners) {}
+
+void ToGpuClonerMultiple::copy_ivf_shard(
+        const IndexIVF* index_ivf,
+        IndexIVF* idx2,
+        long n,
+        long i) {
     if (shard_type == 2) {
         long i0 = i * index_ivf->ntotal / n;
         long i1 = (i + 1) * index_ivf->ntotal / n;
 
-        if(verbose)
-            printf("IndexShards shard %ld indices %ld:%ld\n",
-                   i, i0, i1);
+        if (verbose)
+            printf("IndexShards shard %ld indices %ld:%ld\n", i, i0, i1);
         index_ivf->copy_subset_to(*idx2, 2, i0, i1);
         FAISS_ASSERT(idx2->ntotal == i1 - i0);
     } else if (shard_type == 1) {
-        if(verbose)
-            printf("IndexShards shard %ld select modulo %ld = %ld\n",
-                   i, n, i);
+        if (verbose)
+            printf("IndexShards shard %ld select modulo %ld = %ld\n", i, n, i);
         index_ivf->copy_subset_to(*idx2, 1, n, i);
     } else {
-        FAISS_THROW_FMT ("shard_type %d not implemented", shard_type);
+        FAISS_THROW_FMT("shard_type %d not implemented", shard_type);
     }
-
 }
 
-Index * ToGpuClonerMultiple::clone_Index_to_shards (const Index *index)
-{
+Index* ToGpuClonerMultiple::clone_Index_to_shards(const Index* index) {
     long n = sub_cloners.size();
 
-    auto index_ivfpq =
-        dynamic_cast<const faiss::IndexIVFPQ *>(index);
-    auto index_ivfflat =
-        dynamic_cast<const faiss::IndexIVFFlat *>(index);
+    auto index_ivfpq = dynamic_cast<const faiss::IndexIVFPQ*>(index);
+    auto index_ivfflat = dynamic_cast<const faiss::IndexIVFFlat*>(index);
     auto index_ivfsq =
-        dynamic_cast<const faiss::IndexIVFScalarQuantizer *>(index);
-    auto index_flat =
-        dynamic_cast<const faiss::IndexFlat *>(index);
-    FAISS_THROW_IF_NOT_MSG (
-        index_ivfpq || index_ivfflat || index_flat || index_ivfsq,
-        "IndexShards implemented only for "
-        "IndexIVFFlat, IndexIVFScalarQuantizer, "
-        "IndexFlat and IndexIVFPQ");
+            dynamic_cast<const faiss::IndexIVFScalarQuantizer*>(index);
+    auto index_flat = dynamic_cast<const faiss::IndexFlat*>(index);
+    FAISS_THROW_IF_NOT_MSG(
+            index_ivfpq || index_ivfflat || index_flat || index_ivfsq,
+            "IndexShards implemented only for "
+            "IndexIVFFlat, IndexIVFScalarQuantizer, "
+            "IndexFlat and IndexIVFPQ");
 
     std::vector<faiss::Index*> shards(n);
 
-    for(long i = 0; i < n; i++) {
+    for (long i = 0; i < n; i++) {
         // make a shallow copy
-        if(reserveVecs)
-            sub_cloners[i].reserveVecs =
-                (reserveVecs + n - 1) / n;
+        if (reserveVecs)
+            sub_cloners[i].reserveVecs = (reserveVecs + n - 1) / n;
 
