import itertools from typing import Any import torch from torch.futures import Future from collections import defaultdict, namedtuple from operator import attrgetter from typing import List, Dict, Tuple, Optional 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 EventList(list): """A list of Events (for pretty printing)""" def __init__(self, *args, **kwargs): use_cuda = kwargs.pop('use_cuda', True) profile_memory = kwargs.pop('profile_memory', False) super(EventList, self).__init__(*args, **kwargs) self._cpu_children_populated = False self._use_cuda = use_cuda self._profile_memory = profile_memory def __str__(self): return self.table() def populate_cpu_children(self): """Populates 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. """ if self.cpu_children_populated: return # Some events can be async (i.e. start and end on different threads), # since it's generally undefined how to attribute children ranges to # async ranges, we do not use them when calculating nested ranges and stats sync_events = [evt for evt in self if not evt.is_async] events = sorted( sync_events, key=attrgetter("thread"), ) # Group by both thread and node_id, so that events that happen to have # the same thread_id but are from different nodes aren't incorrectly # grouped together. threads = itertools.groupby( events, key=lambda event: (event.thread, event.node_id) ) # For each thread we keep a stack of current nested parents. # We maintain the invariant that each interval is a subset of all other # intervals lower in the stack. # # First we sort the intervals by their start time. Then we iterate over them. # Every time we see a new interval we remove several parents from # the top until we restore the invariant. Then parent child relationship # if recorded if the stack is not empty. # Finally we add new interval to the list # # Algorithm has O(N * log(N)) complexity where N is number of # intervals for thread_id, thread_events in threads: thread_events_ = sorted( thread_events, key=lambda event: [event.cpu_interval.start, -event.cpu_interval.end], ) current_events: List[FunctionEvent] = [] cur_end = 0 for event in thread_events_: while len(current_events) > 0: parent = current_events[-1] if event.cpu_interval.start >= parent.cpu_interval.end or \ event.cpu_interval.end > parent.cpu_interval.end: # this can't be a parent current_events.pop() else: parent.append_cpu_child(event) assert ( event.cpu_parent is None ), "There is already a CPU parent event for {}".format( event.key ) event.set_cpu_parent(parent) break current_events.append(event) self._cpu_children_populated = True def set_backward_stacktraces(self): self.populate_cpu_children() def bw_parent(evt): if evt is None: return None elif evt.scope == 1: return evt else: return bw_parent(evt.cpu_parent) fwd_stacks = {} for evt in self: if bw_parent(evt) is None: t = (evt.sequence_nr, evt.thread) if t not in fwd_stacks: fwd_stacks[t] = evt.stack for evt in self: p = bw_parent(evt) if p is not None: assert p.fwd_thread is not None t = (p.sequence_nr, p.fwd_thread) if t in fwd_stacks: evt.stack = fwd_stacks[t] else: evt.stack = [] @property def self_cpu_time_total(self): return sum([event.self_cpu_time_total for event in self]) @property def cpu_children_populated(self): return self._cpu_children_populated def table(self, sort_by=None, row_limit=100, header=None, top_level_events_only=False): """Prints an EventList as a nicely formatted table. Arguments: 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``, ``cpu_time_total``, ``cuda_time_total``, ``cpu_memory_usage``, ``cuda_memory_usage``, ``self_cpu_memory_usage``, ``self_cuda_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 ops events are omitted for profiler result readability. Returns: A string containing the table. """ return build_table( self, sort_by=sort_by, row_limit=row_limit, header=header, use_cuda=self._use_cuda, profile_memory=self._profile_memory, top_level_events_only=top_level_events_only) def