Question (image processing code)

Hello,

I have a grey level image (resolution w x h) stored as “dev_im_l” on the device. I want to calculate average values of a 5x5 neighborhood around each pixel and store them in “dev_luminance”. My plan is to store each pixel plus its 24 neighbors in shared memory with each pixel having a block.

I’m new in Cuda, so even if I made beginner mistakes, I appreciate your help.

#define PATCH_SIZE 5

__global__ void compute_luminance(uchar *dev_im_l, float *dev_luminance, int w, int h)
{
		__shared__ uchar patch[PATCH_SIZE*PATCH_SIZE];
	int patch_index = blockIdx.x*blockDim.x + threadIdx.x;
	int pix_index = threadIdx.x;
	int y = (blockIdx.x*blockDim.x + threadIdx.x) / w;
	int x = (blockIdx.x*blockDim.x + threadIdx.x) % w;

	//center row of patch
	int row = threadIdx.x / PATCH_SIZE;
	if (y >= 2 && y < h - 2 && x >= 0 && x < w - 2)
	{
		if (threadIdx.x >= (PATCH_SIZE / 2)*PATCH_SIZE & threadIdx.x < (PATCH_SIZE / 2 + 1)*PATCH_SIZE)
			patch[pix_index] = dev_im_l[patch_index];
		else
			patch[pix_index] = dev_im_l[patch_index + w*(row - PATCH_SIZE / 2)];
	}
	__syncthreads();
	dev_luminance[patch_index] = 0;
	for (int i = 0; i < PATCH_SIZE*PATCH_SIZE; i++)
		dev_luminance[patch_index] += (float)((int)patch[i]);
	dev_luminance[patch_index] /= (float)(PATCH_SIZE*PATCH_SIZE);}

int main()
{
...
	compute_luminance << <(w*h + PATCH_SIZE*PATCH_SIZE - 1) / (PATCH_SIZE*PATCH_SIZE), PATCH_SIZE*PATCH_SIZE >> >(dev_im_l, dev_im_r, dev_luminance, dev_contrast, dev_d, w, h);
...
        return 0;
}

here is code using texture references:

Furthermore, NPP provides also a box filter routine

to bind 2D textures, how should the argument cudaChannelFormatDesc look like?