o Qe̙ @sddlZddlZddlZddlmZddlmZmZmZm Z m Z ddl m Z ddl Z ddlZddlZddlmZmZmZmZmZmZmZmZddlmZmZddlmZmZmZdd l mZdd l!m"Z"Gd d d eZ#Gd ddeZ$GdddeZ%dddde&de&de&de&de&def ddZ'de&fddZ( d)de)de e)defdd Z* d)de)de e)defd!d"Z+dee%fd#d$Z,Gd%d&d&Z-d'd(Z.dS)*N)Enum)AnyCallableIterableOptionalUnion)warn) _Profiler_ProfilerResultProfilerOptionsTracerEventTypeenable_memory_recorderenable_input_shape_recorderdisable_memory_recorderdisable_input_shape_recorder) RecordEventwrap_optimizers) StatisticData _build_table SortedKeys)utils) benchmarkc@s4eZdZdZdZdZdZdZdZdZ dZ d Z d Z d S) SummaryViewa SummaryView define the summary view of different contents. - **SummaryView.DeviceView** : The device summary view. - **SummaryView.OverView** : The overview summary view. - **SummaryView.ModelView** : The model summary view. - **SummaryView.DistributedView** : The distributed summary view. - **SummaryView.KernelView** : The kernel summary view. - **SummaryView.OperatorView** : The operator summary view. - **SummaryView.MemoryView** : The memory summary view. - **SummaryView.MemoryManipulationView** : The meomory manipulation summary view. - **SummaryView.UDFView** : The user defined summary view. rrN) __name__ __module__ __qualname____doc__Z DeviceViewZOverViewZ ModelViewZDistributedViewZ KernelViewZ OperatorViewZ MemoryViewZMemoryManipulationViewZUDFViewr%r%HD:\Projects\ConvertPro\env\Lib\site-packages\paddle/profiler/profiler.pyr*src@ eZdZdZdZdZdZdZdS) ProfilerStatea ProfilerState is used to present the state of :ref:`Profiler ` . The meaning of each ProfilerState is as following - **ProfilerState.CLOSED** : The profiler is closed, and no profiling data will be recorded. - **ProfilerState.READY** : The profiler is open, but the data will not be recorded. This state is used for reducing overhead influence when profiler starts. - **ProfilerState.RECORD** : The profiler is open, and the data will be recorded. - **ProfilerState.RECORD_AND_RETURN** : The profiler is open, and this state stands for the last batch of "RECORD" state in current profiling period. The collected data will be returned in this state. rrrrN)r!r"r#r$CLOSEDREADYRECORDRECORD_AND_RETURNr%r%r%r&r(Ks  r(c@r')ProfilerTargetaz ProfilerTarget is used to specify target device for :ref:`profiling ` . Only CPU, GPU and MLU are supported currently. The meaning of each ProfilerState is as following - **ProfilerTarget.CPU** : Profile events on CPU. - **ProfilerTarget.GPU** : Profile events on GPU. - **ProfilerTarget.MLU** : Profile events on MLU. rrrrN)r!r"r#r$CPUGPUMLU CUSTOM_DEVICEr%r%r%r&r-_s  r-)repeat skip_firstclosedreadyrecordr2r3returncsbdtdtffdd }dkr#dkr#dkr#dkr#dks'Jddkr/td|S)a\ Return a scheduler function, which scheduler the :ref:`state ` according to the setting. The state transform confirms to: .. code-block:: text (CLOSED) (CLOSED) (CLOSED) (READY) (RECORD,last RETURN) (CLOSED) START -> skip_first -> closed -> ready -> record -> END | | | | (if has_repeated < repeat) - - - - - - - - - - - - Note that repeat <= 0 means the cycle will continue until the profiler exits. Args: closed(int): The number of steps in state ProfilerState.CLOSED. ready(int): The number of steps in state ProfilerState.READY. record(int): The number of steps in state ProfilerState.RECORD, and the state in last step will be set as ProfilerState.RECORD_AND_RETURN. repeat(int, optional): The number of cycles to repeat above state transform. Default value is 0, which means it will repeat this cycle until profiler exits. skip_first(int, optional): The number of first steps to drop, not participate in the state transform, and at ProfilerState.CLOSED state. Default value is 0. Returns: A scheduler function, conforms to above state transform setting. The function will takes one parameter `step_num`, and returns corresponding ProfilerState. Examples: 1. profiling range [2, 5]. Assume batch 0: closed, batch 1: ready, batch [2, 5] record. .. code-block:: python :name: code-example1 import paddle.profiler as profiler