Can input matrices also be used to store the output matrix with CUBLAS?

Question

For instance, cublas<t>geam() will do:

But what if I want to store the result in A anyway? Can I call it with pointers *C = *A so that:

without fear that I may be writing output to a matrix still being read as an input??

If so, are guaranteed that we do this with all other CUBLAS matrix operations safely?

Answer 1

The answer is alright in the cublas documentation, as you posted it - cublasgeam() - If C overlaps A or B, then behaviour is undefined. Nvidia won't guarantee that this will work, if C == A.