ó \ K]c@sLddlmZddlmZddlmZddlZdd„ZdS(iÿÿÿÿ(tcuda(tdriver(t numpy_supportNc sWt|ddƒ}|ss|j\}}|jj||jjf}tjjj||f|d|jd|ƒ}ntj |jƒ‰t j ƒj }t tjdtj|dƒdƒƒ}t ||ƒ}||df‰tj‡‡fd†ƒ} t |jd|dƒt |jd|dƒf} ||f} | | | |f||ƒ|S(sáCompute the transpose of 'a' and store it into 'b', if given, and return it. If 'b' is not given, allocate a new array and return that. This implements the algorithm documented in http://devblogs.nvidia.com/parallelforall/efficient-matrix-transpose-cuda-cc/ :param a: an `np.ndarray` or a `DeviceNDArrayBase` subclass. If already on the device its stream will be used to perform the transpose (and to copy `b` to the device if necessary). tstreamitdtypeiic stjjdˆdˆƒ}tjj}tjj}tjjtjj}tjjtjj}||}||}|||jdkrÆ|||jdkrÆ|||||f|||fs