Error with sparse matrix vector multiplication in HYB format in cusparse

Accelerated Computing GPU-Accelerated Libraries
1

Hi! I’m trying to do two operations y=Ax and x=A^Ty, where A -size MxN, x -size N, y -size M. A sparse in HYB format(csr to hyb - matrix hybA) cusparseDcsr2hyb. Using cusparseDhybmv for y=A*x works, but cusparseDhybmv don’t work with transpose of A. As with cusparseDcsrmv I convert csr to csc, then csc to hyb - hybAt matrix, and use cusparseDhybmv with hybAt - this does not work! Part of the code

cusparseDcsr2hyb(cuhandle, m, n, descr,d_vals, d_ptr,d_col, hybA, userEllWidth, CUSPARSE_HYB_PARTITION_AUTO); work

cusparseDcsr2csc(cuhandle, m, n, nnz,d_vals,d_ptr,d_col, d_valst, d_colt, d_ptrt,CUSPARSE_ACTION_NUMERIC, CUSPARSE_INDEX_BASE_ZERO); work

cusparseDcsc2hyb(cuhandle, n, m, descr,d_valst, d_ptrt,d_colt, hybAt, userEllWidth1, CUSPARSE_HYB_PARTITION_AUTO); work

cusparseDhybmv(cuhandle, CUSPARSE_OPERATION_NON_TRANSPOSE, &alpha, descr, hybA, d_x,&beta, d_y); y=A*x work

cusparseDhybmv(cuhandle, CUSPARSE_OPERATION_NON_TRANSPOSE, &alpha, descr, hybAt, d_y,&beta, d_x); x=A^T*y - oes not work!

How can i do x=A^T*y in HYB format?

2

I solved the problem. I have to keep transpose of matrix A in csr firmst, then csr2hyb and then Dhybmv. But use scr2csc, then csc2hyb and Dhybmv don’t work!