infer_batch_buffer.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 __NVDSINFERSERVER_BATCH_BUFFER_H__
#define __NVDSINFERSERVER_BATCH_BUFFER_H__
 
#include <algorithm>
#include <set>
#include <unordered_map>
#include "nvbufsurftransform.h"
#include "infer_common.h"
#include "infer_datatypes.h"
 
namespace nvdsinferserver {
 
class INFER_EXPORT_API BaseBatchBuffer : public IBatchBuffer {
protected:
    explicit BaseBatchBuffer(uint32_t batchSize) : m_BatchSize(batchSize) {}
 
public:
    ~BaseBatchBuffer() override = default;
    const InferBufferDescription& getBufDesc() const final { return m_Desc; }
    uint32_t getBatchSize() const final { return m_BatchSize; }
    uint64_t getTotalBytes() const override
    {
        assert(m_Desc.dataType != InferDataType::kNone);
        assert(m_Desc.elementSize >= 0);
        uint32_t b = getBatchSize();
        return m_Desc.elementSize * m_Desc.dims.numElements * (b ? b : 1);
    }
    void setBufDesc(const InferBufferDescription& desc) { m_Desc = desc; }
    InferBufferDescription& mutableBufDesc() { return m_Desc; }
    virtual void setBatchSize(uint32_t size) { m_BatchSize = size; }
    const SharedCuEvent& cuEvent() const { return m_CuEvent; }
    void setCuEvent(SharedCuEvent e) { m_CuEvent = std::move(e); }
    void setSyncObj(NvBufSurfTransformSyncObj_t SyncObj) { assert(m_SyncObj==nullptr); m_SyncObj = SyncObj; }
    NvBufSurfTransformSyncObj_t& getSyncObj() { return m_SyncObj; }
    void waitForSyncObj()
    {
        if (m_SyncObj)
        {
            NvBufSurfTransformSyncObjWait (m_SyncObj, -1);
            NvBufSurfTransformSyncObjDestroy (&m_SyncObj);
            m_SyncObj = nullptr;
        }
    }
 
    // attached buffers
    void attach(SharedBatchBuf buf)
    {
        m_Attaches.emplace_back(std::move(buf));
        assert(!hasAttachLoop());
    }
    void detach() { m_Attaches.clear(); }
    bool hasAttachedBufs() const { return !m_Attaches.empty();}
    const std::vector<SharedBatchBuf>& attachedBufs() const
    {
        return m_Attaches;
    }
    std::vector<SharedBatchBuf>& mutableAttachedBufs() { return m_Attaches; }
    bool hasAttachLoop() const
    {
        std::set<BaseBatchBuffer*> allAttached;
        for (auto const& buf : m_Attaches) {
            assert(buf);
            if (allAttached.count(buf.get())) {
                return true;
            }
            allAttached.emplace(buf.get());
        }
        return false;
    }
 
    void setBufId(uint64_t id) { m_BufId = id; }
    uint64_t bufId() const { return m_BufId; }
 
private:
    InferBufferDescription m_Desc{
        InferMemType::kNone, 0, InferDataType::kNone, {0}};
    uint32_t m_BatchSize = 0;
    SharedCuEvent m_CuEvent;
    std::vector<SharedBatchBuf> m_Attaches;
    uint64_t m_BufId = UINT64_C(0);  // for debug and track
    NvBufSurfTransformSyncObj_t m_SyncObj = nullptr;
};
 
class BaseBatchArray : public IBatchArray {
public:
    ~BaseBatchArray() override = default;
    uint32_t getSize() const final { return m_Bufs.size(); }
    const IBatchBuffer* getBuffer(uint32_t arrayIdx) const final
    {
        assert(arrayIdx < (uint32_t)m_Bufs.size());
        assert(m_Bufs.at(arrayIdx).get());
        return m_Bufs.at(arrayIdx).get();
    }
    SharedIBatchBuffer getSafeBuf(uint32_t arrayIdx) const final
    {
        assert(arrayIdx < (uint32_t)m_Bufs.size());
        assert(m_Bufs.at(arrayIdx).get());
        return buf(arrayIdx);
    }
    void appendIBatchBuf(SharedIBatchBuffer buf) final
    {
        SharedBatchBuf castBuf =
            std::dynamic_pointer_cast<BaseBatchBuffer>(buf);
        assert(castBuf);
        addBuf(castBuf);
    }
    const IOptions* getOptions() const final { return m_Options.get(); }
    SharedIOptions getSafeOptions() const { return m_Options; }
    void setIOptions(SharedIOptions o) final { setOptions(o); }
 
