a OSicw@s<ddlmZmZmZmZmZmZmZddlm Z m Z m Z m Z m Z mZmZmZmZmZmZddlmZddlZddlZddlmZddlmZmZmZmZddlm Z zddl!m"Z"Wn*e#yddl$Z$Gd d d e%Z"Yn0Gd d d e%Z&Gd dde"Z'Gddde%Z(ddZ)Gddde%Z*ddZ+ddZ,dS)) EventList FunctionEvent MemRecordsAccMEMORY_EVENT_NAME _filter_name_filter_stack_entry _rewrite_name) DeviceTypeProfilerActivityProfilerConfig ProfilerStatekineto_available_ProfilerResult_disable_profiler_enable_profiler_prepare_profiler_supported_activities _kineto_step)_ExperimentalConfigN)Future)AnyDictListOptional)warn)ContextDecoratorc@s$eZdZddZddZddZdS)rcCstdSNNotImplementedErrorselfr!S/var/www/html/django/DPS/env/lib/python3.9/site-packages/torch/autograd/profiler.py __enter__szContextDecorator.__enter__cCstdSrrr exc_typeexc_valexc_tbr!r!r"__exit__szContextDecorator.__exit__cstfdd}|S)Ncs6|i|WdS1s(0YdSrr!)argskwargsfuncr r!r"wrapped"sz*ContextDecorator.__call__..wrapped) functoolswraps)r r,r-r!r+r"__call__!szContextDecorator.__call__N)__name__ __module__ __qualname__r#r(r0r!r!r!r"rsrc @seZdZdZd+dddddddddd ddZdd Zd d Zd d ZddZddZ ddZ ddZ ddZ d,ddZ ej je _ddZejje_d-eedd d!Zd.d#d$Zejje_d%d&Zejje_ed'd(Zd)d*ZdS)/profileaContext manager that manages autograd profiler state and holds a summary of results. Under the hood it just records events of functions being executed in C++ and exposes those events to Python. You can wrap any code into it and it will only report runtime of PyTorch functions. Note: profiler is thread local and is automatically propagated into the async tasks Args: enabled (bool, optional): Setting this to False makes this context manager a no-op. use_cuda (bool, optional): Enables timing of CUDA events as well using the cudaEvent API. Adds approximately 4us of overhead to each tensor operation. record_shapes (bool, optional): If shapes recording is set, information about input dimensions will be collected. This allows one to see which dimensions have been used under the hood and further group by them using prof.key_averages(group_by_input_shape=True). Please note that shape recording might skew your profiling data. It is recommended to use separate runs with and without shape recording to validate the timing. Most likely the skew will be negligible for bottom most events (in a case of nested function calls). But for higher level functions the total self cpu time might be artificially increased because of the shape collection. with_flops (bool, optional): If with_flops is set, the profiler will estimate the FLOPs (floating point operations) value using the operator's input shape. This allows one to estimate the hardware performance. Currently, this option only works for the matrix multiplication and 2D convolution operators. profile_memory (bool, optional): track tensor memory allocation/deallocation. with_stack (bool, optional): record source information (file and line number) for the ops. with_modules (bool): record module hierarchy (including function names) corresponding to the callstack of the op. e.g. If module A's forward call's module B's forward which contains an aten::add op, then aten::add's module hierarchy is A.B Note that this support exist, at the moment, only for TorchScript models and not eager mode models. use_kineto (bool, optional): experimental, enable profiling with Kineto profiler. use_cpu (bool, optional): profile CPU events; setting to ``False`` requires ``use_kineto=True`` and can be used to lower the overhead for GPU-only profiling. experimental_config (_ExperimentalConfig) : A set of experimental options used by profiler libraries like Kineto. Note, backward compatibility is not