infer_trtis_backend.h

Go to the documentation of this file.

/*
 * SPDX-FileCopyrightText: Copyright (c) 2020-2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: LicenseRef-NvidiaProprietary
 *
 * NVIDIA CORPORATION, its affiliates and licensors retain all intellectual
 * property and proprietary rights in and to this material, related
 * documentation and any modifications thereto. Any use, reproduction,
 * disclosure or distribution of this material and related documentation
 * without an express license agreement from NVIDIA CORPORATION or
 * its affiliates is strictly prohibited.
 */
 
#ifndef __NVDSINFER_TRTIS_BACKEND_H__
#define __NVDSINFER_TRTIS_BACKEND_H__
 
#include "infer_base_backend.h"
#include "infer_common.h"
#include "infer_post_datatypes.h"
#include "infer_utils.h"
 
namespace nvdsinferserver {
 
class TrtServerAllocator;
class TrtISServer;
class TrtServerRequest;
class TrtServerResponse;
 
class TrtISBackend : public BaseBackend {
public:
    TrtISBackend(
        const std::string& name, int64_t version, TrtServerPtr ptr = nullptr);
 
    ~TrtISBackend() override;
 
    void addClassifyParams(const TritonClassParams& c) {
        m_ClassifyParams.emplace_back(c); }
 
    void setOutputPoolSize(int size) { m_PerPoolSize = size; }
    int outputPoolSize() const { return m_PerPoolSize; }
    void setOutputMemType(InferMemType memType) { m_OutputMemType = memType; }
    InferMemType outputMemType() const { return m_OutputMemType; }
    void setOutputDevId(int64_t devId) { m_OutputDevId = devId; }
    int64_t outputDevId() const { return m_OutputDevId; }
    std::vector<TritonClassParams> getClassifyParams() {
        return m_ClassifyParams;}
    const std::string& model() const { return m_Model; }
    int64_t version() const { return m_ModelVersion; }
    NvDsInferStatus initialize() override;
 
    NvDsInferStatus specifyInputDims(const InputShapes& shapes) override;
 
    NvDsInferStatus enqueue(SharedBatchArray inputs, SharedCuStream stream,
        InputsConsumed bufConsumed, InferenceDone inferenceDone) override;
 
    void setTensorMaxBytes(const std::string& name, size_t maxBytes)
    {
        size_t& bytes = m_TensorMaxBytes[name];
        bytes = std::max<size_t>(maxBytes, bytes);
        bytes = INFER_ROUND_UP(bytes, INFER_MEM_ALIGNMENT);
    }
 
protected:
    // interface for derived class
 
    virtual void requestTritonOutputNames(std::set<std::string>& outNames);
 
    virtual NvDsInferStatus ensureServerReady();
 
    virtual NvDsInferStatus ensureModelReady();
 
    NvDsInferStatus setupReorderThread();
 
    void setAllocator(UniqTritonAllocator allocator)
    {
        m_ResponseAllocator = std::move(allocator);
    }
 
    virtual NvDsInferStatus setupLayersInfo();
 
    TrtServerPtr& server() { return m_Server; }
 
    using AsyncDone = std::function<void(NvDsInferStatus, SharedBatchArray)>;
    virtual NvDsInferStatus Run(
        SharedBatchArray inputs, InputsConsumed bufConsumed,
        AsyncDone asyncDone);
 
    NvDsInferStatus fixateDims(const SharedBatchArray& bufs);
 
    SharedSysMem allocateResponseBuf(
        const std::string& tensor, size_t bytes, InferMemType memType, int64_t devId);
 
    void releaseResponseBuf(const std::string& tensor, SharedSysMem mem);
 
    NvDsInferStatus ensureInputs(SharedBatchArray& inputs);
 
    enum { kName, kGpuId, kMemType };
    using PoolKey = std::tuple<std::string, int64_t, InferMemType>;
    using PoolValue = SharedBufPool<UniqSysMem>;
 
    PoolValue findResponsePool(PoolKey& key);
 
    PoolValue createResponsePool(PoolKey& key, size_t bytes);
 
    void serverInferCompleted(
        std::shared_ptr<TrtServerRequest> request,
        std::unique_ptr<TrtServerResponse> uniqResponse,
        InputsConsumed inputsConsumed, AsyncDone asyncDone);
 
    struct ReorderItem {
        NvDsInferStatus status = NVDSINFER_SUCCESS;
        SharedBatchArray inputs;
        SharedBatchArray outputs;
 
        std::promise<void> promise;
        std::future<void> future;
        InferenceDone inferenceDone;
    };
    /*
     * @brief Pointer to the reorder thread task.
     */
    using ReorderItemPtr = std::shared_ptr<ReorderItem>;
    bool inferenceDoneReorderLoop(ReorderItemPtr item);
 
    bool debatchingOutput(SharedBatchArray& outputs, SharedBatchArray& inputs);
 
private:
    std::string m_Model;
    int64_t m_ModelVersion = -1;
    TrtServerPtr m_Server;
    bool m_NeedUnload = false;
    std::vector<TritonClassParams> m_ClassifyParams;
    ShrTritonAllocator m_ResponseAllocator;
    InferMemType m_OutputMemType = InferMemType::kNone;
    int64_t m_OutputDevId = -1;
    int m_PerPoolSize = 2;
    std::map<PoolKey, PoolValue> m_ResponsePool;
    using SharedMutex = std::shared_timed_mutex;
    SharedMutex m_ResponseMutex;
    std::unordered_map<std::string, size_t> m_TensorMaxBytes;
 
    using ReorderThread = QueueThread<std::vector<ReorderItemPtr>>;
    std::unique_ptr<ReorderThread> m_ReorderThread;
};
 
} // namespace nvdsinferserver
 
#endif