Effective Sparse Matrix Representation for the GPU Architectures