problem of using cusparseScsrmm for big matrix

I came across a problem when using cusparse lib for matrix multiplication.
everything goes well when using a small matrix (80180), but something goes wrong when increase the matrix to (610009003).
the problem is: cusparseScsrmm operation can return CUSPARSE_STATUS_SUCCESS, but after that, when I access GPU memory again, an ‘illegal memory access’ error came out.
does anyone can help me?
btw, cuda memory is enough large.