cuda内核的配置参数

Question

I have to add two square matrices of N x N using cuda program . The book asks to write the configuration parameters for the kernel for the cases :

(a) Each thread must process only 1 matrix element

(b) Each thread producing one output matrix row

(c) Each thread producing one output matrix column

My solutions for the above :

(a)

dim3 threadPerBlocks(1,1,1);
dim3 numBlocks(N,N,1);

(b)

dim3 threadPerBlocks(N,1,1);
dim3 numBlocks(1,N,1);

(c)

dim3 threadPerBlocks(1,N,1);
dim3 numBlocks(N,1,1);

I have no idea whether I am right or wrong for parts (b) and (c) . Please tell me about those and give a brief explanation about them ( if they are wrong , please correct me and explain ) .

Answer 1

(a) is somewhat fine but you can write in different ways.. All its required is you need to have N x N threads so each processes one element.

Alternative for (a) is

dim3 threadPerBlocks(N,1,1);
dim3 numBlocks(N,1,1);

And in kernel you process as

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

array[id] = ... ; // process one element.

But for (b) it says you need to each thread producing one out matrix row so you need only N or number of columns number of threads. What you have written with that you will still end up with N x N threads.

So you can write this way. One of the possible way there are other ways too.

dim3 threadPerBlocks(N,1,1);
dim3 numBlocks(1,1,1);

idx = threadIdx.x ; 

Then you use a for loop to process 1 row in each thread.

for (i = 0 ; i < N ; i++)
{
    index = idx * N + i ;
    array [index] = ..... ;   
}

Similarly you can think for (c) case.