How to render images to /dev/video0 using v4l2loopback?

I've been trying to render images to /dev/video. I can get something to sort of display but it's somewhat scrambled.

I first started off trying to render a normal RGB24 image (based off this example https://stackoverflow.com/a/44648382/3818491), but the result (below) was a scrambled image.

#include <stdio.h>
#include <fcntl.h>
#include <errno.h>
#include <unistd.h>
#include <iostream>
#include <sys/ioctl.h>
#include <linux/videodev2.h>

#include <CImg.h>

#define VIDEO_OUT "/dev/video0" // V4L2 Loopack

#define WIDTH  1280
#define HEIGHT 720

int main() {
    using namespace cimg_library;

    CImg<uint8_t> canvas(WIDTH, HEIGHT, 1, 3);
    const uint8_t red[] = {255, 0, 0};
    const uint8_t purple[] = {255, 0, 255};

    int fd;
    if ((fd = open(VIDEO_OUT, O_RDWR)) == -1) {
      std::cerr << "Unable to open video output!\n";
      return 1;
    }

    struct v4l2_format vid_format;
    vid_format.type = V4L2_BUF_TYPE_VIDEO_OUTPUT;

    if (ioctl(fd, VIDIOC_G_FMT, &vid_format) == -1) {
      std::cerr << "Unable to get video format data. Errro: " << errno << '\n';
      return 1;
    }

    size_t framesize = canvas.size();
    int width = canvas.width(), height = canvas.height();

    vid_format.fmt.pix.width       = width;
    vid_format.fmt.pix.height      = height;
    vid_format.fmt.pix.pixelformat = V4L2_PIX_FMT_RGB24;
    vid_format.fmt.pix.sizeimage   = framesize;
    vid_format.fmt.pix.field       = V4L2_FIELD_NONE;

    if (ioctl(fd, VIDIOC_S_FMT, &vid_format) == -1) {
      std::cerr << "Unable to set video format! Errno: " << errno << '\n';
      return 1;
    }

    std::cout << "Stream running!\n";
    while (true) {
      canvas.draw_plasma();
      canvas.draw_rectangle(
        100, 100, 100 + 100, 100 + 100, red, 1);
      canvas.draw_text(5,5, "Hello World!", purple);
      canvas.draw_text(5, 20, "Image freshly rendered with the CImg Library!", red);

      write(fd, canvas.data(), framesize);
    }
}

So I checked what (I think) /dev/video expects which seems to be YUV420P.

v4l2-ctl --list-formats-ext                                                                                                                                                                              130 ↵

ioctl: VIDIOC_ENUM_FMT
    Type: Video Capture

    [0]: 'YU12' (Planar YUV 4:2:0)
        Size: Discrete 1280x720
            Interval: Discrete 0.033s (30.000 fps)

So I attempted to convert the frame that format (using this code to quickly test).

Adjusting the spec to:

    vid_format.fmt.pix.width       = width;
    vid_format.fmt.pix.height      = height;
    vid_format.fmt.pix.pixelformat = V4L2_PIX_FMT_YUV420;
    vid_format.fmt.pix.sizeimage   = width*height*3/2; // size of yuv buffer
    vid_format.fmt.pix.field       = V4L2_FIELD_NONE;

That results in this (which seems to be from what I've gathered the structure of a yuv420 image but still rendered incorrectly).

What does /dev/video0 expect?

After a lot of hacking around, I've managed to generate a valid YUYV video/image to send to /dev/video0.

First I make a buffer to hold the frame:

// Allocate buffer for the YUUV frame
std::vector<uint8_t> buffer;
buffer.resize(vid_format.fmt.pix.sizeimage);

Then I write the current canvas to the buffer in YUYV format.

bool skip = true;
cimg_forXY(canvas, cx, cy) {
  size_t row = cy * width * 2;
  uint8_t r, g, b, y;
  r = canvas(cx, cy, 0);
  g = canvas(cx, cy, 1);
  b = canvas(cx, cy, 2);

  y = std::clamp<uint8_t>(r * .299000 + g * .587000 + b * .114000, 0, 255);
  buffer[row + cx * 2] = y;
  if (!skip) {
    uint8_t u, v;
    u = std::clamp<uint8_t>(r * -.168736 + g * -.331264 + b * .500000 + 128, 0, 255);
    v = std::clamp<uint8_t>(r * .500000 + g * -.418688 + b * -.081312 + 128, 0, 255);
    buffer[row + (cx - 1) * 2 + 1] = u;
    buffer[row + (cx - 1) * 2 + 3] = v;
  }
  skip = !skip;
}

Note: CImg has RGBtoYUV has an in-place RGB to YUV conversion, but for some reason calling this on a uint8_t canvas just zeros it.

It also has get_YUVtoRGB which (allocates and) returns a CImg<float> canvas, which I think you multiply each value by 255 to scale to a byte, however, whatever I tried that did not give the correct colour. Edit: I likely forgot the +128 bias (though I still prefer not reallocating for each frame)

My full code is here (if anyone wants to do something similar) https://gist.github.com/MacDue/36199c3f3ca04bd9fd40a1bc2067ef72