How to solve a system of linear equation(matrix of the order of 10000*100000) using CUBLAS

[jeebujacob]

Hello,

Which is the best method to solve a system of linear equation using CUBLAS library? My matrix size comes of the order of upto 100,000*100,000(diagonally dominant) .

[CudaaduC]

you would have to use the sparse library cuSparse, as that is too big to use cuBLAS in dense format

[jeebujacob]

I tried using iterative method(Jacobi Solver) using cublasdgemv() and a separate kernel call for division using diagonal elements, just as the equation of the Jacobi solver dictates… But this system is very inefficient. When iterations are of the order of thousands, cudamemcpy statements to get the temporary solutions are consuming lot of time.

[CudaaduC]

@jeebujacob

I tried using iterative method(Jacobi Solver) using cublasdgemv() and a separate kernel call for division using diagonal elements, just as the equation of the Jacobi solver dictates… But this system is very inefficient. When iterations are of the order of thousands, cudamemcpy statements to get the temporary solutions are consuming lot of time.

Yes, that is why I recommended using cuSPARSE, assuming of course there is more than a trivial amount of zeros in the matrices.

http://docs.nvidia.com/cuda/cusparse/#axzz3GLYSG7Pb

look through that documentation in particular the Level 3 functions