export_chrome_trace(self, path): """Exports an EventList as a Chrome tracing tools file. The checkpoint can be later loaded and inspected under ``chrome://tracing`` URL. Arguments: path (str): Path where the trace will be written. """ import os with open(path, 'w') as f: chrome_events = [] next_id = 0 # Use file IO over using json.dump since JSON dumping is very slow and # this technique is proven to give a 4x speedup. f.write("[") for evt in self: f.write( '{"name": "%s", ' '"ph": "X", ' '"ts": %s, ' '"dur": %s, ' '"tid": %s, ' '"pid": "CPU functions", ' '"args": {}}, ' % ( evt.name, evt.cpu_interval.start, evt.cpu_interval.elapsed_us(), evt.thread if not evt.is_remote else f'" node_id:{evt.node_id}, thread_id:{evt.thread} "', ) ) for k in evt.kernels: # 's' and 'f' draw Flow arrows from # the CPU launch to the GPU kernel f.write('{"name": "%s", ' '"ph": "s", ' '"ts": %s, ' '"tid": %s, ' '"pid": "CPU functions", ' '"id": %s, ' '"cat": "cpu_to_cuda", ' '"args": {}}, ' % (evt.name, evt.cpu_interval.start, evt.thread, next_id)) f.write('{"name": "%s", ' '"ph": "f", ' '"ts": %s, ' '"tid": %s, ' '"pid": "CUDA functions", ' '"id": %s, ' '"cat": "cpu_to_cuda", ' '"args": {}}, ' % (k.name, k.interval.start, k.device, next_id)) f.write('{"name": "%s", ' '"ph": "X", ' '"ts": %s, ' '"dur": %s, ' '"tid": %s, ' '"pid": "CUDA functions", ' '"args": {}}, ' % (k.name, k.interval.start, k.interval.elapsed_us(), k.device)) next_id += 1 # remove trailing whitespace and comma f.seek(f.tell() - 2, os.SEEK_SET) f.truncate() f.write("]") def key_averages(self, group_by_input_shapes=False, group_by_stack_n=0): """Averages all function events over their keys. Arguments: 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 Returns: An EventList containing FunctionEventAvg objects. """ self.populate_cpu_children() stats: Dict[Tuple[int, Tuple[int, int]], FunctionEventAvg] = defaultdict(FunctionEventAvg) def get_key(event, group_by_input_shapes, group_by_stack_n): key = [str(event.key), str(event.node_id)] if group_by_input_shapes: key.append(str(event.input_shapes)) if group_by_stack_n > 0: key += event.stack[:group_by_stack_n] return tuple(key) for evt in self: stats[get_key(evt, group_by_input_shapes, group_by_stack_n)].add(evt) avg_list = EventList(stats.values(), use_cuda=self._use_cuda, profile_memory=self._profile_memory) for evt in avg_list: evt.stack = evt.stack[:group_by_stack_n] if not group_by_input_shapes: evt.input_shapes = "" return avg_list def total_average(self): """Averages all events. Returns: A FunctionEventAvg object. """ total_stat = FunctionEventAvg() for evt in self: total_stat += evt total_stat.key = None total_stat.key = 'Total' return total_stat 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 Arguments: enabled (bool, optional): Setting this to False makes this context manager a no-op. Default: ``True``. use_cuda (bool, optional): Enables timing of CUDA events as well using the cudaEvent API. Adds approximately 4us of overhead to each tensor operation. Default: ``False`` 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. profile_memory (bool, optional): Whether to report memory usage, default: ``False`` with_stack (bool, optional): record source information (file and line number) for the ops .. 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, profile_memory=False, with_stack=False): self.enabled = enabled self.use_cuda = use_cuda self.function_events = None if not self.enabled: return self.entered = False self.record_shapes = record_shapes self.profile_memory = profile_memory self.with_stack = with_stack def __enter__(self): if not self.enabled: return if self.entered: raise RuntimeError("autograd profiler traces are not reentrant") self.entered = True profiler_kind = torch.autograd.ProfilerState.CUDA if self.use_cuda \ else torch.autograd.ProfilerState.CPU config = torch.autograd.ProfilerConfig( profiler_kind, self.record_shapes, self.profile_memory, self.with_stack) torch.autograd._enable_profiler(config) return self def __exit__(self, exc_type, exc_val, exc_tb): if not self.enabled: return records = torch.autograd._disable_profiler() self.function_events = EventList( parse_event_records(records), use_cuda=self.use_cuda, profile_memory=self.profile_memory) if self.with_stack: self.function_events.set_backward_stacktraces() 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 '' self.function_events.populate_cpu_children() return str(self.function_events) def _check_finish(self): if self.function_events is None: raise RuntimeError("can't export a trace that didn't finish running") self.function_events.populate_cpu_children() def table(self, sort_by=None, row_limit=100, 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, header=header, top_level_events_only=top_level_events_only ) table.__doc__ = EventList.table.__doc__ def export_chrome_trace(self, path): self._check_finish() assert self.function_events is not None return self.function_events.export_chrome_trace(path) export_chrome_trace.__doc__ = EventList.export_chrome_trace.__doc__ def key_averages(self, group_by_input_shape=False, group_by_stack_n=0): self._check_finish() assert self.function_events is not None 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 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 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. Arguments: 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): self.name: str = name # 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) 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. Arguments: 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. Arguments: 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() torch.autograd._enable_profiler( torch.autograd.ProfilerConfig( torch.autograd.ProfilerState.NVTX, self.record_shapes, False, False) ) return self def __exit__(self, exc_type, exc_val, exc_tb): if not self.enabled: return torch.cuda.synchronize() torch.autograd._disable_profiler() return False def load_nvprof(path): """Opens an nvprof trace file and parses autograd annotations. Arguments: path (str): path to nvprof trace """ return EventList(parse_nvprof_trace(path)) ################################################################################ # FunctionEvent def format_time(time_us): """Defines how to format time in FunctionEvent""" US_IN_SECOND = 1000.0 * 1000.0 US_IN_MS = 1000.0 if time_us >= US_IN_SECOND: return '{:.3f}s'.format(time_us / US_IN_SECOND) if time_us >= US_IN_MS: return '{:.3f}ms'.format(time_us / US_IN_MS) return '{:.3f}us'.format(time_us) def format_time_share(time_us, total_time_us): """Defines how to format time in FunctionEvent""" if total_time_us == 0: assert time_us == 0, "Expected time_us == 0 but got {}".format(time_us) return "NaN" return '{:.2f}%'.format(time_us * 100.0 / total_time_us) def format_memory(nbytes): """Returns a formatted memory size string""" KB = 1024 MB = 1024 * KB GB = 1024 * MB if (abs(nbytes) >= GB): return '{:.2f} Gb'.format(nbytes * 1.0 / GB) elif (abs(nbytes) >= MB): return '{:.2f} Mb'.format(nbytes * 1.0 / MB) elif (abs(nbytes) >= KB): return '{:.2f} Kb'.format(nbytes * 1.0 / KB) else: return str(nbytes) + ' b' def attr_formatter(name): return property(lambda self: format_time(getattr(self, name))) class FormattedTimesMixin(object): """Helpers for FunctionEvent and FunctionEventAvg. The subclass should define `*_time_total` and `count` attributes. """ cpu_time_str = attr_formatter('cpu_time') cuda_time_str = attr_formatter('cuda_time') cpu_time_total_str = attr_formatter('cpu_time_total') cuda_time_total_str = attr_formatter('cuda_time_total') self_cpu_time_total_str = attr_formatter('self_cpu_time_total') self_cuda_time_total_str = attr_formatter('self_cuda_time_total') @property def cpu_time(self): return 0.0 if self.count == 0 else 1.0 * self.cpu_time_total / self.count # type: ignore @property def cuda_time(self): return 0.0 if self.count == 0 else 1.0 * self.cuda_time_total / self.count # type: ignore class Interval(object): def __init__(self, start, end): self.start = start self.end = end def elapsed_us(self): return self.end - self.start Kernel = namedtuple('Kernel', ['name', 'device', 'interval']) class FunctionEvent(FormattedTimesMixin): """Profiling information about a single function.""" def __init__( self, id, node_id, name, thread, cpu_start, cpu_end, fwd_thread=None, input_shapes=None, stack=None, scope=0, cpu_memory_usage=0, cuda_memory_usage=0, is_async=False, is_remote=True, sequence_nr=-1): self.id: int = id self.node_id: int = node_id self.name: str = name self.cpu_interval: Interval = Interval(cpu_start, cpu_end) self.thread: int = thread self.fwd_thread: Optional[int] = fwd_thread self.kernels: List[Kernel] = [] self.count: int = 1 self.cpu_children: List[FunctionEvent] = [] self.cpu_parent: Optional[FunctionEvent] = None self.input_shapes: Tuple[int, ...] = input_shapes self.stack: List = stack self.scope: int = scope self.cpu_memory_usage: int = cpu_memory_usage self.cuda_memory_usage: int = cuda_memory_usage self.is_async: bool = is_async self.is_remote: bool = is_remote self.sequence_nr: int = sequence_nr def append_kernel(self, name, device, start, end): self.kernels.append(Kernel(name, device, Interval(start, end))) def append_cpu_child(self, child): """Append 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. """ assert(isinstance(child, FunctionEvent)) self.cpu_children.append(child) def set_cpu_parent(self, parent): """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. """ assert(isinstance(parent, FunctionEvent)) self.cpu_parent = parent # Note: async events don't have children, are not used when computing 'self' # metrics of other events, have only total cpu time @property def self_cpu_memory_usage(self): if self.is_async: return 0 return self.cpu_memory_usage - sum( [child.cpu_memory_usage for child in self.cpu_children] ) @property def self_cuda_memory_usage(self): if self.is_async: return 0 return self.cuda_memory_usage - sum( [child.cuda_memory_usage for child in self.cpu_children] ) @property def self_cpu_time_total(self): if self.is_async: return 0 return self.cpu_time_total - sum( [child.cpu_time_total for child in self.cpu_children] ) @property def cuda_time_total(self): return sum(kinfo.interval.elapsed_us() for kinfo in self.kernels) @property def self_cuda_time_total(self): return sum(kinfo.interval.elapsed_us() for kinfo in self.kernels) - \ sum([child.cuda_time_total for child in self.cpu_children]) @property def cpu_time_total(self): return self.cpu_interval.elapsed_us() @property def key(self): return self.name def __repr__(self): return ( ''.format( self.id, self.node_id, self.cpu_time_str, self.cpu_interval.start, self.cpu_interval.end, str([child.id for child in self.cpu_children]), self.cuda_time_str, self.name, self.thread, str(self.input_shapes), self.cpu_memory_usage, self.cuda_memory_usage, self.is_async, self.is_remote, self.sequence_nr, ) ) class FunctionEventAvg(FormattedTimesMixin): """Used to average stats over multiple FunctionEvent objects.""" def __init__(self): self.key: Optional[str] = None self.count: int = 0 self.node_id: int = 0 self.is_async: bool = False self.is_remote: bool = False self.cpu_time_total: int = 0 self.cuda_time_total: int = 0 self.self_cpu_time_total: int = 0 self.self_cuda_time_total: int = 0 self.input_shapes: Optional[List[List[int]]] = None self.stack: Optional[List] = None self.scope: Optional[int] = None self.cpu_memory_usage: int = 0 self.cuda_memory_usage: int = 0 self.self_cpu_memory_usage: int = 0 self.self_cuda_memory_usage: int = 0 self.cpu_children: Optional[List[FunctionEvent]] = None self.cpu_parent: Optional[FunctionEvent] = None def add(self, other): if self.key is None: # First function being recorded as part of FunctionEventAvg, propagate # fields. self.key = other.key self.node_id = other.node_id self.is_async = other.is_async self.is_remote = other.is_remote self.cpu_parent = other.cpu_parent self.cpu_children = other.cpu_children self.input_shapes = other.input_shapes self.stack = other.stack self.scope = other.scope assert isinstance(other, (FunctionEvent, FunctionEventAvg)) assert other.key == self.key self.cpu_time_total += other.cpu_time_total self.cuda_time_total += other.cuda_time_total self.self_cpu_time_total += other.self_cpu_time_total self.self_cuda_time_total += other.self_cuda_time_total self.cpu_memory_usage += other.cpu_memory_usage self.cuda_memory_usage += other.cuda_memory_usage self.self_cpu_memory_usage += other.self_cpu_memory_usage self.self_cuda_memory_usage += other.self_cuda_memory_usage self.count += other.count return self def __iadd__(self, other): return self.add(other) def __repr__(self): return ( ''.format( self.key, self.self_cpu_time_total_str, self.cpu_time_str, self.self_cuda_time_total_str, self.cuda_time_str, str(self.input_shapes), self.cpu_memory_usage, self.cuda_memory_usage, ) ) ################################################################################ # Utilities class StringTable(defaultdict): def __missing__(self, key): # manage cases like 't' (demangled to 'unsigned short') separately, # for now simply check the length to avoid unexpected results for # the short sequences self[key] = torch._C._demangle(key) if len(key) > 1 else key return self[key] def parse_event_records(thread_records): def get_record_key(record): """ Returns a tuple to be used by parse_event_records for correlating start and end records. """ return (record.handle(), record.node_id()) next_id = 0 start_record = None cuda_records = {} functions = [] record_stack = [] string_table = StringTable() # ignoring the following utility ops filtered_out_names = [ "profiler::_record_function_enter", "profiler::_record_function_exit", "aten::is_leaf", "aten::output_nr", "aten::_version", ] def filter_stack_entry(entry): filtered_entries = [ ("autograd/__init__", "_make_grads"), ("autograd/__init__", "backward"), ("torch/tensor", "backward"), ("_internal/common_utils", "prof_callable"), ("_internal/common_utils", "prof_func_call"), ("_internal/common_utils", "prof_meth_call"), ] return all([not (f[0] in entry and f[1] in entry) for f in filtered_entries]) # cuda start events and the overall profiler start event don't happen # at exactly the same time because we need to record an event on each device # and each record takes ~4us. So we adjust here by the difference # adding the difference in CPU time between the profiler start event # and the CPU time of the cuda start event for the device def adjusted_time(cuda_record, cuda_records_map): assert cuda_record.device() != -1 assert start_record is not None cuda_time_0 = cuda_records_map[(cuda_record.node_id(), cuda_record.device())] return cuda_time_0.cuda_elapsed_us(cuda_record) + start_record.cpu_elapsed_us(cuda_time_0) # '__start_profile' is not guaranteed to be first, so we must find it here for record in itertools.chain(*thread_records): name = record.name() if start_record is None and name == '__start_profile': start_record = record elif '__cuda_start_event' in name: # N.B.: Each CUDA device has its own __cuda_start_event. assert record.device() != -1 # key for cuda_records is (node_id, device) in case of multiple nodes # having the same device cuda_records[(record.node_id(), record.device())] = record assert start_record is not None and not start_record.is_remote() for thread_record_list in thread_records: # accumulated memory allocations per handle cpu_memory_allocs = {} cuda_memory_allocs = {} # ranges per handle range_starts = {} filtered_handles = set() prev_record = None for record in thread_record_list: record_key = get_record_key(record) if (record.name() in filtered_out_names or record_key in filtered_handles): filtered_handles.add(record_key) continue