[英]nvJPEG: encode packed BGR
好吧,我的目標很簡單——嘗試使用打包/交錯的 BGR 數據(也可以是 RGB)從緩沖區創建 JPEG 編碼圖像。
NVidia 文檔包含一個示例, 此處基本上描述了正確的圖像輸入。
所以我嘗試了以下方法:
#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;
}
記錄器說nvjpegEncodeImage
只返回NVJPEG_STATUS_INVALID_PARAMETER
,這意味着沒有任何作用。 如果您懷疑my_bitmap_type
填寫錯誤,這里是類似的 turbojpeg 驅動的編碼:
#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 它工作得很好。
我迫切地想弄清楚代碼中缺少什么。 將不勝感激任何幫助或建議。
UPD:使用 CentOS 7 / libnvjpeg-11-1.x86_64 (CUDA 11.1) / gcc 4.8.5
好吧,這有點奇怪,但經過一段時間的反復試驗后,發現 NVidia 文檔缺少基本細節:
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);
盡管文檔清楚地表明 state JPEG 壓縮的默認子采樣為 4:4:4,但編碼不適用於默認編碼器參數,必須明確設置子采樣。
因此,一行代碼解決了所有問題。
聲明:本站的技術帖子網頁,遵循CC BY-SA 4.0協議,如果您需要轉載,請注明本站網址或者原文地址。任何問題請咨詢:yoyou2525@163.com.