如何在CUDA中实现与子矩阵的接口？

Question

I have a wrapper class CudaMatrix that implements several cuBLAS operations, allowing me to call m1.multiply(m2) that runs the sgemm operation on the internal data pointers.

I would like to extend the class by operations on sub-matrices, something like

CudaMatrix a(100,100);
CudaMatrix b(100,100);
// fill a and b

int i=5, j=15;
CudaSubMatrix sa(a, i, j, i+10, j+10); // sa := a[5:15, 15:25]

i=50, j=60;
CudaSubMatrix sb(b, i, j, i+10, j+10); // sb := b[50:60, 60:70]    

CudaMatrix res;
res.copy(sa);
res.multiply(sb)  // res = sa*sb

In the last row, multiply() needs to operate on a sub-matrix sb , so the rows are not contiguous and I can't call the same sgemm operations as before.

How do I implement an efficient interface to sub-matrices that avoids copying data explicitly? Are there any open-source implementations that I can look for?

Answer 1

The sub-matrices multiply may be performed using the ldx parameter of the API calls.

Indexing is described at the 1.1 DataLayout section:

#define IDX2C(i,j,ld) (((j)*(ld))+(i))

Then use the cublasSgemm for example with lda parameter equal to the number of lines

the cuBLAS library uses column-major storage

of the original matrix, and m , n , k for the sub-matrices.

Note indexing might differ in fortran for C indexing scheme.

Hence what you really need is the size of your sub-matrix (col,rows), and the size of a column in the input matrix (its number of lines).