c++ nvJPEG:编码压缩BGR

yizd12fk  于 2022-12-24  发布在  其他
关注(0)|答案(1)|浏览(392)

好吧,我的目标很简单--尝试从带有压缩/交错BGR数据(也可以是RGB)的缓冲区创建JPEG编码图像。
NVidia文档包含一个示例,此处基本描述了正确的图像输入。
所以我尝试了以下方法:

#include <nvjpeg.h>

// very simple
typedef struct {
    int width;
    int height;
    unsigned char *buffer; 
    unsigned long data_size; 
} my_bitmap_type;

std::vector<unsigned char> BitmapToJpegCUDA(const my_bitmap_type *image) 
{
  nvjpegHandle_t nv_handle;
  nvjpegEncoderState_t nv_enc_state;
  nvjpegEncoderParams_t nv_enc_params;
  cudaStream_t stream = NULL;

  nvjpegStatus_t er;
  nvjpegCreateSimple(&nv_handle);
  nvjpegEncoderStateCreate(nv_handle, &nv_enc_state, stream);
  nvjpegEncoderParamsCreate(nv_handle, &nv_enc_params, stream);

  nvjpegImage_t nv_image;
  nv_image.channel[0] = image->buffer;
  nv_image.pitch[0] = 3 * image->width;

  // Nope, that's for planar images!

  // nv_image.channel[0] = image->buffer;
  // nv_image.channel[1] = image->buffer + image->width * image->height;
  // nv_image.channel[2] = image->buffer + 2 * image->width * image->height;
  // nv_image.pitch[0] = image->width;
  // nv_image.pitch[1] = image->width;
  // nv_image.pitch[2] = image->width;

  er = nvjpegEncodeImage(nv_handle, nv_enc_state, nv_enc_params, &nv_image,
                        NVJPEG_INPUT_BGRI, image->width, image->height, stream);
  LOG(ERROR) << "enc " << er;

  size_t length = 0;
  nvjpegEncodeRetrieveBitstream(nv_handle, nv_enc_state, NULL, &length, stream);

  cudaStreamSynchronize(stream);
  std::vector<unsigned char> jpeg(length);
  nvjpegEncodeRetrieveBitstream(nv_handle, nv_enc_state, jpeg.data(), &length, 0);

  nvjpegEncoderParamsDestroy(nv_enc_params);
  nvjpegEncoderStateDestroy(nv_enc_state);
  nvjpegDestroy(nv_handle);

  return jpeg;
}

日志记录器显示nvjpegEncodeImage只返回NVJPEG_STATUS_INVALID_PARAMETER,这意味着没有任何结果。如果您怀疑my_bitmap_type填错了,下面是类似的turbojpeg支持的编码:

#include <turbojpeg.h>

std::vector<unsigned char> BitmapToJpegBuffer(const my_bitmap_type *image)
{
    std::vector<unsigned char> out_data(3 * image->width * image->height);

    cudaError_t err = cudaMemcpy(out_data.data(), image->buffer, image->data_size, cudaMemcpyDeviceToHost);
    if (cudaSuccess != err) {
        LOG(ERROR) << "failed to copy CUDA memory: " << err;
    }

    tjhandle jpeg = tjInitCompress();
    unsigned char *encoded_buf = nullptr;
    long unsigned int encoded_sz = 0;

    int tjres = tjCompress2(jpeg,
                            out_data.data(),
                            image->width,
                            image->width * 3,
                            image->height,
                            TJPF_BGR,
                            &encoded_buf,
                            &encoded_sz,
                            TJSAMP_444,
                            95,
                            TJFLAG_FASTDCT);

    if (tjres != 0) {
        LOG(ERROR) << "jpeg compession failed!";
        return {};
    }

    std::vector<unsigned char> result(encoded_buf, encoded_buf + encoded_sz);
    tjFree(encoded_buf);
    tjDestroy(jpeg);

    return result;
}

......而且效果很好。
我绝望地试图找出,什么是在代码中丢失。将不胜感激的任何帮助或建议。

    • 统一产品开发:**使用操作系统中心7/libnvjpeg-11 - 1.x86_64(CUDA 11.1)/gcc 4.8.5
rqqzpn5f

rqqzpn5f1#

好吧,这有点奇怪,但经过一段时间的试验和错误,它发生了NVidia的文档缺乏一个基本的细节:

nvjpegCreateSimple(&nv_handle);
    nvjpegEncoderStateCreate(nv_handle, &nv_enc_state, stream);
    nvjpegEncoderParamsCreate(nv_handle, &nv_enc_params, stream);

    // This has to be done, default params are not sufficient
    nvjpegEncoderParamsSetSamplingFactors(nv_enc_params, NVJPEG_CSS_444, stream);

虽然文档明确指出JPEG压缩的默认二次采样是4:4:4,但编码不能使用默认的编码器参数,二次采样必须显式设置。
所以,这一行代码修复了一切。

相关问题