Backslash operator on GPU

Hello,

I need to perform a backslash operation (A \ B = x) to solve Ax = B.

The A matrix has 10^4 - 10^6 D.O.F, so a square matrix of 10^6 x 10^6. It is very sparse, with approx. 21 non-zero entries per row.

Because of its sparseness, I can reach very large matrices on the CPU, by using the sparse() form. However, when I try to use the \ operator on gpuArrays, it says that I cannot have a sparse matrix as a gpuArray variable. This therefore limits me on the card to about ~8000x8000 square matrix.

The increase in performance on the GPU is substantial, above >1000x1000 square matrix, and is absolutely necessary for 10^6 x 10^6 matrices.

Is there any way I can input sparse matrices on the GPU? Any other way of accelerating such a computation? Any suggestions?

Cheers,

Francesco

Answer 1

[Edric Ellis]

Sparse arrays are now supported on the GPU, here's a page describing the support: https://www.mathworks.com/help/parallel-computing/work-with-sparse-arrays-on-a-gpu.html . This support was added several years ago (but was not available at the time the original question was posed).

[Francesco]

Would writing a CUDA native program using cuSPARSE to solve the equation work? Would the input of that kernel when called by matlab have to consist of full or sparse matrices?

I could perhaps pre-process on Matlab the full matrix into a more compact form required by cuSPARSE but still make it "look" like a full matrix when I pass it to the cuda kernel, if that makes sense?

Thank you,

Francesco