         if (index_ivfpq) {
             faiss::IndexIVFPQ idx2(
-                       index_ivfpq->quantizer, index_ivfpq->d,
-                       index_ivfpq->nlist, index_ivfpq->code_size,
-                       index_ivfpq->pq.nbits);
+                    index_ivfpq->quantizer,
+                    index_ivfpq->d,
+                    index_ivfpq->nlist,
+                    index_ivfpq->code_size,
+                    index_ivfpq->pq.nbits);
             idx2.metric_type = index_ivfpq->metric_type;
             idx2.pq = index_ivfpq->pq;
             idx2.nprobe = index_ivfpq->nprobe;
             idx2.use_precomputed_table = 0;
             idx2.is_trained = index->is_trained;
-            copy_ivf_shard (index_ivfpq, &idx2, n, i);
+            copy_ivf_shard(index_ivfpq, &idx2, n, i);
             shards[i] = sub_cloners[i].clone_Index(&idx2);
         } else if (index_ivfflat) {
             faiss::IndexIVFFlat idx2(
-                       index_ivfflat->quantizer, index->d,
-                       index_ivfflat->nlist, index_ivfflat->metric_type);
+                    index_ivfflat->quantizer,
+                    index->d,
+                    index_ivfflat->nlist,
+                    index_ivfflat->metric_type);
             idx2.nprobe = index_ivfflat->nprobe;
             idx2.is_trained = index->is_trained;
-            copy_ivf_shard (index_ivfflat, &idx2, n, i);
+            copy_ivf_shard(index_ivfflat, &idx2, n, i);
             shards[i] = sub_cloners[i].clone_Index(&idx2);
         } else if (index_ivfsq) {
             faiss::IndexIVFScalarQuantizer idx2(
-                       index_ivfsq->quantizer, index->d, index_ivfsq->nlist,
-                       index_ivfsq->sq.qtype,
-                       index_ivfsq->metric_type,
-                       index_ivfsq->by_residual);
+                    index_ivfsq->quantizer,
+                    index->d,
+                    index_ivfsq->nlist,
+                    index_ivfsq->sq.qtype,
+                    index_ivfsq->metric_type,
+                    index_ivfsq->by_residual);
 
             idx2.nprobe = index_ivfsq->nprobe;
             idx2.is_trained = index->is_trained;
             idx2.sq = index_ivfsq->sq;
-            copy_ivf_shard (index_ivfsq, &idx2, n, i);
+            copy_ivf_shard(index_ivfsq, &idx2, n, i);
             shards[i] = sub_cloners[i].clone_Index(&idx2);
         } else if (index_flat) {
-            faiss::IndexFlat idx2 (
-                                   index->d, index->metric_type);
+            faiss::IndexFlat idx2(index->d, index->metric_type);
             shards[i] = sub_cloners[i].clone_Index(&idx2);
             if (index->ntotal > 0) {
                 long i0 = index->ntotal * i / n;
                 long i1 = index->ntotal * (i + 1) / n;
-                shards[i]->add (i1 - i0,
-                                index_flat->xb.data() + i0 * index->d);
+                shards[i]->add(i1 - i0, index_flat->get_xb() + i0 * index->d);
             }
         }
     }
 
     bool successive_ids = index_flat != nullptr;
-    faiss::IndexShards *res =
-        new faiss::IndexShards(index->d, true,
-                               successive_ids);
+    faiss::IndexShards* res =
+            new faiss::IndexShards(index->d, true, successive_ids);
 
     for (int i = 0; i < n; i++) {
         res->add_shard(shards[i]);
@@ -341,33 +346,32 @@ Index * ToGpuClonerMultiple::clone_Index_to_shards (const Index *index)
     return res;
 }
 
-Index *ToGpuClonerMultiple::clone_Index(const Index *index)
-{
+Index* ToGpuClonerMultiple::clone_Index(const Index* index) {
     long n = sub_cloners.size();
     if (n == 1)
         return sub_cloners[0].clone_Index(index);
 
