Improve runtime using gpu matrix calculations

[phbasler]

Using gpu matrix calculations, e.g. https://github.com/rudaoshi/gpumatrix or other librarys, we could try to improve our runrime exporting the matrix calculations on the gpu instead of cpu.

We first need to find a suitable library which would work with our code.