o hO@sFddlZddlZddlZddlmZmZddlmZddlm Z m Z ddl m Z ddl Z ddlmZgdZGdd d eZd d Zd d ZddZddZGdddZGdddZedgdZGdddeZGdddeZGdddeZGdddZd d!Zd"Zd#Z d$d%Z!d/d'd(Z"   ) * + , & & &d0d-d.Z#dS)1N) defaultdict namedtuple) attrgetter)AnyOptional) deprecated) DeviceType) EventListFormattedTimesMixinIntervalKernel FunctionEventFunctionEventAvg StringTable MemRecordsAcccseZdZdZfddZddZddZdd Zd d Zd d Z e ddZ       d%ddZ ddZ ddZdedefddZ  d&d!d"Zd#d$ZZS)'r z'A list of Events (for pretty printing).csR|dd}|dd}|dd}tj|i|||_||_d|_||_dS)N use_deviceprofile_memoryF with_flops)popsuper__init__ _use_device_profile_memory _tree_built _with_flops)selfargskwargsrrr __class__p/var/www/html/construction_image-detection-poc/venv/lib/python3.10/site-packages/torch/autograd/profiler_util.pyrs    zEventList.__init__cCs"|||d|_dS)NT)_populate_cpu_children_remove_dup_nodes_set_backward_stacktracesrrr r r! _build_tree's zEventList._build_treecCs|SN)tabler%r r r!__str__-zEventList.__str__cs ttt|D]C}||jdurM||jj||jkrMt||jjdkrM||j||j_||j||j_||jD]}||j|_q?|q tdkrVdSfddt|D}| | |q)NTrcsg|] \}}|vr|qSr r ).0indev to_deleter r! @z/EventList._remove_dup_nodes..) setrangelen cpu_parentname cpu_childrenkernelsadd enumerateclearextend)ridxchnew_evtsr r/r!r#0s$   zEventList._remove_dup_nodesc Csdd|D}t|tdd}tj|ddd}|D]U\}}t|ddd}g}|D]D}t|dkri|d } |jj| jjksF|jj| jjkrK|n| ||j d us]Jd |j | | nt|dks2| |q*qd S) a4Populate child events into each underlying FunctionEvent object. One event is a child of another if [s1, e1) is inside [s2, e2). Where s1 and e1 would be start and end of the child event's interval. And s2 and e2 start and end of the parent event's interval Example: In event list [[0, 10], [1, 3], [3, 4]] would have make [0, 10] be a parent of two other intervals. If for any reason two intervals intersect only partially, this function will not record a parent child relationship between then. cSs"g|] }|js|jtjkr|qSr )is_async device_typerCPUr,evtr r r!r1Ts z4EventList._populate_cpu_children..thread)keycSs |j|jfSr')rFnode_ideventr r r!as z2EventList._populate_cpu_children..cSs|jj|jj gSr') time_rangestartendrIr r r!rKsrNz(There is already a CPU parent event for )sortedr itertoolsgroupbyr5rLrMrNrappend_cpu_childr6rGset_cpu_parentappend) r sync_eventseventsthreads _thread_id thread_eventsthread_events_current_eventsrJparentr r r!r"Ds@        z EventList._populate_cpu_childrencsfddi}|D]}|dur&|jdur&|j|jf}||vr&|j||<q |D]$}|}|durM|jdus:J|j|jf}||vrJ|||_q)g|_q)dS)Ncs$|durdS|jdkr |S|jSNr+)scoper6rE bw_parentr r!rcs   z6EventList._set_backward_stacktraces..bw_parent)stack sequence_nrrF fwd_thread)r fwd_stacksrEtpr rbr!r$s$     z#EventList._set_backward_stacktracescCstdd|DS)Ncs|]}|jVqdSr')self_cpu_time_totalr,rJr r r! z0EventList.self_cpu_time_total..)sumr%r r r!rkszEventList.self_cpu_time_totalNdK7PFc Cs t||||||||j|j|d S)a(Print an EventList as a nicely formatted table. Args: sort_by (str, optional): Attribute used to sort entries. By default they are printed in the same order as they were registered. Valid keys include: ``cpu_time``, ``cuda_time``, ``xpu_time``, ``cpu_time_total``, ``cuda_time_total``, ``xpu_time_total``, ``cpu_memory_usage``, ``cuda_memory_usage``, ``xpu_memory_usage``, ``self_cpu_memory_usage``, ``self_cuda_memory_usage``, ``self_xpu_memory_usage``, ``count``. top_level_events_only(bool, optional): Boolean flag to determine the selection of events to display. If true, the profiler will only display events at top level like top-level invocation of python `lstm`, python `add` or other functions, nested events like low-level cpu/cuda/xpu ops events are omitted for profiler result readability. Returns: A string containing the table. ) sort_by row_limitmax_src_column_widthmax_name_column_widthmax_shapes_column_widthheaderrrtop_level_events_only) _build_tablerr)rrtrurvrwrxryrzr r r!r(szEventList.tablecCs&ddl}|js dn|j}t|dx}d}|d|D]K}|jdur#q|d|j|jj|j|j s6|j n d|j d|j d |j D]}|d |jd |jjd |j d |d|d |d7}qFqt |dkr|||d|j||dWddS1swYdS)zExport an EventList as a Chrome tracing tools file. The checkpoint can be later loaded and inspected under ``chrome://tracing`` URL. Args: path (str): Path where the trace will be written. rNcudaw[zc{{"name": "{}", "ph": "X", "ts": {}, "dur": {}, "tid": {}, "pid": "CPU functions", "args": {{}}}}, z " node_id:z , thread_id:z "z {"name": "z", "ph": "s", "ts": z , "tid": z , "pid": "CPU functions", "id": z, "cat": "cpu_to_z", "args": {}}, r+])osropenwrite trace_nameformatrLrM elapsed_us is_remoterFrHr9r5seektellSEEK_SETtruncate)rpathr device_namefnext_idrE_r r r!export_chrome_tracesL        "zEventList.export_chrome_tracecCsgdS)N)rkself_cuda_time_totalself_xpu_time_totalself_privateuse1_time_totalr r%r r r!supported_export_stacks_metrics r*z)EventList.supported_export_stacks_metricsrmetricc Cs||vrtdt|tdd}t|dY}|D]M}|jrkt|jdkrkt||ddddd d}t |dkrkd }t |jD] }|| |7}|d 7}qH|dd d tt |}| |dqWddS1swwYdS)Nzmetric should be one of: z ; ____r}rr|devicexpu privateuse1;rP  ) r ValueErrorstr maketransrrdr5getattrreplaceintreversed translater) rrrtranslate_tablerrE metric_value stack_strentryr r r! export_stackss4       "zEventList.export_stacksrcs|jsJtt}dttdfffdd }|D]|||||qt||j|j |j d}|D]j d|_ |sEd_ |sJd_ q6|S)aAverages all function events over their keys. Args: group_by_input_shapes: group entries by (event name, input shapes) rather than just event name. This is useful to see which input shapes contribute to the runtime the most and may help with size-specific optimizations or choosing the best candidates for quantization (aka fitting a roof line) group_by_stack_n: group by top n stack trace entries group_by_overload_name: Differentiate operators by their overload name e.g. aten::add.Tensor and aten::add.out will be aggregated separately Returns: An EventList containing FunctionEventAvg objects. return.csrt|jt|jt|jt|jt|jg}|r|j|r(|t|j|dkr5||j d|7}t |SNr) rrGrHrB is_legacyis_user_annotationrV overload_name input_shapesrdtuple)rJgroup_by_input_shapesgroup_by_stack_ngroup_by_overload_namerGrar r!get_keyCs z'EventList.key_averages..get_keyrrrNr)rrrrrr:r valuesrrrrdrr)rrrrstatsravg_listr