profiler.make_scheduler(closed=1, ready=1, record=4, repeat=1) 2. profiling range [3,6], [9,12], [15,18]. Assume batch 0: skiped, batch 1: closed, batch 2: ready, batch [3,6]: record, repeat. .. code-block:: python :name: code-example2 import paddle.profiler as profiler profiler.make_scheduler(closed=1, ready=1, record=4, skip_first=1) stepr7cs|dksJ|kr tjS|}}||}dkr&|kr&tjS||}|kr1tjS|kr>|kr>tjS||dkrGtjStjS)Nrr)r(r)r*r+r,)r8Z period_stepsZ has_repeatedZmod_stepr4r5r6r2r3r%r&getScheduleStates   z(make_scheduler..getScheduleStaterz$Invalid profiler scheduler argumentszProfiler will record data after enabling profiler immediately, some data collected at the beginning of profiling may be 'noisy' because of overhead.)intr(r)r4r5r6r2r3r:r%r9r&make_schedulerqs6 r<r8cCstjS)zd A default state scheduler, keep recording from the beginning of the profiler until ending. )r(r+)r8r%r%r&_default_state_schedulersr=dir_name worker_namecNtjsz tjddWntytdwfdd}|S)a Return a callable, used for outputing tracing data to chrome tracing format file. The output file will be saved in directory ``dir_name``, and file name will be set as `worker_name`. if `worker_name` is not set, the default name is `[hostname]_[pid]`. Args: dir_name(str): Directory to save profiling data. worker_name(str, optional): Prefix of the file name saved, default is `[hostname]_[pid]`. Returns: A callable, which takes a Profiler object as parameter and calls its export method to save data to chrome tracing format file. Examples: The return value can be used as parameter ``on_trace_ready`` in :ref:`Profiler ` . .. code-block:: python # required: gpu import paddle.profiler as profiler with profiler.Profiler( targets=[profiler.ProfilerTarget.CPU, profiler.ProfilerTarget.GPU], scheduler = (3, 10), on_trace_ready=profiler.export_protobuf('./log')) as p: for iter in range(10): #train() p.step() Texist_ok;Can not create directory '{}' for saving profiling results.cRsdttttj}d|d}| tj |ddS)N host_{}pid_{}z{}_time_{}.paddle_trace.json%Y_%m_%d_%H_%M_%S_%fjson formatsocket gethostnamestrosgetpiddatetimenowstrftimeexportpathjoinZprofrPfilenamer>r?r%r& handle_fn  z(export_chrome_tracing..handle_fnrMrSexistsmakedirs Exception RuntimeErrorrIr>r?rXr%rWr&export_chrome_tracing   r`cr@)az Return a callable, used for outputing tracing data to protobuf file. The output file will be saved in directory ``dir_name``, and file name will be set as ``worker_name``. if ``worker_name`` is not set, the default name is `[hostname]_[pid]`. Args: dir_name(str): Directory to save profiling data. worker_name(str, optional): Prefix of the file name saved, default is `[hostname]_[pid]`. Returns: A callable, which takes a Profiler object as parameter and calls its export method to save data to protobuf file. Examples: The return value can be used as parameter ``on_trace_ready`` in :ref:`Profiler ` . .. code-block:: python # required: gpu import paddle.profiler as profiler with profiler.Profiler( targets=[profiler.ProfilerTarget.CPU, profiler.ProfilerTarget.GPU], scheduler = (3, 10), on_trace_ready = profiler.export_protobuf('./log')) as p: for iter in range(10): #train() p.step() TrArCcrD)NrEz{}_time_{}.paddle_trace.pbrFZpbrHrUrWr%r&rX2rYz"export_protobuf..handle_fnrZr_r%rWr&export_protobuf rarbcCs<tr tjtjtjgStrtjtjtjgStjtjgS)zB Get the current supported profiler target in the system. )r Zis_cupti_supportedr-r.r/r1Zis_cnpapi_supportedr0r%r%r%r&_get_supported_targetsAs rcc@seZdZdZdddddddgddeeedeee ge fe dfdeede fd ee d ee d ee d eefd dZddZddZddZddZd&dee fddZd&ddZddZd'd d!Zejd"dd#dfd$d%ZdS)(Profilera  Profiler context manager, user interface to manage profiling process to start, stop, export profiling data and print summary table. Args: targets (list, optional): specify target devices to profile, and all existing and supported devices will be chosen by default. Currently supported values, :ref:`ProfilerTarget.CPU ` , :ref:`ProfilerTarget.GPU ` and :ref:`ProfilerTarget.MLU ` . scheduler (Callable|tuple, optional): If