True 4D Image Denoising on the GPU

Algorithm 2

//  Code  that  is  executed  before  the  kernel  is  launched

int threadsInX  =  32;

int threadsInY  =  16;

int threadsInZ = 1;

int blocksInX = (DATA_W+threadsInX-1)/threadsInX;

int blocksInY = (DATA_H+threadsInY-1)/threadsInY;

int blocksInZ = (DATA_D+threadsInZ-1)/threadsInZ;

dim3  dimGrid  =  dim3(blocksInX, blocksInY*blocksInZ);

dim3  dimBlock  =  dim3(threadsInX, threadsInY, threadsInZ);

//  Code  that  is  executed  inside  the  kernel

int blockIdxz  =  __float2uint_rd(blockIdx.y  *  invBlocksInY);

int blockIdxy  =  blockIdx.y  −  blockIdxz  *  blocksInY;

int x  =  blockIdx.x  *  blockDim.x  +  threadIdx.x;

int y  =  blockIdxy  *  blockDim.y  +  threadIdx.y;

int z  =  blockIdxz  *  blockDim.z  +  threadIdx.z;