guaranteed. .. warning: Enabling memory profiling or source attribution incurs additional profiler overhead .. warning: This context managers should not be called recursively, i.e. no nested instances are allowed .. warning: Due to some CUDA multiprocessing limitations (multiprocessing-cuda-note_), one cannot use the profiler with ``use_cuda = True`` to benchmark DataLoaders with ``num_workers > 0``. If you wish to benchmark data loading, please use ``use_cuda = False`` or ``num_workers = 0``. Example: >>> x = torch.randn((1, 1), requires_grad=True) >>> with torch.autograd.profiler.profile() as prof: >>> for _ in range(100): # any normal python code, really! >>> y = x ** 2 >> y.backward() >>> # NOTE: some columns were removed for brevity >>> print(prof.key_averages().table(sort_by="self_cpu_time_total")) ----------------------------------- --------------- --------------- --------------- Name Self CPU total CPU time avg Number of Calls ----------------------------------- --------------- --------------- --------------- mul 32.048ms 32.048ms 200 pow 27.041ms 27.041ms 200 PowBackward0 9.727ms 55.483ms 100 torch::autograd::AccumulateGrad 9.148ms 9.148ms 100 torch::autograd::GraphRoot 691.816us 691.816us 100 ----------------------------------- --------------- --------------- --------------- TFN) use_cuda record_shapes with_flopsprofile_memory with_stack with_modules use_kinetouse_cpuexperimental_configc Cs||_|jsdS||_d|_d|_||_||_|j|jO_||_||_||_| |_ | durdt } | |_ d|_ |j s|sJd|jrt jstdd|_t|_|j r|jtjtj|_|jr|rtjtvr|j sJdtj|_n|jtjt|jdksJddS)NFz?Device-only events supported only with Kineto (use_kineto=True)z/CUDA is not available, disabling CUDA profilingz+Legacy CUDA profiling requires use_cpu=Truerz(No activities specified for the profiler)enabledr5function_eventsenteredr6r7r8r9r:r<rr=kineto_resultstorchcuda is_availablersetkineto_activitiesaddr CPUr KINETO profiler_kindCUDArKINETO_GPU_FALLBACKlen) r r>r5r6r7r8r9r:r;r<r=r!r!r"__init__|sL  zprofile.__init__cCs"t|j|j|j|j|j|j|jSr)r rJr6r8r9r7r:r=rr!r!r"configszprofile.configcCs,|js dS|jrtd|||S)Nz)Profiler context manager is not reentrant)r>r@ RuntimeError_prepare_trace _start_tracerr!r!r"r#szprofile.__enter__cCsd|_t||jdSNT)r@rrOrFrr!r!r"rQszprofile._prepare_tracecCsd|_t||jdSrS)r@rrOrFrr!r!r"rRszprofile._start_tracecCsT|js dS|jrtjt|_||j}t||j|j |j d|_ |j dS)N)r5r8r7F) r>r5rBrC synchronizerrA_parse_kineto_resultsrr8r7r? _build_tree)r r%r&r'Zparsed_resultsr!r!r"r(s   zprofile.__exit__cCs|jdurdSt|jSNz#)r?reprrr!r!r"__repr__s zprofile.__repr__cCs|jdurdSt|jSrW)r?strrr!r!r"__str__s zprofile.__str__cCs|jdurtddS)NzProfiler didn't finish running)r?rPrr!r!r" _check_finishs zprofile._check_finishdKcCs,||jdusJ|jj|||||dS)N)sort_by row_limitmax_src_column_widthheadertop_level_events_only)r\r?table)r r_r`rarbrcr!r!r"rds z profile.tablecCs,|tr|j|n |j|SdSr)r\r rAsaver?export_chrome_trace)r pathr!r!r"rfszprofile.export_chrome_traceself_cpu_time_total)rgmetriccCs6||jdusJd|js(Jd|j||S)NExpected profiling resultsz(export_stacks() requires with_stack=True)r\r?r9 export_stacks)r rgrir!r!r"rkszprofile.export_stacksrcCs(||jdusJd|j||SNrj)r\r? key_averages)r Zgroup_by_input_shapeZgroup_by_stack_nr!r!r"rmszprofile.key_averagescCs$||jdusJd|jSrl)r\r? total_averagerr!r!r"rn szprofile.total_averagecCs||jdusJ|jjS)zl Returns total time spent on CPU obtained as a sum of all self times across all the events. N)r\r?rhrr!r!r"rhszprofile.self_cpu_time_totalcCs|}dd|D}t|}dd}dd}g}i}d} |D]} t| rZqF| |} | | } | | } d}d}| tj kr| | | D],}|||d7}|||d7}d|d <q| p| | k}t| t| dd t| d d | | | | | d d| D| |||| | | | d }|j| krt|jn| } |jtj