Get access to OpenCV GpuMat channels - c++

Question

I am working on a raytracer. I wanted to optimize my code by saving data of each pixel in an OpenCV Mat, using GPU.
For now, I save pixel values in buffer fb which is a vector of three values (RGB):
__global__ void render(vec3 *fb, int max_x, int max_y, Camera **cam, Triangle *data, size_t n, )
{
int i = threadIdx.x + blockIdx.x * blockDim.x;
int j = threadIdx.y + blockIdx.y * blockDim.y;
if ((i >= max_x) || (j >= max_y)) return;
int pixel_index = j * max_x + i;
float u = float(i) / float(max_x);
float v = float(j) / float(max_y);
Ray r = (*cam)->get_ray(u,v);
fb[pixel_index] = color(r, data,n);
}
and then I save data in Mat on CPU:
for (int j = ny - 1; j >= 0; j--)
{
for (int i = 0; i < nx; i++)
{
size_t pixel_index = j * nx + i;
int ir = int(255.99*fb[pixel_index].r());
int ig = int(255.99*fb[pixel_index].g());
int ib = int(255.99*fb[pixel_index].b());
output.at<Vec3b>(j, i)[0] = (uchar)ib;
output.at<Vec3b>(j, i)[1] = (uchar)ig;
output.at<Vec3b>(j, i)[2] = (uchar)ir;
//std::cout << ir << " " << ig << " " << ib << "\n";
}
}
but it is a very slow process when I have a large pixel array. That why I want to use an OpenCV GpuMat and save the data directly on the GPU.
The problem is that I can't really find an example of how I can save data in each channel of GPU Mat. Is it an easy way to do it, similar to saving data on the CPU?

Answer 1

See the documentation. There it says
no functions that return references to their data (because references on GPU are not valid for CPU)
The only way to access the data is through the datafunction. But the pointer can only be dereferenced in (cuda)kernel code. And there is no
at function as far as I see. So you will have to calculate the offset from data.

Answer 2

Thank you for your answers. They make me think how to do it in another way. I am not sure if it is the best solution to deal with it but it works and in my opinion is a quiet easy way to fill matrix on GPU.
Reserve memory on GPU for matrix
Mat output(ny, nx, CV_8UC3);
const size_t numBytes = output.step * output.rows;
unsigned char *d_output;
cudaMalloc<unsigned char>(&d_output, numBytes);
Fill matrix on GPU
_global__ void render(vec3 *fb, int max_x, int max_y, Camera **cam, Triangle *data, size_t n, unsigned char* input, int step)
{
int i = threadIdx.x + blockIdx.x * blockDim.x;
int j = threadIdx.y + blockIdx.y * blockDim.y;
if ((i >= max_x) || (j >= max_y)) return;
int pixel_index = j * max_x + i;
int index = j * step + 3 * i;
float u = float(i) / float(max_x);
float v = float(j) / float(max_y);
Ray r = (*cam)->get_ray(u,v);
fb[pixel_index] = color(r, data,n);
int ir = int(255.99*fb[pixel_index].r());
int ig = int(255.99*fb[pixel_index].g());
int ib = int(255.99*fb[pixel_index].b());
input[index] = ib;
input[index+1] = ig;
input[index+2] = ir;
}
I will be grateful for any advice and comments to this code.