How can I properly use nppiYUV422ToRGB_8u_C2C3R()?
I am trying to convert a YUV422 image into RGB image using CPU and GPU.
[On CPU] I tried using cv::cvtColor() as shown below:
cv::cvtColor(mat_UYVY, mat_bgr, cv::COLOR_YUV2BGR_UYVY);
And it works fine. See the picture attached here.
(I removed the image because it contains somewhat private informations..)
[But on GPU] I tried using nppiYUV422ToRGB_8u_C2C3R() as shown below:
NppStatus status = nppiYUV422ToRGB_8u_C2C3R(gpu_buff_UYVY,
img_size.width*2,
gpu_buff_RGB,
img_size.width*3,
roi);
But it produces some weird image as attached here.
I know opencv works with BGR image and the image nppiYUV422ToRGB_8u_C2C3R() produces is RGB image. Right. But the problem seem more than this. I tried converting RGB to BGR but prolem remained. There is more than something...
Can someone provide me some advice? I wish there be some advice for me. Thank you!
#include <iterator>
#include <fstream>
#include <iostream>
#include "opencv2/opencv.hpp"
#include "nppdefs.h"
#include "nppi_support_functions.h"
#include "nppi_color_conversion.h"
int main()
{
cv::Size img_size(1920, 1080);
unsigned char *buff_UYVY =
new unsigned char[img_size.width * img_size.height * 2];
unsigned char *buff_RGB =
new unsigned char[img_size.width * img_size.height * 3];
//file >>> buff_UYVY
{
std::string file_uyvy("uyvy422.raw");
std::ifstream stream_uyvy;
stream_uyvy.open(file_uyvy, std::ios::in | std::ios::binary);
if (!stream_uyvy.is_open())
{
std::cerr << "[ERROR] cannot open the raw file " << file_uyvy
<< std::endl;
std::cerr << std::endl;
assert(0);
}
stream_uyvy.read((char*)buff_UYVY, img_size.width*img_size.height*2);
stream_uyvy.close();
}
//buff_UYVY >>> mat_bgr
cv::Mat mat_UYVY(img_size, CV_8UC2, buff_UYVY);
cv::Mat mat_bgr(img_size, CV_8UC3);
cv::cvtColor(mat_UYVY, mat_bgr, cv::COLOR_YUV2BGR_UYVY);
cv::imshow("BGR Image from CPU", mat_bgr);
cv::imwrite("mat_bgr.bmp", mat_bgr);
cv::imwrite("mat_bgr.jpg", mat_bgr);
//buff_UYVY >>> buff_RGB
{
Npp8u* gpu_buff_UYVY;
Npp8u* gpu_buff_RGB;
cudaError_t err_cu_api;
err_cu_api = cudaMalloc((void**)&gpu_buff_UYVY,
img_size.width*img_size.height*2);
std::cout << "cudaMalloc1 : " << err_cu_api << std::endl;
err_cu_api = cudaMemcpy((void*)gpu_buff_UYVY,
(const void*)buff_UYVY,
img_size.width*img_size.height*2,
cudaMemcpyHostToDevice);
std::cout << "cudaMemcpy2 : " << err_cu_api << std::endl;
err_cu_api = cudaMalloc((void**)&gpu_buff_RGB,
img_size.width*img_size.height*3);
std::cout << "cudaMalloc3 : " << err_cu_api << std::endl;
NppiSize roi = {img_size.width, img_size.height};
NppStatus status = nppiYUV422ToRGB_8u_C2C3R(gpu_buff_UYVY,
img_size.width*2,
gpu_buff_RGB,
img_size.width*3,
roi);
std::cout << "NppStatus : " << status << std::endl;
err_cu_api = cudaMemcpy((void*) buff_RGB,
(const void*)gpu_buff_RGB,
img_size.width*img_size.height*3,
cudaMemcpyDeviceToHost);
std::cout << "cudaMemcpy4 : " << err_cu_api << std::endl;
cudaFree(gpu_buff_UYVY);
cudaFree(gpu_buff_RGB);
}
cv::Mat mat_rgb(img_size, CV_8UC3, buff_RGB);
//cv::cvtColor(mat_rgb, mat_rgb, cv::COLOR_RGB2BGR);
std::cout << "depth : " << mat_rgb.depth() << std::endl;
std::cout << "channels : " << mat_rgb.channels() << std::endl;
std::cout << "elemSize : " << mat_rgb.elemSize() << std::endl;
std::cout << "step1 : " << mat_rgb.step1() << std::endl;
std::cout << "type : " << mat_rgb.type() << std::endl;
try {
cv::imshow("RGB Image from GPU", mat_rgb);
cv::imwrite("mat_rgb.bmp", mat_rgb);
cv::imwrite("mat_rgb.jpg", mat_rgb);
} catch( cv::Exception& e ) {
const char* err_msg = e.what();
std::cout << "exception caught #2: " << err_msg << std::endl;
}
// cv::waitKey(0);
delete[] buff_UYVY;
delete[] buff_RGB;
return 0;
}
Output message is as shown below:
cudaMalloc1 : 0
cudaMemcpy2 : 0
cudaMalloc3 : 0
NppStatus : 0
cudaMemcpy4 : 0
depth : 0
channels : 3
elemSize : 3
step1 : 5760
type : 16
Solution
I think the main problem you have is that the openCV UYVY format does not match the storage order of the NPP YUV422 format.