kr|j s| }|dkr||j|j|d|_||| }|dkrF||vrg||<|||qF|D]~}|jtj kr|j s|j|vr||jD]J}|jtjkrZ||j|j|jj |jj!n|jtj kr(|j"|_"q(q|D]|}|d s||d|} | d 7} t| t#d|d | | |d ggd||d||dd dtj dd}||q||j$ddd|S)NcSs g|]}|tkr|dgqS)F)namer).0evtr!r!r" z1profile._parse_kineto_results..cSs$|tjtjtjfvr |SdSNr) device_typer rHMKLDNNIDEEPnbytes mem_recordr!r!r"_cpu_memory_usages z8profile._parse_kineto_results.._cpu_memory_usagecSs |tjtjfvr|SdSrt)rur rKHIPrxryr!r!r"_cuda_memory_usage$s z9profile._parse_kineto_results.._cuda_memory_usagerT)roZ with_wildcardFcSsg|]}t|r|qSr!)r)rpentryr!r!r"rrJrs)idro trace_namethreadstart_usend_us fwd_thread input_shapesstackscopecpu_memory_usagecuda_memory_usageis_async sequence_nrru device_indexflops)rrorrrrrrrrrrrrrurcSs|jj|jj gSr) time_rangestartendrqr!r!r"rsz/profile._parse_kineto_results..key)%trace_start_useventsrrror duration_usrur rHZ in_intervalrstart_thread_id end_thread_idrcorrelation_idr fwd_thread_idshapesrrrrrrcuda_elapsed_us append_kernel is_legacyappendlinked_correlation_idrKrrrrrsort)r resultrZ mem_recordsZmem_records_accr{r}r?Z cuda_corr_mapZ max_evt_idZ kineto_eventZ rel_start_usZ rel_end_usZ abs_end_usrrrzrZfeZ cuda_timeZcorr_idZf_evtr!r!r"rUs            zprofile._parse_kineto_results)T)Nr]r^NF)rh)Fr)r1r2r3__doc__rNrOr#rQrRr(rYr[r\rdrrfrZrkrmrnpropertyrhrUr!r!r!r"r4*sBS 8         r4c@sXeZdZdZdeeedddZddZeeedd d Z e ee ed d d Z dS)record_functionaContext manager/function decorator that adds a label to a block of Python code (or function) when running autograd profiler. It is useful when tracing the code profile. Args: name (str): Label assigned to the block of code. node_id (int): ID of node, for distributed profiling. Unset in non-distributed cases. Example: >>> x = torch.randn((1, 1), requires_grad=True) >>> with torch.autograd.profiler.profile() as prof: ... y = x ** 2 ... with torch.autograd.profiler.record_function("label-z"): # label the block ... z = y ** 3 ... y.backward() ... >>> # NOTE: some columns were removed for brevity >>> print(prof.key_averages().table(sort_by="self_cpu_time_total")) ----------------------------------- --------------- --------------- --------------- Name Self CPU total % CPU time avg Number of Calls ----------------------------------- --------------- --------------- --------------- pow 60.77% 47.470us 3 mul 21.73% 25.465us 2 PowBackward0 12.03% 121.891us 1 torch::autograd::AccumulateGrad 2.70% 6.324us 1 label-z 2.13% 12.421us 1 torch::autograd::GraphRoot 0.64% 1.503us 1 ----------------------------------- --------------- --------------- --------------- Self CPU time total: 234.344us CUDA time total: 0.000us N)ror)cCs"||_||_d|_td|_dS)NTr~)ror)run_callbacks_on_exitrBzeroshandle)r ror)r!r!r"rNszrecord_function.__init__cCstjj|j|j|_|Sr)rBopsprofilerZ_record_function_enterror)rrr!r!r"r#szrecord_function.__enter__)r% exc_value tracebackcCs|jrtjj|jdSr)rrBrrZ_record_function_exitr)r r%rrr!r!r"r(szrecord_function.__exit__)futreturncCs*|jstdd|_tjj|j|}|S)a _call_end_callbacks_on_future is meant to be used for profiling async calls that return a future. Calling this function will extend recording beyond this scope, until the future is satisfied. It is useful for profiling the end to end time of asynchronous calls. This function should only be called once to attach the callback onto the future, and will throw if called multiple times. Args: fut: (torch._C.Future): future for which to schedule callback for. Returns: A future that completes with the value of the passed in future when the profiling callbacks have ran. z6_call_end_callbacks_on_future can only be called