U hd @s:ddlmZddlmZddlZddlmZdddZdS))cuda)driverN) numpy_supportc st|dd}|sJ|j\}}|jj||jjf}tjjj||f||j|d}t |jt j }t tdt|dd}t ||}||dftjfdd} t |jd|dt |jd|df} ||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)dtyperc stjjd}tjj}tjj}tjjtjj}tjjtjj}||}||}|||jdkr|||jdkr|||||f|||f<t ||jdkr||jdkr|||f|||f<dS)N)shaperrr) rZsharedarrayZ threadIdxxyZblockIdxZblockDimr Z syncthreads) inputoutputZtileZtxtyZbxZbyr r dtZ tile_shapeU/home/sam/Atlas/atlas_env/lib/python3.8/site-packages/numba/cuda/kernels/transpose.pykernel)s$ztranspose..kernel)getattrr ritemsizerZcudadrvZ devicearrayZ DeviceNDArraynpsZ from_dtyperZ get_deviceZMAX_THREADS_PER_BLOCKintmathpowlogZjit) abrcolsrowsstridesZtpbZ tile_widthZ tile_heightrblocksthreadsrrr transposes*      ,r#)N) ZnumbarZnumba.cuda.cudadrv.driverrrZnumba.nprrr#rrrrs