Ph UdZddlmZddlmZmZmZmZmZddl Z ddl m Z ddl m Z m Z mZmZmZmZgdZeeed <ee ZGd d ZGd d e j.j0ZGdde j.j0Zy)zlAutograd functions for stream-aware CUDA copy. It is used to overlap copy and computation on the same GPU. )deque)DequeListOptionalTupleSequenceN)Tensor)AbstractStreamcurrent_stream get_device record_stream use_stream wait_stream)ContextCopyWait__all__c"eZdZUeed<eed<y)r prev_stream next_streamN)__name__ __module__ __qualname__r __annotations__oC:\Users\daisl\Desktop\realtime-object-detection\venv\Lib\site-packages\torch/distributed/pipeline/sync/copy.pyrrsrrc `eZdZdZededededefdZedede de e e dffd Z y ) rz#Copies tensors on specific streams.ctxrrreturnc||_||_g}tt|}t |5t |5|D]n}t j |rF|jt|d}|j|t||t||^|j|p ddddddt|S#1swYxYw#1swYt|SxYw)NT non_blocking) rrr r rtorch is_tensortoappendrtuple)r rrinputoutput output_streamxys rforwardz Copy.forward"s&%&z+'>?  $j&=??1%Z 44HAMM!$"![1"!]3MM!$'> $V}'>&= $V}s# CA4C 1C C CC+ grad_output.c|j}|j}tt|}t t |}t |5t |5t|D]G}|jt |d}|j|t||t||I ddddddd}|t|zS#1swY!xYw#1swY%xYw)N)maxlenTr#NN) rrrlenr r rreversedr' appendleftrr)) r r0rr grad_input input_streamr-r. grad_streamss rbackwardz Copy.backward;soo oo $)[1A$B %j&=>  $j&=k*DDK0tDD%%a(a-a.+'> $6B eJ///'>&= $ $s% CAC0CC CC%N rrr__doc__ staticmethodrr Tensorsr/r rrr:rrrrrsm-W>dk.0g0V0x?OQT?T9U00rrc `eZdZdZededededefdZedede de e e dffd Z y ) rzSynchronizes a stream to another stream. Place it just before you want to start an operation on the next stream, provided that all operations on the previous stream are done. r rrr!cZ||_||_t||td|DS)Nc3jK|]+}tj|r|jn|-yw)N)r%r&detach).0r-s r zWait.forward..bs'Le5??1#5QXXZ1MMcjMM)g)F)hv>NPS>S8T))rr)r< collectionsrtypingrrrrrr%r streamr r r rrrrstrrr>rautogradFunctionrrrrrrLsr99 ff0c0 6   005>> " "00f)5>> " ")r