-    if(dynamic_cast<const IndexFlat *>(index) ||
-       dynamic_cast<const faiss::IndexIVFFlat *>(index) ||
-       dynamic_cast<const faiss::IndexIVFScalarQuantizer *>(index) ||
-       dynamic_cast<const faiss::IndexIVFPQ *>(index)) {
-        if(!shard) {
-            IndexReplicas * res = new IndexReplicas();
-            for(auto & sub_cloner: sub_cloners) {
+    if (dynamic_cast<const IndexFlat*>(index) ||
+        dynamic_cast<const faiss::IndexIVFFlat*>(index) ||
+        dynamic_cast<const faiss::IndexIVFScalarQuantizer*>(index) ||
+        dynamic_cast<const faiss::IndexIVFPQ*>(index)) {
+        if (!shard) {
+            IndexReplicas* res = new IndexReplicas();
+            for (auto& sub_cloner : sub_cloners) {
                 res->addIndex(sub_cloner.clone_Index(index));
             }
             res->own_fields = true;
             return res;
         } else {
-            return clone_Index_to_shards (index);
+            return clone_Index_to_shards(index);
         }
-    } else if(auto miq = dynamic_cast<const MultiIndexQuantizer *>(index)) {
+    } else if (auto miq = dynamic_cast<const MultiIndexQuantizer*>(index)) {
         if (verbose) {
             printf("cloning MultiIndexQuantizer: "
                    "will be valid only for search k=1\n");
         }
-        const ProductQuantizer & pq = miq->pq;
-        IndexSplitVectors *splitv = new IndexSplitVectors(pq.d, true);
+        const ProductQuantizer& pq = miq->pq;
+        IndexSplitVectors* splitv = new IndexSplitVectors(pq.d, true);
         splitv->own_fields = true;
 
         for (int m = 0; m < pq.M; m++) {
@@ -377,12 +381,13 @@ Index *ToGpuClonerMultiple::clone_Index(const Index *index)
             long i1 = pq.M <= n ? (m + 1) * n / pq.M : i0 + 1;
             std::vector<ToGpuCloner> sub_cloners_2;
             sub_cloners_2.insert(
-                                 sub_cloners_2.begin(), sub_cloners.begin() + i0,
-                                 sub_cloners.begin() + i1);
+                    sub_cloners_2.begin(),
+                    sub_cloners.begin() + i0,
+                    sub_cloners.begin() + i1);
             ToGpuClonerMultiple cm(sub_cloners_2, *this);
-            IndexFlatL2 idxc (pq.dsub);
-            idxc.add (pq.ksub, pq.centroids.data() + m * pq.d * pq.ksub);
-            Index *idx2 = cm.clone_Index(&idxc);
+            IndexFlatL2 idxc(pq.dsub);
+            idxc.add(pq.ksub, pq.centroids.data() + m * pq.d * pq.ksub);
+            Index* idx2 = cm.clone_Index(&idxc);
             splitv->add_sub_index(idx2);
         }
         return splitv;
@@ -391,17 +396,39 @@ Index *ToGpuClonerMultiple::clone_Index(const Index *index)
     }
 }
 
-
-
-faiss::Index * index_cpu_to_gpu_multiple(
-       std::vector<GpuResourcesProvider*> & provider,
-       std::vector<int> &devices,
-       const faiss::Index *index,
-       const GpuMultipleClonerOptions *options)
-{
+faiss::Index* index_cpu_to_gpu_multiple(
+        std::vector<GpuResourcesProvider*>& provider,
+        std::vector<int>& devices,
+        const faiss::Index* index,
+        const GpuMultipleClonerOptions* options) {
     GpuMultipleClonerOptions defaults;
     ToGpuClonerMultiple cl(provider, devices, options ? *options : defaults);
     return cl.clone_Index(index);
 }
 
-} } // namespace
+GpuProgressiveDimIndexFactory::GpuProgressiveDimIndexFactory(int ngpu) {
+    FAISS_THROW_IF_NOT(ngpu >= 1);
+    devices.resize(ngpu);
+    vres.resize(ngpu);
+
+    for (int i = 0; i < ngpu; i++) {
+        vres[i] = new StandardGpuResources();
+        devices[i] = i;
+    }
+    ncall = 0;
+}
+
+GpuProgressiveDimIndexFactory::~GpuProgressiveDimIndexFactory() {
+    for (int i = 0; i < vres.size(); i++) {
+        delete vres[i];
+    }
+}
+
+Index* GpuProgressiveDimIndexFactory::operator()(int dim) {
+    IndexFlatL2 index(dim);
+    ncall++;
+    return index_cpu_to_gpu_multiple(vres, devices, &index, &options);
+}
+
+} // namespace gpu
+} // namespace faiss