0

我正在编写快速的 JPEG 读取代码，我将把它用作一个更大项目的一部分。我决定将 CUDA 和 NPP 一起用于该任务，因为 NPP 已经实现了所有编码和解码功能。一切正常，直到我尝试使用nppiDCTQuantInv8x8LS_JPEG_16s8u_C1R_NEW函数运行逆 DCT。这似乎破坏了代码的完整性。运行 DCT 后，报告了几个 cudaFree 调用cudaErrorLaunchFailure。使用 NSIGHT CUDA 调试器后，我可以看到启动 IDCT 函数报告CUDA Grid launch failed错误。可能的原因是什么？如果我不使用 NSIGHT CUDA 调试器 IDCT 函数以NPP_NO_ERROR，但仍会损坏设备指针。我附上了我认为相关的代码片段，但我可以根据要求提供更多。我有一种感觉，我可能在某些时候对指针感到困惑。不过，我已经花费了大量时间在 vs 调试器中检查和调试主机端内存。

实际IDCT部分：

void CJPEGDecoder::InverseDCT(CJPGFile* file, NppiDCTState* pDCTState, CJPEGDeviceData* dataNPP)
{
    cudaError_t quantError, huffmanError, quantAllocError;
    NppStatus DCTstatus;
    Npp8u* deviceQuantizationTables;

    quantAllocError = cudaMalloc(&deviceQuantizationTables, 64 * file->m_quantizationTables.size());
    for (int i = 0; i < file->m_quantizationTables.size(); i++)
    {
        quantError = cudaMemcpyAsync(deviceQuantizationTables + i * 64, file->m_quantizationTables.at(i).aTable, 64, cudaMemcpyHostToDevice);
    }
    for (int i = 0; i < m_numComponent; i++)
    {
        int blockHeight = dataNPP[i].m_srcSize.height / 8;
        huffmanError = cudaMemcpyAsync(dataNPP[i].m_devDCT, m_hostDCT[i], dataNPP[i].m_DCTStep*blockHeight, cudaMemcpyHostToDevice);
    }

    // Inverse DCT
    for (int i = 0; i < m_numComponent; i++)
    {
        DCTstatus = nppiDCTQuantInv8x8LS_JPEG_16s8u_C1R_NEW(dataNPP[i].m_devDCT, dataNPP[i].m_DCTStep,
            dataNPP[i].m_srcImage, dataNPP[i].m_srcImageStep,
            deviceQuantizationTables + file->m_frameHeader.quantizationSelector[i] * 64,
            dataNPP[i].m_srcSize,
            pDCTState);
    }

    cudaFree(deviceQuantizationTables);
}

释放霍夫曼表时报告错误：

void CJPEGDecoder::HuffmanDealloc()
{
    NppStatus DCerror, ACerror;
    cudaError_t error;
    for (int i = 0; i < m_numComponent; i++)
    {
        DCerror = nppiDecodeHuffmanSpecFreeHost_JPEG(apHuffmanDCTable[i]); //NPP_OK
        ACerror = nppiDecodeHuffmanSpecFreeHost_JPEG(apHuffmanACTable[i]); //NPP_OK
        error = cudaFreeHost(m_hostDCT[i]); // cudaErrorLaunchFailure if DCT was launched, cudaSuccess otherwise
    }
}

销毁 CJPEGDeviceData 时报告的错误：

void CJPEGDeviceData::ClearData()
{
    cudaError_t errorDCT, errorImg;
    m_DCTStep = 0;
    m_srcImageStep = 0;

    errorDCT = cudaFree(m_devDCT); // cudaErrorLaunchFailure if DCT was launched, cudaSuccess otherwise
    errorImg = cudaFree(m_srcImage); // cudaErrorLaunchFailure if DCT was launched, cudaSuccess otherwise

    m_allocated = false;
}

dct 计算的实际调用及其周围环境：

void CJPEGWrapper::DecodeJPG()
{
    int numComponents = m_JPGFile->m_frameHeader.numberOfComponents;
    m_deviceData = new CJPEGDeviceData[numComponents];

    uint8_t maxV{ 0 }, maxH{ 0 };
    for (int i = 0; i < numComponents; i++)
    {
        uint8_t testH = m_JPGFile->m_frameHeader.samplingFactor[i] & 0x0F;
        uint8_t testV = m_JPGFile->m_frameHeader.samplingFactor[i] >> 4;
        if (testH > maxH)
            maxH = testH;
        if (testV > maxV)
            maxV = testV;
    }

    m_JPGdecoder.SetImgSize(m_JPGFile->m_frameHeader.width, m_JPGFile->m_frameHeader.height,numComponents);
    m_JPGdecoder.SetMaxMCUSize(maxH, maxV);

    for (int i = 0; i < numComponents; i++)
    {
        m_JPGdecoder.DecodeMCU(m_JPGFile->m_frameHeader.samplingFactor[i],m_deviceData[i]);
    }

    m_JPGdecoder.HuffmanAlloc(m_JPGFile);
    m_JPGdecoder.HuffmanDecode(m_JPGFile, m_deviceData);
    m_JPGdecoder.InverseDCT(m_JPGFile, m_pDCTState, m_deviceData); // IDCT is launched here
    m_JPGdecoder.HuffmanDealloc();
}

CJPEGDeviceData 类：

class CJPEGDeviceData
{
public:
    NppiSize m_blockSize;
    NppiSize m_srcSize;
    Npp16s* m_devDCT;
    Npp32s m_DCTStep;

    Npp8u* m_srcImage;
    Npp32s m_srcImageStep;
public:
    CJPEGDeviceData();
    CJPEGDeviceData(const CJPEGDeviceData& object);
    CJPEGDeviceData(CJPEGDeviceData&& object);
    ~CJPEGDeviceData();

    void AllocDevicePointers(NppiSize blocksSize);
    void ClearData();
    bool IsAllocated() const;
private:
    bool m_allocated;
};

谁能帮助我了解发生了什么以及我做错了什么？即使我启动有问题的 IDCT 部分，cuda-memcheck 也不会报告任何错误，我只能在 VS 调试器中检测到错误。我相信读取文件本身可以正常工作，我已经对其进行了许多测试，因此初始数据应该没问题。问题始于设备数据。我还可以补充一点，在打开 IDCT 的情况下启动 CUDA 分析器会使应用程序崩溃并引发非零退出代码错误。否则它运行良好。

4

0 回答 0