| / /代码那是执行之前的内核是推出了 |
| intthreadsInX=32; |
| intthreadsInY=16; |
| intthreadsIn
= 1; |
| intblocksInX = (DATA_W + threadsInX-1) / threadsInX; |
| intblocksInY = (DATA_H + threadsInY-1) / threadsInY; |
| intblocksInZ = (DATA_D + threadsInZ-1) / threadsInZ; |
| dim3dimGrid=dim3 (blocksInX blocksInY * blocksInZ); |
| dim3dimBlock=dim3 (threadsInX threadsInY threadsInZ); |
| / /代码那是执行内部的内核 |
| intblockIdxz=__float2uint_rd (blockIdx.y*invBlocksInY); |
| intblockIdxy=blockIdx.y−blockIdxz*blocksInY; |
| intx=blockIdx.x*blockDim.x+threadIdx.x; |
| inty=blockIdxy*blockDim.y+threadIdx.y; |
| intz=blockIdxz*blockDim.z+threadIdx.z; |