once.F)rrPrBrrZ_call_end_callbacks_on_jit_futr)r rZprofiled_futurer!r!r"_call_end_callbacks_on_futures z-record_function._call_end_callbacks_on_future)N) r1r2r3rrZrrNr#rr(rrr!r!r!r"rs ! rc@s*eZdZdZd ddZddZdd Zd S) emit_nvtxaContext manager that makes every autograd operation emit an NVTX range. It is useful when running the program under nvprof:: nvprof --profile-from-start off -o trace_name.prof -- Unfortunately, there's no way to force nvprof to flush the data it collected to disk, so for CUDA profiling one has to use this context manager to annotate nvprof traces and wait for the process to exit before inspecting them. Then, either NVIDIA Visual Profiler (nvvp) can be used to visualize the timeline, or :func:`torch.autograd.profiler.load_nvprof` can load the results for inspection e.g. in Python REPL. .. warning: This context manager should not be called recursively, i.e. at most one instance should be enabled at any given time. Args: enabled (bool, optional, default=True): Setting ``enabled=False`` makes this context manager a no-op. Default: ``True``. record_shapes (bool, optional, default=False): If ``record_shapes=True``, the nvtx range wrapping each autograd op will append information about the sizes of Tensor arguments received by that op, in the following format: ``[[arg0.size(0), arg0.size(1), ...], [arg1.size(0), arg1.size(1), ...], ...]`` Non-tensor arguments will be represented by ``[]``. Arguments will be listed in the order they are received by the backend op. Please note that this order may not match the order in which those arguments were passed on the Python side. Also note that shape recording may increase the overhead of nvtx range creation. Example: >>> with torch.cuda.profiler.profile(): ... model(x) # Warmup CUDA memory allocator and profiler ... with torch.autograd.profiler.emit_nvtx(): ... model(x) **Forward-backward correlation** When viewing a profile created using :class:`emit_nvtx` in the Nvidia Visual Profiler, correlating each backward-pass op with the corresponding forward-pass op can be difficult. To ease this task, :class:`emit_nvtx` appends sequence number information to the ranges it generates. During the forward pass, each function range is decorated with ``seq=``. ``seq`` is a running counter, incremented each time a new backward Function object is created and stashed for backward. Thus, the ``seq=`` annotation associated with each forward function range tells you that if a backward Function object is created by this forward function, the backward object will receive sequence number N. During the backward pass, the top-level range wrapping each C++ backward Function's ``apply()`` call is decorated with ``stashed seq=``. ``M`` is the sequence number that the backward object was created with. By comparing ``stashed seq`` numbers in backward with ``seq`` numbers in forward, you can track down which forward op created each backward Function. Any functions executed during the backward pass are also decorated with ``seq=``. During default backward (with ``create_graph=False``) this information is irrelevant, and in fact, ``N`` may simply be 0 for all such functions. Only the top-level ranges associated with backward Function objects' ``apply()`` methods are useful, as a way to correlate these Function objects with the earlier forward pass. **Double-backward** If, on the other hand, a backward pass with ``create_graph=True`` is underway (in other words, if you are setting up for a double-backward), each function's execution during backward is given a nonzero, useful ``seq=``. Those