What is the most efficient way to transpose a matrix in CUDA?

Question

I have a M*N host memory matrix, and upon copying into a device memory, I need it to be transposed into a N*M matrix. Is there any cuda (cuBLAS...)

Answer 1

In the cublas API:

cublasgeam()

This function performs the matrix-matrix addition/transposition
the user can transpose matrix A by setting *alpha=1 and *beta=0.  

(and specifying the transa operator as CUBLAS_OP_T for transpose)