o 0 ir2 @sDddlZddlZddlmZmZddlZddlmZddlmZddlmZe e e e fZ e e e fZ e e e e e ffZee e e e fZe e dfZejGdddZd ed ed efd d Zded ed ed efddZde edfde e dfd e e dffddZde e dfd e edffddZde e dfde e ee edffd e eeffddZdddee e dfge e dffdee edfge edffd dfdd Zd3d!d"Zd3d#d$Zd3d%d&Zd3d'd(Z d4d)dd*dd+edd dfd,d-Zd.ee d/ee d0eee e d e e ee edfffd1d2Z!dS)5N)CallableOptional)_array)_chunk)_modes.c@s2eZdZUeeed<eeed<eeed<dS) _Blocking i_partitions j_partitions k_partitionsN)__name__ __module__ __qualname__listint__annotations__rrk/home/app/PaddleOCR-VL-test/.venv_paddleocr/lib/python3.10/site-packages/cupyx/distributed/array/_linalg.pyrs rlocation_map_alocation_map_breturnc Csg}g}g}dd}|D]\}}||||||q|D]\}}||||||q!dd} | |}| |}| |}dd} |D]\}}| ||| ||qH|D]\}}| ||| ||q[t|||S)NcSs2|\}}}|dkr td||||dS)Nz"Step other than 1 is not supported) RuntimeErrorappend)indices partitionsstartstopsteprrradd_to_partitions6s  z)_find_blocking..add_to_partitionscSsXt|dkr td||dg}t||ddD] \}}||kr)||q|S)NrzArray has no chunkr)lenrsortzipr)rresxyrrrto_unique_sortedGs   z(_find_blocking..to_unique_sortedcSs.|\}}}||d||krtddS)NrzInconsistent index mapping)indexr)rrrr_rrr check_indicesXs z%_find_blocking..check_indices)keysr) rrrr r rZ i_indicesZ k_indicesZ j_indicesr%r(rrr_find_blocking-s,          r*blockingcCs|j}|j}|j}g}t||ddD]V}t||ddD]J}t||ddD]>} |d| df} | d|df} t|| } t|| } | | @}|r_|}|| | |fq*td| d| qq|S)NrrzAThere is no device that can perform multiplication between block z and ) rr r r!setr)poprr)r+rrrr r planZi_rangeZj_rangeZk_rangeblock_ablock_bZ devices_aZ devices_b intersectiondevrrr_make_execution_planhs2r4slicesshapecCs,t|t|ks Jtddt||DS)Ncss|] \}}||VqdSN)r).0slengthrrr sz%_convert_to_tuples..)rtupler!)r5r6rrr_convert_to_tuplessr=tuplescCstdd|DS)Ncss|]}t|VqdSr7slice)r8trrrr;sz%_convert_to_slices..)r<)r>rrr_convert_to_slicessrB index_mapc Cs|i}|D]5\}}t|D],\}}t||}|dd|dd}} tt| } ||i} | | i} || |<qq|S)N)items enumerater=typingcast _BlockIdx setdefault) r6rCZ location_mapsr3ZidxsZchunk_iidxZ idx_tuples batch_idxZ block_idxZ location_maplocationrrr_group_by_batchs      rNarr_array.DistributedArrayf_shapef_idxcsldtjdtjffdd i}|jD]\}}fdd|D||<q|j}t||j||j|j S)NchunkrcsH|j|jj}|j}fdd|jD}t||j|||jS)Ncsg|] \}}||fqSrr)r8datarK)rRrr sz>_reshape_array_with..reshape_chunk..) arrayZreshaper6r&updatesr_Chunkready prevent_gc)rSrTr&rW)rRrQrr reshape_chunks  z*_reshape_array_with..reshape_chunkcsg|]}|qSrr)r8rS)r[rrrUsz'_reshape_array_with..) rrX _chunks_maprEr6rDistributedArraydtype_mode_comms)rOrQrR chunks_mapr3chunksr6r)rRrQr[r_reshape_array_withs rccCt|ddddS)NcSsd|SNr,rr6rrrz'_prepend_one_to_shape..cSstdf|Sr7r?rKrrrrgrcrOrrr_prepend_one_to_shape rmcCrd)NcSs|dSrerrfrrrrgrhz&_append_one_to_shape..cSs|tdfSr7r?rirrrrgrjrkrlrrr_append_one_to_shapernrocC&|jddks Jt|ddddS)NrcS |ddSNrqrrfrrrrg z!_pop_from_shape..cSrrrsrrirrrrgrtr6rcrlrrr_pop_from_shape rvcCrp)NrrcS |ddSNrrrfrrrrgrtz'_pop_front_from_shape..cSrxryrrirrrrgrtrurlrrr_pop_front_from_shaperwrzaboutc" Ks|durtddD]}||vrtd|dq t|tjr%t|tjs)td|tj}|tj}d}}|jdkrDd }t|}|jdkrOd }t |}|j d d\}}|j d d\} } || kss|j dd |j dd krwt d t |j |j } t |j |j } | | krtd d d|jD} d}| D]}| |}| |}t||}t|||}t|}|D]y\}}}||}||}|j|||}|j|||}|tj|tj|t|dt|df}|3}||jtjjj|j|jfi|}tj|| |||fd}| |!