functions may themselves create Function objects to be executed later during double-backward, just as the original functions in the forward pass did. The relationship between backward and double-backward is conceptually the same as the relationship between forward and backward: The functions still emit current-sequence-number-tagged ranges, the Function objects they create still stash those sequence numbers, and during the eventual double-backward, the Function objects' ``apply()`` ranges are still tagged with ``stashed seq`` numbers, which can be compared to `seq` numbers from the backward pass. .. warning: The sequence number is thread-local, and some forward functions don't create an associated backward Function object (instead delegating that to sub-functions further down the call chain). For these reasons, the correspondence of stashed sequence numbers in backward Function ``apply()`` ranges with `seq` numbers in forward-pass ranges is not guaranteed to be 1 to 1. The sequence numbers alone may not be enough to fully disambiguate which forward function created which backward Function object. You may need to make a judgment based on analytic knowledge of what the expected correspondence should be. TFcCs||_d|_||_dSNF)r>r@r6)r r>r6r!r!r"rN3szemit_nvtx.__init__c CsN|js dS|jrtdd|_tjtttj |j ddddt t |S)Nz0NVTX annotation context manager is not reentrantTF) r>r@rPrBrCrTrr r NVTXr6rrErr!r!r"r#8s&  zemit_nvtx.__enter__cCs|js dStjtdSr)r>rBrCrTrr$r!r!r"r(Ls  zemit_nvtx.__exit__N)TF)r1r2r3rrNr#r(r!r!r!r"rsP rcCs tt|S)ztOpens an nvprof trace file and parses autograd annotations. Args: path (str): path to nvprof trace )rparse_nvprof_trace)rgr!r!r" load_nvprofTsrc@s eZdZdZddZddZdS) EnforceUniquez0Raises an error if a key is seen more than once.cCs t|_dSr)rEseenrr!r!r"rN_szEnforceUnique.__init__cGs*||jvrtdt||j|dS)Nzduplicate key: )rrPrZrG)r rr!r!r"seebs zEnforceUnique.seeN)r1r2r3rrNrr!r!r!r"r]src Cs.ddl}||}|j|_i}|dD]}tj|d||d<q(d}g}i}t}||D]N} | | dt | dd|| d| d| d dd } | | | || j <qbd } t}|| D]R} | | d| d | d dksJ|| d} | | dd| d| dq|jddd|S)Nrz)SELECT _id_ as id, value FROM StringTablevaluera& SELECT start.id AS marker_id, start.name, start.timestamp AS start_time, end.timestamp AS end_time FROM CUPTI_ACTIVITY_KIND_MARKER AS start INNER JOIN CUPTI_ACTIVITY_KIND_MARKER AS end ON start.id = end.id WHERE start.name != 0 AND end.name = 0 Z marker_idro start_timeend_time)rnode_idrorrra SELECT start.id AS marker_id, start.name, start.timestamp, end.timestamp, runtime._id_ AS runtime_id, runtime.cbid, runtime.start AS runtime_start, runtime.end AS runtime_end, kernel.start AS kernel_start, kernel.end AS kernel_end, kernel.name AS kernel_name FROM CUPTI_ACTIVITY_KIND_MARKER AS start INNER JOIN CUPTI_ACTIVITY_KIND_MARKER AS end ON start.id = end.id INNER JOIN CUPTI_ACTIVITY_KIND_RUNTIME as runtime ON (start.timestamp < runtime.start AND runtime.end < end.timestamp) INNER JOIN CUPTI_ACTIVITY_KIND_CONCURRENT_KERNEL AS kernel ON kernel.correlationId = runtime.correlationId Z runtime_idZcbid kernel_nameZ kernel_endZ kernel_startcSs|jjSr)rrrr!r!r"rrsz$parse_nvprof_trace..r)sqlite3connectRowZ row_factoryexecuterB_C _demanglerrrrrrr) rgrconnstringsrZ marker_query functionsZ functions_mapuniquerowrqZ kernel_queryr!r!r"rhsB       rcCs tdS)zc Notify kineto so it is aware of iteration boundaries for asynchronous trace requests. N)rr!r!r!r" kineto_stepsr)-Ztorch.autograd.profiler_utilrrrrrrrtorch.autogradr r r r r rrrrrrtorch._C._autogradrrB torch.cuda torch.futuresrtypingrrrrwarningsr contextlibr ImportErrorr.objectr4rrrrrrr!r!r!r"s*$4    jPr  @