it is a callable object, it takes a step number as parameter and return the corresponding :ref:`ProfilerState `. This callable object can be generated by :ref:`make_scheduler ` function. If not provided (None), the default scheduler will keep tracing until the profiler exits. If it is a tuple, it has two values start_batch and end_batch, which means profiling range [start_batch, end_batch). on_trace_ready (Callable, optional): Callable object, serves as callback function, and takes the Profiler object as parameter, which provides a way for users to do post-processing. This callable object will be called when ``scheduler`` returns ``ProfilerState.RECORD_AND_RETURN``. The default value is :ref:`export_chrome_tracing `. timer_only (bool, optional): If it is True, the cost of Dataloader and every step of the model will be count without profiling. Otherwise, the model will be timed and profiled. Default: False. record_shapes (bool, optional): If it is True, collect op's input shape information. Default: False. profile_memory (bool, optional): If it is True, collect tensor memory allocation and release information. Default: False. Examples: 1. profiling range [2, 5). .. code-block:: python :name: code-example1 # required: gpu import paddle.profiler as profiler with profiler.Profiler( targets=[profiler.ProfilerTarget.CPU, profiler.ProfilerTarget.GPU], scheduler = (2, 5), on_trace_ready = profiler.export_chrome_tracing('./log')) as p: for iter in range(10): #train() p.step() 2. profiling range [2,4], [7, 9], [11,13]. .. code-block:: python :name: code-example2 # required: gpu import paddle.profiler as profiler with profiler.Profiler( targets=[profiler.ProfilerTarget.CPU, profiler.ProfilerTarget.GPU], scheduler = profiler.make_scheduler(closed=1, ready=1, record=3, repeat=3), on_trace_ready = profiler.export_chrome_tracing('./log')) as p: for iter in range(10): #train() p.step() 3. Use profiler without context manager, and use default parameters. .. code-block:: python :name: code-example3 # required: gpu import paddle.profiler as profiler p = profiler.Profiler() p.start() for iter in range(10): #train() p.step() p.stop() p.summary() 4. Use profiler to get throughput and cost of the model. .. code-block:: python :name: code-example-timer1 import paddle import paddle.profiler as profiler class RandomDataset(paddle.io.Dataset): def __init__(self, num_samples): self.num_samples = num_samples def __getitem__(self, idx): image = paddle.rand(shape=[100], dtype='float32') label = paddle.randint(0, 10, shape=[1], dtype='int64') return image, label def __len__(self): return self.num_samples class SimpleNet(paddle.nn.Layer): def __init__(self): super(SimpleNet, self).__init__() self.fc = paddle.nn.Linear(100, 10) def forward(self, image, label=None): return self.fc(image) dataset = RandomDataset(20 * 4) simple_net = SimpleNet() opt = paddle.optimizer.SGD(learning_rate=1e-3, parameters=simple_net.parameters()) BATCH_SIZE = 4 loader = paddle.io.DataLoader( dataset, batch_size=BATCH_SIZE) p = profiler.Profiler(timer_only=True) p.start() for i, (image, label) in enumerate(loader()): out = simple_net(image) loss = paddle.nn.functional.cross_entropy(out, label) avg_loss = paddle.mean(loss) avg_loss.backward() opt.minimize(avg_loss) simple_net.clear_gradients() p.step(num_samples=BATCH_SIZE) if i % 10 == 0: step_info = p.step_info(unit='images') print("Iter {}: {}".format(i, step_info)) # The average statistics for 10 steps between the last and this call will be # printed when the "step_info" is called at 10 iteration intervals. # The values you get may be different from the following. # Iter 0: reader_cost: 0.51946 s batch_cost: 0.66077 s ips: 6.054 images/s # Iter 10: reader_cost: 0.00014 s batch_cost: 0.00441 s ips: 907.009 images/s p.stop() # The performance summary will be automatically printed when the "stop" is called. # Reader Ratio: 2.658% # Time Unit: s, IPS Unit: images/s # | | avg | max | min | # | reader_cost | 0.00011 | 0.00013 | 0.00007 | # | batch_cost | 0.00405 | 0.00434 | 0.00326 | # | ips | 1086.42904 | 1227.30604 | 959.92796 | NF)targets scheduleron_trace_ready record_shapesprofile_memory timer_only emit_nvtxcustom_device_typesrerfrg.rhrjrkrlcCst} |r!t||_|D]} | | vr|j| td| q n| |_t} tj|jvr4| j dO_ tj |jvrA| j dO_ tj |jvrN| j dO_ tj |jvrb| j dO_ |sbt j}tt| ||_t|rt||_nDt|ttfrt|dkr|d|dksJ|\} } t| d} | dkrtt| ddd| | dd|_ntdd| | dd|_nt|_|dkrtd|_n||_d|_t j!