rar! key_averages)s4   zEventList.key_averagescCs(t}|D] }||7}d|_qd|_|S)zVAverages all events. Returns: A FunctionEventAvg object. NTotal)rrG)r total_statrEr r r! total_averagejs zEventList.total_average)NrprqrrrsNF)FrF)__name__ __module__ __qualname____doc__rr&r)r#r"r$propertyrkr(rrrrrr __classcell__r r rr!r s4  E  *8 Ar cCsDd}d}||kr||ddS||kr||ddS|ddS)+Define how to format time in FunctionEvent.g.Ag@@z.3fsmsusr )time_us US_IN_SECONDUS_IN_MSr r r! _format_timexs rcCs6|dkr|dksJd|dS|d|ddS)rrzExpected time_us == 0 but got NaNgY@.2f%r )r total_time_usr r r!_format_time_sharesrcCsd}d|}d|}t||kr|d|ddSt||kr*|d|ddSt||kr:|d|ddSt|dS)z&Return a formatted memory size string.i?rz Gbz Mbz Kbz b)absr)nbytesKBMBGBr r r!_format_memorys    rcstfddS)Ncstt|Sr')rrr%r7r r!rKsz!_attr_formatter..)rrr rr!_attr_formattersrc@speZdZdZedZedZedZedZedZ edZ e dd Z e d d Z e ed ed ddZdS)r z{Helpers for FunctionEvent and FunctionEventAvg. The subclass should define `*_time_total` and `count` attributes. cpu_time device_timecpu_time_totaldevice_time_totalrkself_device_time_totalcC|jdkrdSd|j|jSNrgr)countrr%r r r!rzFormattedTimesMixin.cpu_timecCrr)rrr%r r r!rrzFormattedTimesMixin.device_timez<`cuda_time` is deprecated, please use `device_time` instead.categorycC|jSr')rr%r r r! cuda_timezFormattedTimesMixin.cuda_timeN)rrrrr cpu_time_strdevice_time_strcpu_time_total_strdevice_time_total_strself_cpu_time_total_strself_device_time_total_strrrrr FutureWarningrr r r r!r s$  r c@seZdZddZddZdS)r cCs||_||_dSr')rMrN)rrMrNr r r!rs zInterval.__init__cCs |j|jS)z4 Returns the length of the interval )rNrMr%r r r!rs zInterval.elapsed_usN)rrrrrr r r r!r s r r )r7rdurationc@seZdZdZddddddddddddejddddddddfddZdd Zd d Zd d Z e ddZ e ddZ e e dedddZe ddZe ddZe ddZe e dedddZe dd Ze e d!edd"d#Ze d$d%Zd&d'ZdS)(r z.Profiling information about a single function.NrFrPcCs||_||_||_||_||_t|||_||_||_g|_ d|_ g|_ d|_ ||_ ||_||_| |_| |_| |_| |_| |_||_||_||_||_||_|durT|n||_||_||_||_d|_d|_d|_ dS)Nr+rP)!idrHr7rrr rLrFrfr9rr8r6rconcrete_inputskwinputsrdr`rcpu_memory_usagedevice_memory_usagerArrerB device_indexdevice_resource_idrflopsrself_cpu_percenttotal_cpu_percenttotal_device_percent)rrr7rFstart_usend_usrrfrrdr`rrrrArrerHrBrrrrrrrrr r r!rsB  zFunctionEvent.__init__cCs(|jtjksJ|jt|||dSr')rBrrCr9rVr )rr7rrr r r! append_kernelszFunctionEvent.append_kernelcCs>|jtjksJt|tsJ|jtjksJ|j|dS)zAppend a CPU child of type FunctionEvent. One is supposed to append only direct children to the event to have correct self cpu time being reported. N)rBrrC isinstancer r8rV)rchildr r r!rTszFunctionEvent.append_cpu_childcCs8|jtjksJt|tsJ|jtjksJ||_dS)a$Set the immediate CPU parent of type FunctionEvent. One profiling FunctionEvent should have only one CPU parent such that the child's range interval is completely inside the parent's. We use this connection to determine the event is from top-level op or not. N)rBrrCrr r6)rr^r r r!rUs zFunctionEvent.set_cpu_parentcC0|js |jtjkr dS|jtdd|jDS)Nrcsrjr')rr,rr r r!rm/ z6FunctionEvent.self_cpu_memory_usage..)rArBrrCrror8r%r