Source code for executorch.devtools.inspector._inspector
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
# pyre-unsafe
import dataclasses
import logging
import sys
import warnings
from collections import defaultdict, OrderedDict
from dataclasses import dataclass
from functools import cached_property
from typing import (
Any,
Callable,
Dict,
IO,
List,
Mapping,
Optional,
Sequence,
Tuple,
TypeAlias,
TypedDict,
Union,
)
import executorch.devtools.etdump.schema_flatcc as flatcc
import numpy as np
import pandas as pd
from executorch.devtools.debug_format.et_schema import OperatorGraph, OperatorNode
from executorch.devtools.etdump.schema_flatcc import (
DebugEvent,
ETDumpFlatCC,
ProfileEvent,
)
from executorch.devtools.etrecord import ETRecord, parse_etrecord
from executorch.devtools.inspector._inspector_utils import (
calculate_time_scale_factor,
compare_intermediate_outputs,
create_debug_handle_to_op_node_mapping,
DebugHandle,
display_or_print_df,
EDGE_DIALECT_GRAPH_KEY,
EXCLUDED_COLUMNS_WHEN_PRINTING,
EXCLUDED_EVENTS_FOR_INTERMEDIATE_OUTPUT,
EXCLUDED_EVENTS_WHEN_PRINTING,
find_op_names,
find_populated_event,
FORWARD,
gen_etdump_object,
gen_graphs_from_etrecord,
get_aot_debug_handle_to_op_name_mapping,
inflate_runtime_output,
is_debug_output,
is_inference_output_equal,
map_runtime_aot_intermediate_outputs,
merge_runtime_overlapping_debug_handles,
ProgramOutput,
propagate_back_debug_handle,
RESERVED_FRAMEWORK_EVENT_NAMES,
TimeScale,
verify_debug_data_equivalence,
)
from executorch.devtools.inspector._intermediate_output_capturer import (
IntermediateOutputCapturer,
)
from executorch.devtools.inspector.numerical_comparator import (
L1Comparator,
MSEComparator,
SNRComparator,
)
from executorch.exir import ExportedProgram
log: logging.Logger = logging.getLogger(__name__)
# Signature of an InstructionEvent
@dataclass(frozen=True, order=True)
class InstructionEventSignature:
instruction_id: int
chain_index: int
delegate_id: Optional[int] = None
delegate_id_str: Optional[str] = None
# Aggregated Runtime Events for a single instruction
@dataclass
class InstructionEvent:
signature: InstructionEventSignature
profile_events: Optional[List[ProfileEvent]] = None
debug_events: Optional[List[DebugEvent]] = None
@staticmethod
def gen_from_events(run_events: List[flatcc.Event]) -> List["InstructionEvent"]:
"""
Given a list of events from a run in ETDump, collate the ProfileEvent
and DebugEvents by instruction id and return a list of InstructionEvents
constructed from collated events (ignoring run_output events)
"""
instruction_events: Dict[InstructionEventSignature, InstructionEvent] = (
OrderedDict()
)
for event in run_events:
# Find the event that was logged
populated_event: Union[DebugEvent, ProfileEvent] = find_populated_event(
event
)
# Get existing InstructionEvent or insert a new one
signature = InstructionEventSignature(
instruction_id=populated_event.instruction_id,
chain_index=populated_event.chain_index,
delegate_id=populated_event.delegate_debug_id_int,
delegate_id_str=populated_event.delegate_debug_id_str,
)
instruction_event = instruction_events.setdefault(
signature, InstructionEvent(signature=signature)
)
# Update InstructionEvent based on event type
if isinstance(populated_event, ProfileEvent):
if instruction_event.profile_events is None:
instruction_event.profile_events = []
instruction_event.profile_events.append(populated_event)
elif isinstance(populated_event, DebugEvent):
# Ignore run_output events
if not is_debug_output(populated_event.debug_entry):
if instruction_event.debug_events is None:
instruction_event.debug_events = []
instruction_event.debug_events.append(populated_event)
return list(instruction_events.values())
# Signature of a ProfileEvent
@dataclass(frozen=True, order=True)
class ProfileEventSignature:
name: str
instruction_id: Optional[int]
delegate_id: Optional[int] = None
delegate_id_str: Optional[str] = None
@staticmethod
def _gen_from_event(event: ProfileEvent) -> "ProfileEventSignature":
"""
Given a ProfileEvent, extract the fields into a signature
ProfileEvents from ETDump default to "" and -1 when the field is not populated
The Signature will convert these back to the intended None value
"""
return ProfileEventSignature(
event.name or "",
event.instruction_id if event.instruction_id != -1 else None,
event.delegate_debug_id_int if event.delegate_debug_id_int != -1 else None,
event.delegate_debug_id_str if event.delegate_debug_id_str != "" else None,
)
# Signature of a DebugEvent
@dataclass(frozen=True, order=True)
class DebugEventSignature:
name: str = ""
instruction_id: Optional[int] = -1
delegate_id: Optional[int] = None
delegate_id_str: Optional[str] = None
@staticmethod
def _gen_from_event(event: DebugEvent) -> "DebugEventSignature":
"""
Given a DebugEvent, extract the fields into a signature
DebugEvents from ETDump default to "" and -1 when the field is not populated
The Signature will convert these back to the intended None value
"""
return DebugEventSignature(
event.name or "",
event.instruction_id if event.instruction_id != -1 else None,
event.delegate_debug_id_int if event.delegate_debug_id_int != -1 else None,
event.delegate_debug_id_str if event.delegate_debug_id_str != "" else None,
)
# Signature of an Event inside of a Run
@dataclass(frozen=True, order=True)
class EventSignature:
"""
Note that (profile_event_signature, debug_event_signature) are sufficient
signature identifiers.
