Unexpected bugs when using nvjpeg to decode image

Accelerated Computing GPU-Accelerated Libraries
,
1

To accelerate the whole model inference pipeline, I have tried to use nvjpeg to decode my entire datasets. Followed on the official samples provided at here, my code have been written as follows:

       // create library handle
        nvjpegJpegState_t nvjpeg_state;
        nvjpegHandle_t nvjpeg_handle;
        nvjpegCreateSimple(&nvjpeg_handle);//step 1
        // create bitstream object
        nvjpegJpegStateCreate(nvjpeg_handle, &nvjpeg_state);//step 2
       //init
        std::vector<nvjpegImage_t> out;out.resize(img_names.size());
        for (int i = 0; i < out.size(); i++) {
            for (int c = 0; c < NVJPEG_MAX_COMPONENT; c++) {
            out[i].channel[c] = NULL;
            out[i].pitch[c] = 0;
            }
        }
       int widths[NVJPEG_MAX_COMPONENT];
       int heights[NVJPEG_MAX_COMPONENT];
       int channels;
       nvjpegChromaSubsampling_t subsampling;

       for(int i = 0; i < img_names.size(); i++) {
            std::tuple<size_t, std::vector<char>> raw_len_data = read_next_batch(img_names[i]);
            auto raw_len = std::get<0>(raw_len_data);
            auto raw_data = std::get<1>(raw_len_data);
            std::cout<<raw_len<<std::endl;
            assert(raw_data.size()>0);
            NVJPEG_CHECK(nvjpegGetImageInfo(nvjpeg_handle, (unsigned char *)raw_data.data(), raw_len, &channels, &subsampling, widths, heights));//step 3
                
            int ori_width = widths[0];
            int ori_height = heights[0];
            std::cout<<"width: "<<ori_width<<"  ";
            std::cout<<"height: "<<ori_height<<std::endl;
    

            NVJPEG_CHECK(nvjpegDecode(nvjpeg_handle, nvjpeg_state,
                                            (unsigned char *)raw_data.data(),
                                            raw_len, NVJPEG_OUTPUT_BGRI, &out[i],
                                            mCudaStream));//step 4
            
            assert(out[i].channel[0] !=nullptr);
}

However, the compiler indicates that "assertion failed",which means out[i].channel[0] is equal to nullptr. At the same time, the image info was generated successfully provided by this functionnvjpegGetImageInfo" .

I have been struggling with this problem for a couple of days. Any hints or solutions would be highly appreciated,

2

Hi hantengfei013,

We take a look in the code snippet shared and found that out[i].channel[0] not allocated properly.

It is the applications responsibility to make sure that the pointers in nvjpegImage_t are allocated using cudaMalloc and need to allocate output memory before calling nvjpegDecode().

Sample example for allocating memory for all batch images here -

Hope this will help.