r r!self_cpu_memory_usage+   z#FunctionEvent.self_cpu_memory_usagecCr)Nrcsrjr')rrr r r!rm7rz9FunctionEvent.self_device_memory_usage..)rArBrrCrror8r%r r r!self_device_memory_usage3rz&FunctionEvent.self_device_memory_usagezO`self_cuda_memory_usage` is deprecated. Use `self_device_memory_usage` instead.rcCrr'rr%r r r!self_cuda_memory_usage;rz$FunctionEvent.self_cuda_memory_usagecCs|jtjkr |jSdSr)rBrrCrLrr%r r r!rCs  zFunctionEvent.cpu_time_totalcCr)Nrcsrjr')rrr r r!rmNrz4FunctionEvent.self_cpu_time_total..)rArBrrCrror8r%r r r!rkJrz!FunctionEvent.self_cpu_time_totalcCs|js|jsdS|jtjkr/|js%tdd|jDtdd|jDStdd|jDS|jtj tj tj fvs.csrjr'r)r,r?r r r!rmYrcsrjr'r r r r r!rm^rn)rArrBrrCrror9r8CUDA PrivateUse1MTIArLrr%r r r!rRs    zFunctionEvent.device_time_totalzA`cuda_time_total` is deprecated. Use `device_time_total` instead.cCrr'r r%r r r!cuda_time_totalgrzFunctionEvent.cuda_time_totalcCsV|js|jsdS|jtjkr|jtdd|jDS|jtjtj tj fvs(J|jS)Nrcsrjr'r rr r r!rmtrz7FunctionEvent.self_device_time_total..) rArrBrrCrror8r rrr%r r r!ros     z$FunctionEvent.self_device_time_totalzK`self_cuda_time_total` is deprecated. Use `self_device_time_total` instead.cCrr'rr%r r r!rrz"FunctionEvent.self_cuda_time_totalcCrr'rr%r r r!rGszFunctionEvent.keycCs*|j}|j}|j}dgd|jd|jd|jd|jd|jd|j d|j j d |j j d t d d |jDd |d|d|jd|jdt |jd|jd |d|d|jd|jd|jd|jdS)Nrz.r_time=z thread= input_shapes= cpu_memory_usage=_memory_usage=z is_async=z is_remote=z seq_nr=z is_legacy=>)rrrjoinrr7rrBrHrrLrMrNrr8rFrrrArrer)rrrrr r r!__repr__sR. zFunctionEvent.__repr__)rrrrrrCrrrTrUrrrrrrrrkrrrrrGrr r r r!r sx  @            r c@s2eZdZdZd ddZddZdd Zd d ZdS) rz:Used to average stats over multiple FunctionEvent objects.rNcCsd|_d|_d|_d|_d|_d|_d|_d|_d|_d|_ d|_ d|_ d|_ d|_ d|_d|_d|_d|_d|_d|_tj|_d|_d|_dS)NrF)rGrrHrArrrrrkrrrrdr`rrrrr8r6rrCrBrrr%r r r!rs. zFunctionEventAvg.__init__cCs`|jdur=|j|_|j|_|j|_|j|_|j|_|j|_|j|_|j|_|j|_|j |_ |j |_ |j |_ |j |_ |j |_ t|ttfsFJ|j|jksNJ|j|j7_|j|j7_|j|j7_|j|j7_|j|j7_|j|j7_|j|j7_|j|j7_|j|j7_|jdur|j|_|S|jdur|j|j7_|Sr')rGrHrArr6r8rrrdr`rBrrrrr rrrrkrrrrrrrrotherr r r!r:s@   zFunctionEventAvg.addcCs ||Sr')r:rr r r!__iadd__s zFunctionEventAvg.__iadd__cCsx|jsdn|j}|j}|j}|j}d|jd|jd|jd|d|d|d|dt|jd |j d|d |d S) Nr|z.) _mem_records _start_nses_indicesr5rQr;zip)r mem_recordstmpr r r!rs zMemRecordsAcc.__init__ccsLt|j|d}t|j|d}t||D] }|j|j|VqdS)z Return all records in the given interval To maintain backward compatibility, convert us to ns in function iN)bisect bisect_leftr( bisect_rightr4r'r))rrr start_idxend_idxr%r r r! in_intervals zMemRecordsAcc.in_intervalN)rrrrrr2r r r r!rs rcsgd}tfdd|DS)N))autograd/__init__ _make_grads)r3backward)z torch/tensorr5)_internal/common_utils prof_callable)r6prof_func_call)r6prof_meth_callc3s*|]}|dvo|dv VqdS)rr+Nr )r,rrr r!rms(z&_filter_stack_entry..)all)rfiltered_entriesr r:r!