:i :SSKJr SSKJr SSKrSSKJr SSjrg))cuda)driverN) numpy_supportc ^ ^ [USS5nU(drURup4URRU-URR4n[R R RXC4UURUS9n[R"UR5m [R"5Rn[[R"S[R "US5S- 55n[Xg- 5nXS-4m [R"U U 4Sj5n [URSU- S-5[URSU- S-54n X4n XX4"X5 U$)aCompute the transpose of 'a' and store it into 'b', if given, and return it. If 'b' is not given, allocate a new array and return that. This implements the algorithm documented in http://devblogs.nvidia.com/parallelforall/efficient-matrix-transpose-cuda-cc/ :param a: an `np.ndarray` or a `DeviceNDArrayBase` subclass. If already on the device its stream will be used to perform the transpose (and to copy `b` to the device if necessary). streamr)dtyperc>[RRT T S9n[RRn[RR n[R R[RR-n[R R [RR -nXc-nXT-nXd-URS:a$XS-URS:aXU-XS-4X$U4'[R"5 XRS:aXqRS:a X#U4XU4'ggg)N)shaperrr ) rsharedarray threadIdxxyblockIdxblockDimr syncthreads) inputoutputtiletxtybxbyrrdt tile_shapes _/srv/projetos/modelo_ml_acdoc/venv/lib/python3.13/site-packages/numba/cuda/kernels/transpose.pykerneltranspose..kernel)s{{  z < ^^   ^^   ]]__t}} . ]]__t}} . G G 7U[[^ #%++a.(@ b"'!12DRL  ||A 1||A#6Br?s, ): r!