from torch.autograd.profiler_util import ( EventList, FunctionEvent, MemRecordsAcc, MEMORY_EVENT_NAME, _filter_name, _filter_stack_entry, _rewrite_name ) from torch.autograd import ( DeviceType, ProfilerActivity, ProfilerConfig, ProfilerState, kineto_available, _ProfilerResult, _disable_profiler, _enable_profiler, _prepare_profiler, _supported_activities, _kineto_step, ) from torch._C._autograd import _ExperimentalConfig import torch import torch.cuda from torch.futures import Future from typing import Any, Dict, List, Optional from warnings import warn try: # Available in Python >= 3.2 from contextlib import ContextDecorator except ImportError: import functools class ContextDecorator(object): # type: ignore[no-redef] def __enter__(self): raise NotImplementedError def __exit__(self, exc_type, exc_val, exc_tb): raise NotImplementedError def __call__(self, func): @functools.wraps(func) def wrapped(*args, **kwargs): with self: return func(*args, **kwargs) return wrapped class profile(object): """Context 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 ----------------------------------- --------------- --------------- --------------- """ def __init__( self, enabled=True, *, use_cuda=False, record_shapes=False, with_flops=False, profile_memory=False, with_stack=False, with_modules=False, use_kineto=False, use_cpu=True, experimental_config=None): self.enabled: bool = enabled if not self.enabled: return self.use_cuda = use_cuda self.function_events: Optional[EventList] = None self.entered = False self.record_shapes = record_shapes self.with_flops = with_flops self.record_shapes |= self.with_flops self.profile_memory = profile_memory self.with_stack = with_stack self.with_modules = with_modules self.use_cpu = use_cpu if experimental_config is None: experimental_config = _ExperimentalConfig() self.experimental_config = experimental_config self.kineto_results: Optional[_ProfilerResult] = None if not self.use_cpu: assert use_kineto, \ "Device-only events supported only with Kineto (use_kineto=True)" if self.use_cuda and not torch.cuda.is_available(): warn("CUDA is not available, disabling CUDA profiling") self.use_cuda = False self.kineto_activities = set() if self.use_cpu: self.kineto_activities.add(ProfilerActivity.CPU) self.profiler_kind = ProfilerState.KINETO if self.use_cuda: if (not use_kineto or ProfilerActivity.CUDA not in _supported_activities()): assert self.use_cpu, "Legacy CUDA profiling requires use_cpu=True" self.profiler_kind = ProfilerState.KINETO_GPU_FALLBACK else: self.kineto_activities.add(ProfilerActivity.CUDA) assert len(self.kineto_activities) > 0, \ "No activities specified for the profiler" def config(self): return ProfilerConfig( self.profiler_kind, self.record_shapes, self.profile_memory, self.with_stack, self.with_flops, self.with_modules, self.experimental_config) def __enter__(self): if not self.enabled: return if self.entered: raise RuntimeError("Profiler context manager is not reentrant") self._prepare_trace() self._start_trace() return self def _prepare_trace(self): self.entered = True _prepare_profiler(self.config(), self.kineto_activities) def _start_trace(self): self.entered = True _enable_profiler(self.config(), self.kineto_activities) def __exit__(self, exc_type, exc_val, exc_tb): if not self.enabled: return if self.use_cuda: torch.cuda.synchronize() self.kineto_results = _disable_profiler() parsed_results = self._parse_kineto_results(self.kineto_results) self.function_events = EventList( parsed_results, use_cuda=self.use_cuda, profile_memory=self.profile_memory, with_flops=self.with_flops) self.function_events._build_tree() return False def __repr__(self): if self.function_events is None: return '' return repr(self.function_events) def __str__(self): if self.function_events is None: return '' return str(self.function_events) def _check_finish(self): if self.function_events is None: raise RuntimeError("Profiler didn't finish running") def table(self, sort_by=None, row_limit=100, max_src_column_width=75, header=None, top_level_events_only=False): self._check_finish() assert self.function_events is not None return self.function_events.table( sort_by=sort_by, row_limit=row_limit, max_src_column_width=max_src_column_width, header=header, top_level_events_only=top_level_events_only ) table.__doc__ = EventList.table.__doc__ def export_chrome_trace(self, path): self._check_finish() if kineto_available(): self.kineto_results.save(path) # type: ignore[union-attr] else: return self.function_events.export_chrome_trace(path) # type: ignore[union-attr] export_chrome_trace.