if record.kind() == 'push': # workaround to reduce double logging from operator # wrappers and redispatch if prev_record is not None: duplicate = ( prev_record.name() == record.name() and prev_record.kind() == record.kind() and prev_record.node_id() == record.node_id() ) if duplicate: filtered_handles.add(record_key) continue range_starts[record_key] = record cpu_memory_allocs[record_key] = 0 cuda_memory_allocs[record_key] = 0 elif record.kind() == 'pop': assert ( record_key in range_starts ), """Expected record with key {} to exist in range_starts. This means that the pop event did not have a corresponding push.""".format( record_key ) start = range_starts[record_key] cpu_memory_usage = cpu_memory_allocs[record_key] cuda_memory_usage = cuda_memory_allocs[record_key] is_async = start.thread_id() != record.thread_id() is_remote_event = record.is_remote() fe = FunctionEvent( id=record.handle(), node_id=record.node_id(), name=string_table[start.name()], thread=start.thread_id(), cpu_start=start_record.cpu_elapsed_us(start), cpu_end=start_record.cpu_elapsed_us(record), fwd_thread=start.fwd_thread_id(), input_shapes=start.shapes(), stack=[entry for entry in start.stack() if filter_stack_entry(entry)], scope=start.scope(), cpu_memory_usage=cpu_memory_usage, cuda_memory_usage=cuda_memory_usage, is_async=is_async, is_remote=is_remote_event, sequence_nr=start.sequence_nr(), ) # note: async events have only cpu total time if not is_async and start.has_cuda(): cuda_start = adjusted_time(start, cuda_records) cuda_end = adjusted_time(record, cuda_records) if (cuda_end - cuda_start) > 0: fe.append_kernel( start.name(), start.device(), cuda_start, cuda_end) functions.append(fe) del range_starts[record_key] del cpu_memory_allocs[record_key] del cuda_memory_allocs[record_key] elif record.kind() == 'memory_alloc': for handle in cpu_memory_allocs.keys(): cpu_memory_allocs[handle] += record.cpu_memory_usage() for handle in cuda_memory_allocs.keys(): cuda_memory_allocs[handle] += record.cuda_memory_usage() prev_record = record # Sort functions by start time then by end time ascending. # This ensures that--in the case of nested events which # have the same start time (which may happen due to the # granularity of the given clock tick)--we always show # the outermost nested call first. This adds stability # in how FunctionEvents appear functions.sort(key=lambda evt: [evt.cpu_interval.start, -evt.cpu_interval.end]) return functions ################################################################################ # CUDA checkpoints 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']], cpu_start=row['start_time'], cpu_end=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; 13 is for older cuda versions assert (row['cbid'] == 211) or (row['cbid'] == 13) evt = functions_map[row['marker_id']] evt.append_kernel(row['kernel_name'], 0, row['kernel_start'], row['kernel_end']) functions.sort(key=lambda evt: evt.cpu_interval.start) return functions ################################################################################ # Pretty printer def build_table( events, sort_by=None, header=None, row_limit=100, use_cuda=True, profile_memory=False, top_level_events_only=False): """Prints a summary of events (which can be a list of FunctionEvent or FunctionEventAvg).""" if len(events) == 0: return "" if sort_by is not None: events = EventList(sorted( events, key=lambda evt: getattr(evt, sort_by), reverse=True ), use_cuda=use_cuda, profile_memory=profile_memory) has_input_shapes = any( [(event.input_shapes is not None and len(event.input_shapes) > 0) for event in events]) name_column_width = max([len(evt.key) for evt in events]) + 4 DEFAULT_COLUMN_WIDTH = 12 shapes_column_width = max([len(str(evt.input_shapes)) for evt in events]) + 4 shapes_column_width = min(shapes_column_width, 45) src_column_width = None stacks = [] for evt in events: if evt.stack is not None and