||j"}Wdn 1s.wYqq|j dd || f} t| || tj#|j$}!|rRt%|!}!|rYt&|!}!|!S)akMatrix multiplication between distributed arrays. The arguments must have compatible :attr:`~DistributedArray.shape` and :attr:`~DistributedArray.index_map`. This operation converts its operands into the replica mode, and compute their product in the sum mode. Args: a, b: Input distributed arrays. out (optional): A location into which the result is stored. This option is currently not supported. Returns: The matrix product of the inputs. Example: >>> A = distributed_array( ... cupy.arange(6).reshape(2, 3), ... make_2d_index_map([0, 2], [0, 1, 3], ... [[{0}, {1, 2}]])) >>> B = distributed_array( ... cupy.arange(12).reshape(3, 4), ... make_2d_index_map([0, 1, 3], [0, 2, 4], ... [[{0}, {0}], ... [{1}, {2}]])) >>> C = A @ B >>> C.mode 'sum' >>> C.all_chunks() {0: [array([[0, 0], [0, 3]]), array([[0, 0], [6, 9]])], 1: [array([[20, 23], [56, 65]])], 2: [array([[26, 29], [74, 83]])]} >>> C array([[20, 23, 26, 29], [56, 68, 80, 92]]) .. seealso:: :obj:`numpy.matmul` NzArgument `out` is not supported)ZsubokZaxesZaxisz Argument `z` is not supportedzHMixing a distributed array with a non-distributed array is not supportedFrTrDzShapes are incompatiblezMismatched batch shapescSsi|]}|gqSrr)r8r3rrr ,szmatmul..r)rZ)'r isinstancerr]Z _to_op_moderZREPLICAndimrmror6 ValueErrorrNrCr)devicesr*r4rBr\flushr@Zon_readyZ wait_eventrYcupyZlinalgZ_productmatmulrVrrXrecordrr^ZSUMr`rzrv)"r{r|r}kwargsparamZ one_prependedZ one_appendednmm2pZlocation_maps_aZlocation_maps_brar^rLrrr+r/Z index_prefixr0r1r3Zloc_aZloc_bZchunk_aZchunk_br&streamZchunk_ab_arrayZchunk_abr6r"rrrrs/      $        rrr rc Cs |ddksJtt||ksJ|ddksJtt||ks$Ji}t|t|dks2Jtt|D]K}t||t|dksHJtt||D]2}||}||d}||}||d} t||t|| f} |||D] } || g| qvqPq8|S)aCreate an ``index_map`` for a 2D matrix with a specified blocking. Args: i_partitions (list of ints): boundaries of blocks on the `i` axis j_partitions (list of ints): boundaries of blocks on the `j` axis devices (2D list of sets of ints): devices owning each block Returns: dict from int to array indices: index_map Indices for the chunks that devices with designated IDs are going to own. Example: >>> index_map = make_2d_index_map( ... [0, 2, 4], [0, 3, 5], ... [[{0}, {1}], ... [{2}, {0, 1}]]) >>> pprint(index_map) {0: [(slice(0, 2, None), slice(0, 3, None)), (slice(2, 4, None), slice(3, 5, None))], 1: [(slice(0, 2, None), slice(3, 5, None)), (slice(2, 4, None), slice(3, 5, None))], 2: [(slice(2, 4, None), slice(0, 3, None))]} rr)sortedr-rranger@rJr) rr rrCijZi_startZi_stopZj_startZj_stoprKr3rrrmake_2d_index_mapXs&   r)rrPr7)" dataclassesrGrrrZcupyx.distributed.arrayrrrr<rZ _SliceIndicesrIdictZ_BlockLocationMaprZ_ExecutionPlanZ _BatchIdx dataclassrr*r4r@r=rBrNrcrmrorvrzrr-rrrrrs       ; #                {