_filter_stack_entry sr=z[memory]z [OutOfMemory]cCsttddddddg}||vS)Nz profiler::_record_function_enterz$profiler::_record_function_enter_newzprofiler::_record_function_exitz aten::is_leafzaten::output_nrzaten::_version)MEMORY_EVENT_NAMEOUT_OF_MEMORY_EVENT_NAME)r7filtered_out_namesr r r! _filter_names rAFcCs$t}||}|r|drd}|S)Nz ProfilerStep#z ProfilerStep*)r startswith)r7 with_wildcard string_tabler r r! _rewrite_name-s  rErprqrrrsc .st|dkrdStdd|D} tdd|D} |dj} | s'| r'tdtdd|D} tdd|D}d urNtt|fd d d d | ||d}tdd|Dd}|d urbt||}tdd|Dd}|d urvt||}d}|}d }dd|D}t|dk}|rtdd|Dd}|d urt||}dg}|r|d|gd7}| d ur| nd}| r| d|d|d|d|dg|r| ddg| r| r| |d d|d g|d!td"d|D}|r|d#d$dgdg gd>fd&d' }d(d)}|||r'|||d*|d D]}||q/| rC|d+|||rQ|d,||d-d.|ryd/d|D}t|dkrw|t|\}}|d0|||nd1}d}d} d}!d }gfd2d3}"d}#d}$|D]/}%|#|%j 7}#|%j t jkr|%jr|$|%j7}$q|%j t jt jt jfvr|%js|$|%j7}$q|d ur|"d4|!|"|| r|"d4|!|"d5|"| |"|j||"| d6d7}&d}'|D]?}%|'|krn6| r|%jd urq|'d*7}'|%j}(|d ur-t|(|d8kr-|(d |d8d9}(t|%j |#|%_|%js>t|%j|#nd|%_|(g})|rg|%j}*|d urbt|*|d8krb|*d |d8d9}*|)|*g7})|)|%j|%j|%j|%j|%j g7})| rt|%j|$|%_!|) |%j"|%j!|%j#|%j$g|r|) t%|%j&t%|%j'g| r| r|) t%|%j(t%|%j)g|)|%j*|r|)|%j+| r|)t,|%j-d ||r|%j.dkr|)d:n |)|%j.|d;|rd}+t|%j/dkr|&|%j/d|}+|)|+|"|j|)|r7dgt|d*},|%j/d*d D]}-|"|j|,|&|-|gq|,d|"|j|,q|"| |"d.csrFrGrrlr r r!rmGrHz9use_device is None, but there is device performance data.cs(|]}|jduot|jdkVqdSr)rr5rlr r r!rmO  csrIr)rr5rlr r r!rmTrJNcs"t|ddddddS)Nr|rrr)rrra)rtr r!rK]s  z_build_table..T)rGreversercss|]}t|jVqdSr')r5rGrDr r r!rmjrHcss|] }tt|jVqdSr')r5rrrDr r r!rmns cSs*g|]}|jdurt|jdkr|jqSr)rdr5rDr r r!r1vs$z _build_table..css"|] }tdd|DVqdS)css|]}t|VqdSr'r5)r,rr r r!rm|sz)_build_table...N)max)r,rdr r r!rm|s Namez Overload Name)z Self CPU %zSelf CPUz CPU total %z CPU totalz CPU time avgNonezSelf z %z totalz time avgzCPU Memz Self CPU Memz Memz # of Callscss|]}|jdkVqdS)rPN)rHrDr r r!rmrHzNode IDrrcs\dd|t|dd7<dd|d7<d|7<dS)Nrz{: }r-)r)paddingtext_dir) SPACING_SIZEheader_sep_lstline_length_lstrow_format_lstr r! add_columns z _build_table..add_columncSsrgd}|dks Jtdtt|dtt|d}|dkr'|t|ks)Jtdt|d|t|fS)N)FLOPsKFLOPsMFLOPsGFLOPsTFLOPsPFLOPsrr+ g) rOminmathlog10floatr5powfloorr)r flop_headers log_flopsr r r!auto_scale_flopss  & z&_build_table..auto_scale_flopsr+z Input ShapeszSource Location<)rUcSsg|] }|jdkr|jqSr#)rrDr r r!r1r2zTotal Fcs|ddS)Nr)rV)r)resultr r!rVs z_build_table..append=z1This report only display top-level ops statisticscSsDt||kr t||}||d}t|dkr d|dd}|S)Nra...rN)rsrc_column_widthoffsetr r r! trim_paths    z_build_table..trim_pathraroz--z8.3fzSelf CPU time total: z time total: )r)2r5anyr RuntimeErrorr rQrOrcrVupperr=rkrBrrCrrr rrrrr6rGrrrArrrrrrrrrrrrrrrrrHrrrrdrr).rXrtryrurvrwrxrrrzhas_device_timehas_device_memrhas_input_shapeshas_overload_namesname_column_widthshapes_column_widthDEFAULT_COLUMN_WIDTHflops_column_widthrpstacks has_stackheadersrappend_node_idrZrkr raw_flops flops_scale flops_header row_format header_sep line_lengthrVsum_self_cpu_time_totalsum_self_device_time_totalrErr event_limitr7 row_valuesr src_field empty_headersrr )rVrWrXrmrYrtr!r{6s                                      r{)F) NNrprqrrrsFFF)$r-rRrd collectionsrroperatorrtypingrrtyping_extensionsrrtorch.autogradr__all__listr rrrrr r r r rrrr=r>r?rArEr{r r r r!sP    `  PQ