len(evt.stack) > 0: stacks.append(evt.stack) has_stack = len(stacks) > 0 if has_stack: src_column_width = max([max([len(entry) for entry in stack]) for stack in stacks]) + 4 src_column_width = min(src_column_width, 75) headers = [ 'Name', 'Self CPU %', 'Self CPU', 'CPU total %', 'CPU total', 'CPU time avg', ] if use_cuda: headers.extend([ 'Self CUDA', 'Self CUDA %', 'CUDA total', 'CUDA time avg', ]) if profile_memory: headers.extend([ 'CPU Mem', 'Self CPU Mem', ]) if torch.cuda.is_available(): headers.extend([ 'CUDA Mem', 'Self CUDA Mem', ]) headers.append( '# of Calls' ) # Only append Node ID if any event has a valid (>= 0) Node ID append_node_id = any([evt.node_id != -1 for evt in events]) if append_node_id: headers.append('Node ID') # Have to use a list because nonlocal is Py3 only... SPACING_SIZE = 2 row_format_lst = [""] header_sep_lst = [""] line_length_lst = [-SPACING_SIZE] MAX_STACK_ENTRY = 5 def add_column(padding, text_dir='>'): row_format_lst[0] += '{: ' + text_dir + str(padding) + '}' + (' ' * SPACING_SIZE) header_sep_lst[0] += '-' * padding + (' ' * SPACING_SIZE) line_length_lst[0] += padding + SPACING_SIZE add_column(name_column_width) for _ in headers[1:]: add_column(DEFAULT_COLUMN_WIDTH) if has_input_shapes: headers.append('Input Shapes') add_column(shapes_column_width) if has_stack: headers.append('Source Location') add_column(src_column_width, text_dir='<') row_format = row_format_lst[0] header_sep = header_sep_lst[0] line_length = line_length_lst[0] add_column = None # type: ignore # Have to use a list because nonlocal is Py3 only... result = [] def append(s): result.append(s) result.append('\n') # Yes, newline after the end as well self_cpu_time_total = sum([event.self_cpu_time_total for event in events]) cuda_time_total = sum([evt.self_cuda_time_total for evt in events]) # Actual printing if header is not None: append('=' * line_length) append(header) if top_level_events_only: append('=' * line_length) append('This report only display top-level ops statistics') append(header_sep) append(row_format.format(*headers)) append(header_sep) event_limit = 0 for evt in events: if event_limit == row_limit: break if top_level_events_only and evt.cpu_parent is not None: continue else: event_limit += 1 row_values = [ evt.key, # Name # Self CPU total, 0 for async events. % format_time_share(evt.self_cpu_time_total, self_cpu_time_total), evt.self_cpu_time_total_str, # Self CPU total # CPU total %, 0 for async events. format_time_share(evt.cpu_time_total, self_cpu_time_total) if not evt.is_async else 0, evt.cpu_time_total_str, # CPU total evt.cpu_time_str, # CPU time avg ] if use_cuda: row_values.extend([ evt.self_cuda_time_total_str, # CUDA time total % format_time_share(evt.self_cuda_time_total, cuda_time_total), evt.cuda_time_total_str, evt.cuda_time_str, # Cuda time avg ]) if profile_memory: row_values.extend([ # CPU Mem Total format_memory(evt.cpu_memory_usage), # Self CPU Mem Total format_memory(evt.self_cpu_memory_usage), ]) if torch.cuda.is_available(): row_values.extend([ # CUDA Mem Total format_memory(evt.cuda_memory_usage), # Self CUDA Mem Total format_memory(evt.self_cuda_memory_usage), ]) row_values.append( evt.count, # Number of calls ) if append_node_id: row_values.append(evt.node_id) if has_input_shapes: row_values.append(str(evt.input_shapes)[:shapes_column_width]) if has_stack: src_field = "" if len(evt.stack) > 0: src_field = evt.stack[0][:src_column_width] row_values.append(src_field) append(row_format.format(*row_values)) if has_stack: empty_headers = [""] * (len(headers) - 1) for entry in evt.stack[1:MAX_STACK_ENTRY]: append(row_format.format(*(empty_headers + [entry[:src_column_width]]))) empty_headers.append("") append(row_format.format(*empty_headers)) append(header_sep) append("Self CPU time total: {}".format(format_time(self_cpu_time_total))) if use_cuda: append("CUDA time total: {}".format(format_time(cuda_time_total))) return ''.join(result)