__doc__ = EventList.export_chrome_trace.__doc__ def export_stacks(self, path: str, metric: str = "self_cpu_time_total"): self._check_finish() assert self.function_events is not None, "Expected profiling results" assert self.with_stack, "export_stacks() requires with_stack=True" return self.function_events.export_stacks(path, metric) def key_averages(self, group_by_input_shape=False, group_by_stack_n=0): self._check_finish() assert self.function_events is not None, "Expected profiling results" return self.function_events.key_averages(group_by_input_shape, group_by_stack_n) key_averages.__doc__ = EventList.key_averages.__doc__ def total_average(self): self._check_finish() assert self.function_events is not None, "Expected profiling results" return self.function_events.total_average() total_average.__doc__ = EventList.total_average.__doc__ @property def self_cpu_time_total(self): """ Returns total time spent on CPU obtained as a sum of all self times across all the events. """ self._check_finish() assert self.function_events is not None return self.function_events.self_cpu_time_total def _parse_kineto_results(self, result): # result.events() has most of the events - PyTorch op-level and device-level events trace_start_us = result.trace_start_us() mem_records = [[evt, False] for evt in result.events() if evt.name() == MEMORY_EVENT_NAME] mem_records_acc = MemRecordsAcc(mem_records) def _cpu_memory_usage(mem_record): return mem_record.nbytes() if \ mem_record.device_type() in [DeviceType.CPU, DeviceType.MKLDNN, DeviceType.IDEEP] \ else 0 def _cuda_memory_usage(mem_record): return mem_record.nbytes() if \ mem_record.device_type() in [DeviceType.CUDA, DeviceType.HIP] \ else 0 # Create and return FunctionEvent list function_events = [] cuda_corr_map: Dict[int, List[FunctionEvent]] = {} max_evt_id = 0 for kineto_event in result.events(): if _filter_name(kineto_event.name()): continue rel_start_us = kineto_event.start_us() - trace_start_us rel_end_us = rel_start_us + kineto_event.duration_us() abs_end_us = kineto_event.start_us() + kineto_event.duration_us() cpu_memory_usage = 0 cuda_memory_usage = 0 if kineto_event.device_type() == DeviceType.CPU: # find the corresponding memory allocation events for mem_record in mem_records_acc.in_interval(kineto_event.start_us(), abs_end_us): cpu_memory_usage += _cpu_memory_usage(mem_record[0]) cuda_memory_usage += _cuda_memory_usage(mem_record[0]) mem_record[1] = True is_async = kineto_event.is_async() or ( kineto_event.start_thread_id() != kineto_event.end_thread_id() ) fe = FunctionEvent( id=kineto_event.correlation_id(), name=_rewrite_name(name=kineto_event.name(), with_wildcard=True), trace_name=_rewrite_name(name=kineto_event.name(), with_wildcard=False), thread=kineto_event.start_thread_id(), start_us=rel_start_us, end_us=rel_end_us, fwd_thread=kineto_event.fwd_thread_id(), input_shapes=kineto_event.shapes(), stack=[entry for entry in kineto_event.stack() if _filter_stack_entry(entry)], scope=kineto_event.scope(), cpu_memory_usage=cpu_memory_usage, cuda_memory_usage=cuda_memory_usage, is_async=is_async, sequence_nr=kineto_event.sequence_nr(), device_type=kineto_event.device_type(), device_index=kineto_event.device_index(), flops=kineto_event.flops(), ) max_evt_id = fe.id if fe.id > max_evt_id else max_evt_id if fe.device_type == DeviceType.CPU and not fe.is_async: # Check if we have CUDA time as a fallback cuda_time = kineto_event.cuda_elapsed_us() if cuda_time > 0: fe.append_kernel( fe.name, fe.device_index, cuda_time) fe.is_legacy = True function_events.append(fe) corr_id = kineto_event.linked_correlation_id() if corr_id > 0: if corr_id not in cuda_corr_map: cuda_corr_map[corr_id] = [] cuda_corr_map[corr_id].append(fe) # associate CUDA kernels and CUDA runtime (CPU) with CPU events for fe in function_events: if (fe.device_type == DeviceType.CPU and not fe.is_async and fe.id in cuda_corr_map): for f_evt in cuda_corr_map[fe.id]: if f_evt.device_type == DeviceType.CUDA: fe.append_kernel( f_evt.name, f_evt.device_index, f_evt.time_range.end - f_evt.time_range.start) elif f_evt.device_type == DeviceType.CPU: # make sure that 'thread' of a CPU Kineto (e.g. CUDA Runtime) event is associated # with the 'thread' of the corresponding linked PyTorch event to properly track # parents and children f_evt.thread = fe.thread # output top-level memory events for mem_record in mem_records: if not mem_record[1]: rel_start_us = mem_record[0].start_us() - trace_start_us max_evt_id += 1 fe = FunctionEvent( id=max_evt_id, name=MEMORY_EVENT_NAME, trace_name=None, # not