How to convert BGRA2BGR using OpenCV Transparent API

I want to achieve colorspace conversion of a given image using available GPU on my hardware. To achieve this, we have Opencv's Transparent api's (UMat) that utilize the GPU to perform processing.

I have IMX8MP EVK kit, that has GPU support enabled.

I have a video that has BGRA format and I am reading the video using gstreamer appsrc. I want to change the colorspace conversion to BGR for further processing using opencv. I have tried using opencv's Mat with cvtColor() api which consumes lot of cpu resources. Hence, I want to perform the color conversion activity on GPU.

I convert the received Gstreamer buffer to cv::Mat. Using the code snippet.

cv::Mat image = cv::Mat(cv::Size(width, height), CV_8UC4, (char*)map.data, cv::Mat::AUTO_STEP);

I further convert the cv::Mat to cv::UMat using

cv::UMat uframe = image.getUMat(cv::ACCESS_READ);

After getting the UMat type uframe, I perform colorspace conversion.

cv::cvtColor(uframe, uframe, cv::COLOR_BGRA2BGR);

At this point, my hardware stops responding to any command as if it is in deadlock when my application is executed. I want to understand where I am going wrong.

Does UMat support BGRA2BGR conversion?
Also, will OpenCL will be utilized internally by UMat to do the processing on GPU?

Looking at the sources of ocl_cvtColor (in color.cpp), it seems like there is no support for "in-place" processing.

We can't use uframe both as input and an output as in cv::cvtColor(uframe, uframe, cv::COLOR_BGRA2BGR);.

Selecting cv::ACCESS_READ may also cause issues, when the function tries to write to uframe.

Replace cv::cvtColor(uframe, uframe_bgr, cv::COLOR_BGRA2BGR); with something like:

cv::cvtColor(uframe, uframe_bgr, cv::COLOR_BGRA2BGR);

Where uframe_bgr is another cv::UMat with cv::ACCESS_WRITE (or cv::ACCESS_RW) access flags.

Note:

• The issue is not reproducible using a PC, so we can't be sure that the solution is going to work with IMX8MP.

Sample code (map.data is replaced with map.data()):

#include <opencv2/opencv.hpp>
#include <opencv2/videoio.hpp>

int main()
{
    int width = 640;
    int height = 480;

    std::vector<uint8_t> map = std::vector<uint8_t>(width*height*4, 0); //Place holder for map buffer from the question.
    std::vector<uint8_t> map_bgr = std::vector<uint8_t>(width*height*3, 0); // Allocate buffer for storing BGR image (we don't really need it).

    //Fill the image with some content (for testing):
    cv::Mat image = cv::Mat(cv::Size(width, height), CV_8UC4, (char*)map.data(), cv::Mat::AUTO_STEP);
    image = cv::Scalar(60, 60, 60, 255);
    cv::putText(image, "Text", cv::Point(width / 2 - 70 * (int)(std::string("Text").length()), height / 2 + 100), cv::FONT_HERSHEY_DUPLEX, 8, cv::Scalar(30, 255, 30), 15);

    cv::UMat uframe = image.getUMat(cv::ACCESS_READ);
    
    cv::Mat image_bgr = cv::Mat(cv::Size(width, height), CV_8UC3, (char*)map_bgr.data(), cv::Mat::AUTO_STEP); //Create cv::Mat for storing BGR image.
    cv::UMat uframe_bgr = image_bgr.getUMat(cv::ACCESS_WRITE); //Get cv::UMat with accessFlags = cv::ACCESS_WRITE

    //cv::cvtColor(uframe, uframe_bgr, cv::COLOR_BGRA2BGR);
    cv::cvtColor(uframe, uframe_bgr, cv::COLOR_BGRA2BGR);  //Use uframe_bgr as output.

    cv::imshow("uframe_bgr", uframe_bgr);
    cv::waitKey();
    cv::destroyAllWindows();
    

    return 0;
}

The above code sample uses std::vector for allocating the memory buffers.
Replace the std::vector with the allocation you are using for map.data.