|_"||j|_#d|_$d|_%||_&||_'||_(||_)dS) Nz1Profiling {} is not supported in current context.rrrr r)r4r5r6r2z./profiler_log/)*rcsetreremoverrIr r-r.Z trace_switchr/r0r1paddleZdeviceZget_all_custom_device_typerr createprofilercallablerf isinstancetuplelistlenmaxr<r=r`rgstep_numr(r)previous_state current_state record_eventprofiler_resultrjrhrirk)selfrerfrgrhrirjrkrlZsupported_targetstargetZ profileoptionZ start_batchZ end_batchr%r%r&__init__sv                zProfiler.__init__cCs ||SN)startr}r%r%r& __enter__szProfiler.__enter__cCs |dSr)stop)r}exc_typeexc_valexc_tbr%r%r&__exit__ s zProfiler.__exit__cCst|jr |jrdt_|jrdS|jrt|jrt |j t j kr+|j n!|j t jkr<|j |j n|j t jkrL|j |j td|jtjd|_|jdS)a Start profiler and enter the first profiler step(0). State transformed from CLOSED to self.current_state and trigger corresponding action. Examples: .. code-block:: python :name: code-example4 # required: gpu import paddle.profiler as profiler prof = profiler.Profiler( targets=[profiler.ProfilerTarget.CPU, profiler.ProfilerTarget.GPU], scheduler = (1, 9), on_trace_ready = profiler.export_chrome_tracing('./log')) prof.start() for iter in range(10): #train() prof.step() prof.stop() TNProfileStep#{}nameZ event_type)rbeginrjrkr_is_profiler_usedrhrrir rzr(r*rqpreparer+rr,rrIrxr ProfileStepr{rr%r%r&r#s,           zProfiler.startcCst|jr dS|jrt|jrt|jr!|jd|_|jt j kr5t d|j |j |jt jksA|jt jkrO|j |_|jrO||dt_dS)a Stop profiler and State transformed from self.current_state to CLOSED. Trigger corresponding action and post-process profiler result using self.on_trace_ready if result exists. Examples: .. code-block:: python :name: code-example5 # required: gpu import paddle.profiler as profiler prof = profiler.Profiler( targets=[profiler.ProfilerTarget.CPU, profiler.ProfilerTarget.GPU], scheduler = (1, 7), on_trace_ready = profiler.export_chrome_tracing('./log')) prof.start() for iter in range(10): #train() prof.step() prof.stop() NzbInproper Profiler state transform: READY->CLOSED, profiler will start and stop without saving dataF)rendrjrhrrirr{rzr(r*rrqrrr+r,r|rgrrrr%r%r&rRs,          z Profiler.stop num_samplescCs~t||jr dS|jr|jd|_|j|_|jd7_||j|_| t d |jt j d|_|jdS)a Signals the profiler that the next profiling step has started. Get the new ProfilerState and trigger corresponding action. Args: num_samples (int|None, optional): Specifies the batch size of every step of the model that is used to compute throughput when `timer_only` is True. Default: None. Examples: .. code-block:: python :name: code-example6 # required: gpu import paddle.profiler as profiler prof = profiler.Profiler( targets=[profiler.ProfilerTarget.CPU, profiler.ProfilerTarget.GPU], scheduler = (3, 7), on_trace_ready = profiler.export_chrome_tracing('./log')) prof.start() for iter in range(10): #train() prof.step() prof.stop() Nrrr)rr8rjr{rrzryrxrf_trigger_actionrrIr rr)r}rr%r%r&r8s   z Profiler.stepcCs|durd}t|S)a Get statistics for current step. If the function is called at certain iteration intervals, the result is the average of all steps between the previous call and this call. Statistics are as follows: 1. reader_cost: the cost of loading data measured in seconds. 2. batch_cost: the cost of step measured in seconds. 