    const std::vector<SharedBatchBuf>& bufs() const { return m_Bufs; }
    std::vector<SharedBatchBuf>& mutableBufs() { return m_Bufs; }
    void addBuf(SharedBatchBuf buf) { assert(buf); m_Bufs.emplace_back(std::move(buf)); }
    const SharedBatchBuf& buf(uint32_t idx) const
    {
        assert(idx < (uint32_t)m_Bufs.size());
        return m_Bufs.at(idx);
    }
    SharedBatchBuf& buf(uint32_t idx)
    {
        assert(idx < (uint32_t)m_Bufs.size());
        return m_Bufs.at(idx);
    }
    const SharedCuEvent& cuEvent() const { return m_CuEvent; }
    void setCuEvent(SharedCuEvent e) { m_CuEvent = std::move(e); }
    int findFirstGpuId() const {
        int gpuId = -1;
        auto iter = std::find_if(m_Bufs.cbegin(), m_Bufs.cend(),
            [&gpuId](const SharedBatchBuf& buf) {
                assert(buf);
                const InferBufferDescription& desc = buf->getBufDesc();
                if (desc.memType == InferMemType::kGpuCuda) {
                    gpuId = desc.devId;
                    return true;
                }
                return false;
            });
        (void) iter;
 
        assert((iter != m_Bufs.cend() && gpuId >= 0) ||
               (iter == m_Bufs.cend() && gpuId == -1));
        return gpuId;
    }
 
    void setBufId(uint64_t id) { m_BufId = id; }
    uint64_t bufId() const { return m_BufId; }
 
    // set options
    void setOptions(SharedIOptions o) { m_Options = std::move(o); }
 
private:
    std::vector<SharedBatchBuf> m_Bufs;
    SharedCuEvent m_CuEvent;
    uint64_t m_BufId = UINT64_C(0);  // for debug and track
 
    SharedOptions m_Options;
};
 
extern void normalizeDims(InferDims& dims);
 
class RefBatchBuffer : public BaseBatchBuffer {
public:
    RefBatchBuffer(
        void* bufBase, size_t bufBytes, const InferBufferDescription& desc,
        uint32_t batchSize)
        : BaseBatchBuffer(batchSize), m_BufBase(bufBase), m_BufBytes(bufBytes)
    {
        setBufDesc(desc);
        normalizeDims(mutableBufDesc().dims);
    }
    void* getBufPtr(uint32_t batchIdx) const override {
        assert(batchIdx == 0 || batchIdx < getBatchSize());
        const InferBufferDescription& desc = getBufDesc();
        assert(batchIdx <= 0 || desc.dataType != InferDataType::kString);
        batchIdx = std::max(batchIdx, 0U);
        return (void*)((uint8_t*)m_BufBase +
                       desc.elementSize * desc.dims.numElements * batchIdx);
    }
    uint64_t getTotalBytes() const final { return m_BufBytes; }
    void* basePtr() { return m_BufBase; }
 
private:
    mutable void* m_BufBase = nullptr;
    size_t m_BufBytes = 0;
};
 
class WrapCBatchBuffer : IBatchBuffer {
public:
    template <typename BufPtr>
    WrapCBatchBuffer(BufPtr buf) : m_Impl(std::move(buf)) {
        assert(m_Impl);
    }
 
    ~WrapCBatchBuffer() final = default;
    const InferBufferDescription& getBufDesc() const final {
        return m_Impl->getBufDesc();
    }
    void* getBufPtr(uint32_t batchIdx) const final {
        return m_Impl->getBufPtr(batchIdx);
    }
    uint32_t getBatchSize() const final { return m_Impl->getBatchSize(); }
    uint64_t getTotalBytes() const final { return m_Impl->getTotalBytes(); }
 
private:
    SharedBatchBuf m_Impl;
};
 
} // namespace nvdsinferserver
 
#endif