OpenCV UYVY storage format is: U0 Y0 V0 Y1
NPP format is: Y0 U0 Y1 V0
I wasn't able to locate any "raw" UYVY encoded image files on the internet, and you haven't provided one either. So I elected to use a synthetic image. Here is a test case:
$ cat t30.cu
#include <iterator>
#include <fstream>
#include <iostream>
#include "opencv2/opencv.hpp"
#include "nppdefs.h"
#include "nppi_support_functions.h"
#include "nppi_color_conversion.h"
int main()
{
cv::Size img_size(1920, 1080);
unsigned char *buff_UYVY =
new unsigned char[img_size.width * img_size.height * 2];
unsigned char *buff_RGB =
new unsigned char[img_size.width * img_size.height * 3];
#if 0
//file >>> buff_UYVY
{
std::string file_uyvy("uyvy422.raw");
std::ifstream stream_uyvy;
stream_uyvy.open(file_uyvy, std::ios::in | std::ios::binary);
if (!stream_uyvy.is_open())
{
std::cerr << "[ERROR] cannot open the raw file " << file_uyvy
<< std::endl;
std::cerr << std::endl;
assert(0);
}
stream_uyvy.read((char*)buff_UYVY, img_size.width*img_size.height*2);
stream_uyvy.close();
}
#endif
// create synthetic R,G,B,Black image
unsigned char r,g,b;
for (int i = 0; i < img_size.height; i++)
for (int j = 0; j < img_size.width; j+=3) {
if (j < img_size.width/3) {r = 200; g = 0; b = 0;}
else if (j < img_size.width*2/3) {r = 0; g = 200; b = 0;}
else {r = 0; g = 0; b = 200;}
buff_RGB[i*img_size.width*3+j] = r;
buff_RGB[i*img_size.width*3+j+1] = g;
buff_RGB[i*img_size.width*3+j+2] = b;}
Npp8u* gpu_buff_UYVY;
Npp8u* gpu_buff_RGB;
// use NPP to convert synthetic RGB image into NPP format YUV422
cudaError_t err_cu_api;
err_cu_api = cudaMalloc((void**)&gpu_buff_UYVY,
img_size.width*img_size.height*2);
std::cout << "cudaMalloc1 : " << err_cu_api << std::endl;
err_cu_api = cudaMalloc((void**)&gpu_buff_RGB,
img_size.width*img_size.height*3);
std::cout << "cudaMalloc3 : " << err_cu_api << std::endl;
err_cu_api = cudaMemcpy((void*)gpu_buff_RGB,
(const void*)buff_RGB,
img_size.width*img_size.height*3,
cudaMemcpyHostToDevice);
std::cout << "cudaMemcpy2 : " << err_cu_api << std::endl;
NppiSize roi = {img_size.width, img_size.height};
NppStatus status = nppiRGBToYUV422_8u_C3C2R(gpu_buff_RGB,
img_size.width*3,
gpu_buff_UYVY,
img_size.width*2,
roi);
std::cout << "NppStatus : " << status << std::endl;
err_cu_api = cudaMemcpy((void*) buff_UYVY,
(const void*)gpu_buff_UYVY,
img_size.width*img_size.height*2,
cudaMemcpyDeviceToHost);
std::cout << "cudaMemcpy4 : " << err_cu_api << std::endl;
// convert NPP format YUV422 to UYVY for use by OpenCV
for (int i = 0; i < img_size.width*img_size.height*2; i+=2){
unsigned char v1 = buff_UYVY[i];
unsigned char v2 = buff_UYVY[i+1];
buff_UYVY[i+1] = v1;
buff_UYVY[i] = v2;}
//buff_UYVY >>> mat_bgr
cv::Mat mat_UYVY(img_size, CV_8UC2, buff_UYVY);
cv::Mat mat_bgr(img_size, CV_8UC3);
cv::cvtColor(mat_UYVY, mat_bgr, cv::COLOR_YUV2BGR_UYVY);
cv::imshow("BGR Image from CPU", mat_bgr);
cv::imwrite("mat_bgr.bmp", mat_bgr);
cv::imwrite("mat_bgr.jpg", mat_bgr);
//convert UYVY OpenCV format back to NPP YUV422 format for use by NPP
for (int i = 0; i < img_size.width*img_size.height*2; i+=2){
unsigned char v1 = buff_UYVY[i];
unsigned char v2 = buff_UYVY[i+1];
buff_UYVY[i+1] = v1;
buff_UYVY[i] = v2;}
//buff_UYVY >>> buff_RGB
{
err_cu_api = cudaMemcpy((void*)gpu_buff_UYVY,
(const void*)buff_UYVY,
img_size.width*img_size.height*2,
cudaMemcpyHostToDevice);
std::cout << "cudaMemcpy4 : " << err_cu_api << std::endl;
status = nppiYUV422ToRGB_8u_C2C3R(gpu_buff_UYVY,