3. ips(Instance Per Second): the throughput of the model measured in `samples/s` or others depends on the `unit`. When `num_samples` of `step()` is None, it is measured in `steps/s`. Args: unit (string, optional): The unit of input data is only used When `num_samples` of `step()` is specified as a number. For example, when it is `images`, the unit of throughput is `images/s`. Default: None, the unit of throughput is `samples/s`. Returns: string: A string representing the statistic. Examples: .. code-block:: python :name: code-example-timer2 import paddle.profiler as profiler prof = profiler.Profiler(timer_only=True) prof.start() for iter in range(20): #train() prof.step() if iter % 10 == 0: print("Iter {}: {}".format(iter, prof.step_info())) # The example does not call the DataLoader, so there is no "reader_cost". # Iter 0: batch_cost: 0.00001 s ips: 86216.623 steps/s # Iter 10: batch_cost: 0.00001 s ips: 103645.034 steps/s prof.stop() # Time Unit: s, IPS Unit: steps/s # | | avg | max | min | # | batch_cost | 0.00000 | 0.00002 | 0.00000 | # | ips | 267846.19437 | 712030.38727 | 45134.16662 | NZsamples)r step_info)r}unitr%r%r&rs+ zProfiler.step_infocCs|jtjkr5|jtjkr|j|jtjkr!|j|j|jtj kr3|j|jdSdS|jtjkri|jtjkrOt d|j|j |jtjkrZ|j|jtj krg|jdSdS|jtjkr|jtjkr~t d|j |jtjkrt d|j |j|jtj kr dSdS|jtj ksJ|jtjkr|j |_ |jtjkr|j |_ |j|jtjkr|j |_ |j|j|jtj kr|j |_ |j|j|j r| |dSdS)NzRImproper schedule: READY->CLOSED, profiler will start and stop without saving dataz@Improper schedule: RECORD->CLOSED, profiler will not saving datazCImproper schedule: RECORD->READY, profiler will stop and re-prepare)ryr(r)rzr*rqrr+rr,rrr|rgrr%r%r&rsl                                   zProfiler._trigger_actionrGcCs|jr |j||dSdS)aP Exports the tracing data to file. Args: path(str): file path of the output. format(str, optional): output format, can be chosen from ['json', 'pb'], 'json' for chrome tracing and 'pb' for protobuf, default value is 'json'. Examples: .. code-block:: python :name: code-example7 # required: gpu import paddle.profiler as profiler prof = profiler.Profiler( targets=[profiler.ProfilerTarget.CPU, profiler.ProfilerTarget.GPU], scheduler = (3, 7)) prof.start() for iter in range(10): #train() prof.step() prof.stop() prof.export(path="./profiler_data.json", format="json") N)r|save)r}rSrIr%r%r&rR"szProfiler.exportTmsc CsLt|tr|g}|jr$t|j|j}tt||||||ddSdS)a Print the Summary table. Currently support overview, model, distributed, operator, memory manipulation and userdefined summary. Args: sorted_by( :ref:`SortedKeys ` , optional): how to rank the op table items, default value is SortedKeys.CPUTotal. op_detail(bool, optional): expand each operator detail information, default value is True. thread_sep(bool, optional): print op table each thread, default value is False. time_unit(str, optional): time unit for display, can be chosen form ['s', 'ms', 'us', 'ns'], default value is 'ms'. views(SummaryView|list[SummaryView], optional): summary tables to print, default to None means all views to be printed. Examples: .. code-block:: python :name: code-example8 # required: gpu import paddle.profiler as profiler prof = profiler.Profiler( targets=[profiler.ProfilerTarget.CPU, profiler.ProfilerTarget.GPU], scheduler = (3, 7), on_trace_ready = profiler.export_chrome_tracing('./log')) prof.start() for iter in range(10): #train() prof.step() prof.stop() prof.summary(sorted_by=profiler.SortedKeys.CPUTotal, op_detail=True, thread_sep=False, time_unit='ms') ) sorted_by op_detail thread_sep time_unitviewsN)rsrr|rget_dataget_extra_infoprintr)r}rrrrrZstatistic_datar%r%r&summary>s$ #zProfiler.summaryr)rrG)r!r"r#r$rrr-rrr;r(rtrboolrurrrrrr8rrrRrZCPUTotalrr%r%r%r&rdTsN   K/0 */ Grdc Cszt|d }t|}Wdn1swYWn ty<}ztd|tdtWYd}~Sd}~wwi}d|vrzz*g|d<|dD]}|dkr]|dt j qL|dkrk|dt j qLWn tdd|d<Yd|vrzIt |dt r|dD]*\}}|d }|d } t|} t| |} | s| |d i|d |d<q| |d<qn|dd |ddg|d<Wn tdd|d<Yd|vr"zSet timer_only parameter error, use default parameter instead.r%)openrGloadr]rrIrdlowerappendr-r.r/rsdictitems importlib import_modulegetattrr) Z config_pathZ filehandleZ config_dicteZtranslated_config_dictr~keyvalue module_pathrrmethodr%r%r& get_profilerus                 rr)/rMrJrOenumrtypingrrrrrwarningsrrrGroZpaddle.fluid.corer r r r r rrrrrrZprofiler_statisticrrrZpaddle.profilertimerrrr(r-r;r<r=rLr`rbrcrdrr%r%r%r&sp  (   ! [ 8 7 %