Square matrix factorization to increase sparsity for GPU use

Dear All,

I have a full (no zeros) real asymmetric square matrix A in a loop on the GPU (to increase execution speed compared with the cpu), which limits the use of the code for smaller A sizes that I need. Is there, and if so, which one, an efficient factorization type for matrix A (full, asymetric, real) leading to sparser output f(A)? When I say sparser, I am not referring to the sparsity of individual output matrices (which may be increased, say with triagular or identity matrix factors), but the sparsity of the whole factorization output f(A), so when taken elementwise f(A) is consistently less memory costly than A. Thank you,

Octavian