Commit 37a9e4ed authored by jinhai's avatar jinhai
Browse files

Merge branch 'branch-0.5.0' into 'branch-0.5.0' 

MS-631 IVFSQ8H Index support

See merge request megasearch/milvus!687

Former-commit-id: 9fc0fdae4a4bab09e7bf653dc81a4f882c046925
parents bcd79ce1 14ce9652

cpp/CHANGELOG.md

+1 −0
Original line number Diff line number Diff line
@@ -36,6 +36,7 @@ Please mark all change in change log and use the ticket from JIRA. 


## New Feature

- MS-627 - Integrate new index: IVFSQHybrid

- MS-631 - IVFSQ8H Index support



## Task

- MS-554 - Change license to Apache 2.0

cpp/src/cache/CpuCacheMgr.cpp

+1 −1
Original line number Diff line number Diff line
@@ -32,7 +32,7 @@ CpuCacheMgr::CpuCacheMgr() { 
    server::Config& config = server::Config::GetInstance();

    Status s;



    int32_t cpu_cache_cap;

    int64_t cpu_cache_cap;

    s = config.GetCacheConfigCpuCacheCapacity(cpu_cache_cap);

    if (!s.ok()) {

        SERVER_LOG_ERROR << s.message();

cpp/src/cache/GpuCacheMgr.cpp

+2 −2
Original line number Diff line number Diff line
@@ -36,12 +36,12 @@ GpuCacheMgr::GpuCacheMgr() { 
    server::Config& config = server::Config::GetInstance();

    Status s;



    int32_t gpu_cache_cap;

    int64_t gpu_cache_cap;

    s = config.GetCacheConfigGpuCacheCapacity(gpu_cache_cap);

    if (!s.ok()) {

        SERVER_LOG_ERROR << s.message();

    }

    int32_t cap = gpu_cache_cap * G_BYTE;

    int64_t cap = gpu_cache_cap * G_BYTE;

    cache_ = std::make_shared<Cache<DataObjPtr>>(cap, 1UL << 32);



    float gpu_mem_threshold;

cpp/src/core/knowhere/knowhere/index/vector_index/IndexIVFSQHybrid.cpp

+21 −5
Original line number Diff line number Diff line
@@ -100,16 +100,20 @@ IVFSQHybrid::CopyCpuToGpu(const int64_t& device_id, const Config& config) { 
void

IVFSQHybrid::LoadImpl(const BinarySet& index_binary) {

    FaissBaseIndex::LoadImpl(index_binary);  // load on cpu

    auto* ivf_index = dynamic_cast<faiss::IndexIVF*>(index_.get());

    ivf_index->backup_quantizer();

}



void

IVFSQHybrid::search_impl(int64_t n, const float* data, int64_t k, float* distances, int64_t* labels,

                         const Config& cfg) {

    if (gpu_mode) {

    if (gpu_mode == 2) {

        GPUIVF::search_impl(n, data, k, distances, labels, cfg);

    } else {

    } else if (gpu_mode == 1) {

        ResScope rs(res_, gpu_id_);

        IVF::search_impl(n, data, k, distances, labels, cfg);

    } else if (gpu_mode == 0) {

        IVF::search_impl(n, data, k, distances, labels, cfg);

    }

}
@@ -137,8 +141,12 @@ IVFSQHybrid::LoadQuantizer(const Config& conf) { 
        delete gpu_index;



        auto q = std::make_shared<FaissIVFQuantizer>();

        q->quantizer = index_composition->quantizer;



        auto& q_ptr = index_composition->quantizer;

        q->size = q_ptr->d * q_ptr->getNumVecs() * sizeof(float);

        q->quantizer = q_ptr;

        res_ = res;

        gpu_mode = 1;

        return q;

    } else {

        KNOWHERE_THROW_MSG("CopyCpuToGpu Error, can't get gpu_resource");
@@ -156,7 +164,7 @@ IVFSQHybrid::SetQuantizer(const QuantizerPtr& q) { 


    faiss::gpu::GpuIndexFlat* is_gpu_flat_index = dynamic_cast<faiss::gpu::GpuIndexFlat*>(ivf_index->quantizer);

    if (is_gpu_flat_index == nullptr) {

        delete ivf_index->quantizer;

//        delete ivf_index->quantizer;

        ivf_index->quantizer = ivf_quantizer->quantizer;

    }

}
@@ -199,10 +207,18 @@ IVFSQHybrid::LoadData(const knowhere::QuantizerPtr& q, const Config& conf) { 


        auto gpu_index = faiss::gpu::index_cpu_to_gpu(res->faiss_res.get(), gpu_id_, index_composition, &option);

        index_.reset(gpu_index);

        gpu_mode = true;  // all in gpu

        gpu_mode = 2;  // all in gpu

    } else {

        KNOWHERE_THROW_MSG("CopyCpuToGpu Error, can't get gpu_resource");

    }

}



FaissIVFQuantizer::~FaissIVFQuantizer() {

    if (quantizer != nullptr) {

        delete quantizer;

        quantizer = nullptr;

    }

    // else do nothing

}



}  // namespace knowhere

cpp/src/core/knowhere/knowhere/index/vector_index/IndexIVFSQHybrid.h

+6 −4
Original line number Diff line number Diff line
@@ -27,23 +27,25 @@ namespace knowhere { 


struct FaissIVFQuantizer : public Quantizer {

    faiss::gpu::GpuIndexFlat* quantizer = nullptr;



    ~FaissIVFQuantizer() override;

};

using FaissIVFQuantizerPtr = std::shared_ptr<FaissIVFQuantizer>;



class IVFSQHybrid : public GPUIVFSQ {

 public:

    explicit IVFSQHybrid(const int& device_id) : GPUIVFSQ(device_id) {

        gpu_mode = false;

        gpu_mode = 0;

    }



    explicit IVFSQHybrid(std::shared_ptr<faiss::Index> index) : GPUIVFSQ(-1) {

        index_ = index;

        gpu_mode = false;

        gpu_mode = 0;

    }



    explicit IVFSQHybrid(std::shared_ptr<faiss::Index> index, const int64_t& device_id, ResPtr& resource)

        : GPUIVFSQ(index, device_id, resource) {

        gpu_mode = true;

        gpu_mode = 2;

    }



 public:
@@ -76,7 +78,7 @@ class IVFSQHybrid : public GPUIVFSQ { 
    LoadImpl(const BinarySet& index_binary) override;



 protected:

    bool gpu_mode = false;

    int64_t gpu_mode = 0; // 0,1,2

};



}  // namespace knowhere