i 8ddlmZddlmZddlZddlmZddZy))cuda)driverN) numpy_supportc  t|dd}|sw|j\}}|jj|z|jjf}tj j j||f||j|}tj|j tjj}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)dtyperctjj }tjj}tjj }tj jtjjz}tj j tjj z}||z}||z}||z|jdkr'||z|jdkr|||z||zf|||f<tj||jdkr ||jdkr |||f|||f<yyy)N)shaperrr ) rsharedarray threadIdxxyblockIdxblockDimr syncthreads) inputoutputtiletxtybxbyrrdt tile_shapes q/Volumes/fast/ai/experiments/voice-extract-mac/.venv/lib/python3.12/site-packages/numba/cuda/kernels/transpose.pykernelztranspose..kernel)s{{  z < ^^   ^^   ]]__t}} . ]]__t}} . G G 7U[[^ #R%++a.(@ b"r'!12DRL  v||A 1v||A#6Br>s, ): r