基本/简单公式，用于计算CUDA内核中所需的块数

Question

i was having some difficulty with getting the correct number of blocks per grid in CUDA. can anyone show basic/simple formula to compute number of blocks needed in CUDA kernel for 2D one? (ie gridDim.x and gridDim.y) given the fact that user wants to run N total number of threads, and his blocks are A by B (where A*B<=512 or 1024 depending on compute capability), or for simple case let's assume his blocks are 8 by 8. thanks again. also can you point which things we have to keep in my mind, for example does it really matter whether they are powers of two or not...

 dim3 dimBlock(A,B);
 dim3 dimGrid(Z,T);

i am looking for Z and T. thanks!!!

Answer 1

The total number of threads N is calculated by

N = No_blocks * No_threads_per_block

So you have A*B threads per block, so you should have Z*T=N/(A*B) .

Z and T should be integers of course, but don't need to be powers of 2. Also, depends on your CUDA compute capability, there are limitations on the number of blocks in each dimension of the grid.