nvJPEG: encode packed BGR - c++

Question

Well, my goal is simple -- trying to create a JPEG encoded image from buffer with packed/interleaved BGR data (could be RGB as well).
The NVidia docs contain an example, the proper image input is essentially described here.
So I tried the following:
#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;
}
The logger says that nvjpegEncodeImage just returns NVJPEG_STATUS_INVALID_PARAMETER, meaning nothing works. In case you suspect my_bitmap_type to be filled wrong, here's the similar turbojpeg-powered encoding:
#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;
}
... aaand it works pretty fine.
I'm desperate trying to figure out, what's missing in the code. Would gratefully appreciate any help or advice.
UPD:
Using CentOS 7 / libnvjpeg-11-1.x86_64 (CUDA 11.1) / gcc 4.8.5

Answer 1

Okaaay, that's strange a bit, but after some time spent on trial and error it occured that NVidia docs lack an essential detail:
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);
Although the docs clearly state that the default subsampling for JPEG compression is 4:4:4, encoding's not working with the default encoder params, subsampling has to be explicitly set.
So, that one line of code fixes everything.