Question

Say I have a 1D array converted from a MxN 2D matrix, and I want to parallelize each column and do some operations. How do I assign a thread to each column?

For example, if I have a 3×3 matrix:

1  2  3

4  5  6

7  8  9

And I want to add each number in the column depending on the column # (so 1st column will add 1, 2nd will add 2….), it then becomes:

1+1   2+1   3+1

4+2   5+2   6+2

7+3   8+3   9+3

How do I do this in CUDA? I know how to assign threads to all the elements in the array but I don’t know how to assign thread to each column. So, what I want is to send each column (1 , 2 ,3 ) ( 4 , 5 ,6 ) (7 , 8 ,9) and do the operation.

Answer 1

In your example you are adding numbers based on the row. Still, you know the row/column length of the matrix (you know it’s MxN). What you could do is something like:

__global__ void MyAddingKernel(int* matrix, int M, int N)
{

    int gid = threadIdx.x + blockDim.x*blockIdx.x;
    //Let's add the row number to each element
    matrix[ gid ] += gid % M;
    //Let's add the column number to each element
    matrix[ gid ] += gid % N;

}

If you wanted to add a different number, you could do something like:

matrix[ gid ] += my_col_number_function(gid%N);