cublasXt ................ copy matrix to device explicitly

Hi there,

I am wondering whether there is any option in cublasXt to copy the “A” array of dgemm to the devices explicitly and retain it on the device for repeated multiplications.

I looked at the documentation and it appears as if cublasXt has very limited user interaction and all the copying to and from the device is taken care for by the actual multiplication call. If that is correct, this is superficially pretty handy but will cause a huge unnecessary overhead if say array “A” of dgemm is used repeatedly during an itertative process.

Can somebody confirm whether cublasXt is really that limited or whether I have missed something.