f :ddlmZddlmZddlZddlmZddZdS))cuda)driverN) numpy_supportc  t|dd}|sT|j\}}|jj|z|jjf}tjj||f||j|}tj |j tj j }ttjdtj|ddz }t||z }||dzf tj fd} t|jd|z dzt|jd|z dzf} ||f} | | | |f|||S)aCompute 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). streamr)dtyperc8tj }tjj}tjj}tjjtjjz}tjjtjjz}||z}||z}||z|jdkr)||z|jdkr|||z||zf|||f<tj ||jdkr"||jdkr|||f|||f<dSdSdS)N)shaperrr ) rsharedarray threadIdxxyblockIdxblockDimr syncthreads) inputoutputtiletxtybxbyrrdt tile_shapes .kernel)s{  z << ^  ^  ]_t} . ]_t} . G G 7U[^ # #R%+a.(@(@ b"r'!12DRL  v|A  1v|A#6#6Br>sb,,,,,, )))))): : : : : : r