outputting in the trace thread=mem_record[0].start_thread_id(), start_us=rel_start_us, end_us=rel_start_us, # no duration fwd_thread=mem_record[0].start_thread_id(), input_shapes=[], stack=[], scope=0, # RecordScope::FUNCTION cpu_memory_usage=_cpu_memory_usage(mem_record[0]), cuda_memory_usage=_cuda_memory_usage(mem_record[0]), is_async=False, sequence_nr=-1, device_type=DeviceType.CPU, device_index=0, ) function_events.append(fe) function_events.sort(key=lambda evt: [evt.time_range.start, -evt.time_range.end]) return function_events class record_function(ContextDecorator): """Context 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 """ def __init__(self, name: str, args: Optional[str] = None): self.name: str = name self.args: Optional[str] = args # Whether or not we should run record function's end callbacks when exiting. self.run_callbacks_on_exit: bool = True # Stores underlying RecordFunction as a tensor. TODO: move to custom # class (https://github.com/pytorch/pytorch/issues/35026). self.handle: torch.Tensor = torch.zeros(1) def __enter__(self): self.handle = torch.ops.profiler._record_function_enter(self.name, self.args) return self def __exit__(self, exc_type: Any, exc_value: Any, traceback: Any): if self.run_callbacks_on_exit: torch.ops.profiler._record_function_exit(self.handle) def _call_end_callbacks_on_future(self, fut: Future[Any]) -> Future[Any]: """ _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. """ # Throw if we have already attached a callback onto the future. if not self.run_callbacks_on_exit: raise RuntimeError("_call_end_callbacks_on_future can only be called once.") # We are scheduling to run this RecordFunction's end callbacks when the # passed in future completes, so don't run end callbacks on exit. self.run_callbacks_on_exit = False profiled_future = torch.ops.profiler._call_end_callbacks_on_jit_fut(self.handle, fut) return profiled_future class emit_nvtx(object): """Context 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. """ def __init__(self, enabled=True, record_shapes=False): self.enabled = enabled self.entered = False self.record_shapes = record_shapes def __enter__(self): if not self.enabled: return if self.entered: raise RuntimeError("NVTX annotation context manager is not reentrant") self.entered = True torch.cuda.synchronize() _enable_profiler( ProfilerConfig( ProfilerState.NVTX, self.record_shapes, False, False, False, False, _ExperimentalConfig()), set() ) return self def __exit__(self, exc_type, exc_val, exc_tb): if not self.enabled: return torch.cuda.synchronize() _disable_profiler() return False def load_nvprof(path): """Opens an nvprof trace file and parses autograd annotations. Args: path (str): path to nvprof trace """ return EventList(parse_nvprof_trace(path)) class EnforceUnique(object): """Raises an error if a key is seen more than once.""" def __init__(self): self.seen = set() def see(self, *key): if key in self.seen: raise RuntimeError('duplicate key: ' + str(key)) self.seen.add(key) def parse_nvprof_trace(path): import sqlite3 conn = sqlite3.connect(path) conn.row_factory = sqlite3.Row # Parse strings table strings = {} for r in conn.execute("SELECT _id_ as id, value FROM StringTable"): strings[r["id"]] = torch._C._demangle(r["value"]) # First, find all functions and create FunctionEvents for them marker_query = """ 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 """ functions = [] functions_map = {} unique = EnforceUnique() for row in conn.execute(marker_query): unique.see(row['marker_id']) evt = FunctionEvent(id=row['marker_id'], node_id=0, # missing a node_id when calling FunctionEvent. This is just to ensure # that pytorch doesn't crash when creating a FunctionEvent() object name=strings[row['name']], start_us=row['start_time'], end_us=row['end_time'], thread=0) # TODO: find in sqlite database functions.append(evt) functions_map[evt.id] = evt # Now, correlate all kernels with FunctionEvents kernel_query = """ 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 """ unique = EnforceUnique() for row in conn.execute(kernel_query): unique.see(row['marker_id'], row['runtime_id']) # 211 is cudaKernelLaunch for cuda >= 9.2 assert (row['cbid'] == 211) evt = functions_map[row['marker_id']] evt.append_kernel(row['kernel_name'], 0, row['kernel_end'] - row['kernel_start']) functions.sort(key=lambda evt: evt.time_range.start) return functions def kineto_step(): """ Notify kineto so it is aware of iteration boundaries for asynchronous trace requests. """ _kineto_step()