img_size.width*2,
gpu_buff_RGB,
img_size.width*3,
roi);
std::cout << "NppStatus : " << status << std::endl;
err_cu_api = cudaMemcpy((void*) buff_RGB,
(const void*)gpu_buff_RGB,
img_size.width*img_size.height*3,
cudaMemcpyDeviceToHost);
std::cout << "cudaMemcpy4 : " << err_cu_api << std::endl;
}
cv::Mat mat_rgb(img_size, CV_8UC3, buff_RGB);
//cv::cvtColor(mat_rgb, mat_rgb, cv::COLOR_RGB2BGR);
std::cout << "depth : " << mat_rgb.depth() << std::endl;
std::cout << "channels : " << mat_rgb.channels() << std::endl;
std::cout << "elemSize : " << mat_rgb.elemSize() << std::endl;
std::cout << "step1 : " << mat_rgb.step1() << std::endl;
std::cout << "type : " << mat_rgb.type() << std::endl;
try {
cv::imshow("RGB Image from GPU", mat_rgb);
cv::imwrite("mat_rgb.bmp", mat_rgb);
cv::imwrite("mat_rgb.jpg", mat_rgb);
} catch( cv::Exception& e ) {
const char* err_msg = e.what();
std::cout << "exception caught #2: " << err_msg << std::endl;
}
// cv::waitKey(0);
delete[] buff_UYVY;
delete[] buff_RGB;
return 0;
}
(base) [bob@localhost misc]$ nvcc -o t30 t30.cu -lnppicc -lopencv_core -lopencv_highgui -lopencv_imgcodecs -lopencv_imgproc
(base) [bob@localhost misc]$ ./t30
cudaMalloc1 : 0
cudaMalloc3 : 0
cudaMemcpy2 : 0
NppStatus : 0
cudaMemcpy4 : 0
cudaMemcpy4 : 0
NppStatus : 0
cudaMemcpy4 : 0
depth : 0
channels : 3
elemSize : 3
step1 : 5760
type : 16
The output files are like this:
mat_rgb.png:
mat_bgr.png:
The primary difference being the RGB <-> BGR mismatch that you already know about.
For the remaining color differences, its also possible that the YUV colorspace for OpenCV doesn't quite match the colorspace for NPP. There are alternative YUV functions you can try in NPP if you wish, such as nppiCbYCr422ToBGR_709HDTV_8u_C2C3R.
- Speed up nested loops and bitwise operations with Alea GPU
- Allocating CUDA Unified Memory by overloading operator new causes illegal access
- NVML Header File Missing
- Is it possible to execute multiple instances of a CUDA program on a multi-GPU machine?
- Is there any way that I can convert my cpu parallel code to the cuda?
- nvidia-smi Failed to initialize NVML: GPU access blocked by the operating system
- undefined reference to `cuCtxGetCurrent` while getting CUDA context for OptiX
- What's the capacity of a CUDA stream (=queue)?
- Error compiling CUDA
- Add Permanent Include and Library Path for CUDA C/C++ Compiler
- CUB reduce_by_key
- Do I need to redo cudaHostRegister after switching GPU devices with C++?
- Trying to understand kernel optimization in CUDA. These changes had negligable effect
- CUDA Cooperative Groups : Linking error
- How to change CUDA version
- nvlink error for C++ virtual classes when compiling CUDA with clang
- cuFFT profiling issue
- Printf() 45 arguments in CUDA
- Can we use printf or any other similar function in a CUDA Kernel?
- Cuda atomic (CAS, Exch) loop hangs
- Template __host__ __device__ calling host defined functions
- Cuda Mutex, why deadlock?
- How to trace variables with print from gpu parallel (CUDA)
- Issue with printf on CUDA GPU
- texture is not a template
- Understanting thread utilization in the CUDA reduction examples
- Why does this CUDA kernel not accurately handling non-square matrix multiplication?
- When using thrust, is it legitimate to create a std::array inside a __host__ __device__ functor?
- How to include and link against both kokkos and eigen
- In CUDA, How to do a mergesort pass one shared variable into another