instruction_id is extracted from the signatures (equivalent in both) and
surfaced for convenience
"""
instruction_id: int
profile_event_signature: Optional[ProfileEventSignature] = None
debug_event_signature: Optional[DebugEventSignature] = None
@staticmethod
def gen_from_instruction_event(
instruction_event: InstructionEvent,
) -> List[Tuple["EventSignature", InstructionEvent]]:
"""
Construct EventSignatures from the given InstructionEvent
and return tuples of (1) EventSignature and (2) related subset
InstructionEvent
"""
# Generate the DebugEventSignature
debug_events = instruction_event.debug_events
debug_signature = (
DebugEventSignature._gen_from_event(debug_events[0])
if debug_events is not None and len(debug_events) > 0
else None
)
# If no ProfileEvents, return a singleton EventSignature
if (profile_events := instruction_event.profile_events) is None:
return [
(
EventSignature(
instruction_id=instruction_event.signature.instruction_id,
debug_event_signature=debug_signature,
),
instruction_event,
)
]
# Generate the ProfileEventSignature
return [
(
EventSignature(
instruction_id=instruction_event.signature.instruction_id,
profile_event_signature=ProfileEventSignature._gen_from_event(
profile_event
),
debug_event_signature=debug_signature,
),
dataclasses.replace(instruction_event, profile_events=[profile_event]),
)
for profile_event in profile_events
]
# Signature of a Run
@dataclass(frozen=True, order=True)
class RunSignature:
"""
Args:
name: Name of the run
events: List of EventSignatures that correspond to the run
bundled_input_index: Index of the bundled input used to generate the debug output
"""
name: str
events: Optional[Tuple[EventSignature]] = None
bundled_input_index: Optional[int] = None
# Typing for mapping Event.delegate_debug_identifiers to debug_handle(s)
DelegateIdentifierDebugHandleMap: TypeAlias = Union[
Mapping[int, DebugHandle], Mapping[str, DebugHandle]
]
# Typing for Dict containig delegate metadata
DelegateMetadata = TypedDict(
"DelegateMetadata",
{"name": str, "delegate_map": DelegateIdentifierDebugHandleMap},
)
@dataclass
class PerfData:
def __init__(self, raw: List[float]):
self.raw: List[float] = raw
@property
def p10(self) -> float:
return np.percentile(self.raw, 10)
@property
def p50(self) -> float:
return np.percentile(self.raw, 50)
@property
def p90(self) -> float:
return np.percentile(self.raw, 90)
@property
def avg(self) -> float:
return np.mean(self.raw)
@property
def min(self) -> float:
return min(self.raw)
@property
def max(self) -> float:
return max(self.raw)
[docs]@dataclass
class Event:
"""
An Event corresponds to an operator instance with perf data retrieved from the runtime and other metadata from `ETRecord`.
Args:
name: Name of the profiling `Event`, empty if no profiling event.
perf_data: Performance data associated with the event retrived from the runtime (available attributes: p10, p50, p90, avg, min and max).
op_type: List of op types corresponding to the event.
delegate_debug_identifier: Supplemental identifier used in combination with instruction id.
debug_handles: Debug handles in the model graph to which this event is correlated.
num_outputs: Indicates the number of outputs generated by the node.
Right now only used for call_delegate nodes that output more than one tensor.
stack_trace: A dictionary mapping the name of each associated op to its stack trace.
module_hierarchy: A dictionary mapping the name of each associated op to its module hierarchy.
is_delegated_op: Whether or not the event was delegated.
delegate_backend_name: Name of the backend this event was delegated to.
_delegate_debug_metadatas: A list of raw delegate debug metadata in string, one for each profile event.
Available parsed (if parser provided) as Event.delegate_debug_metadatas
Available as Event.raw_delegate_debug_metadatas
debug_data: A list containing intermediate data collected.
_instruction_id: Instruction Identifier for Symbolication
_delegate_metadata_parser: Optional Parser for _delegate_debug_metadatas
"""
name: str
perf_data: Optional[PerfData] = None
op_types: List[str] = dataclasses.field(default_factory=list)
delegate_debug_identifier: Optional[Union[int, str]] = None
debug_handles: Optional[Union[int, Sequence[int]]] = None
num_outputs: int = 1
stack_traces: Dict[str, str] = dataclasses.field(default_factory=dict)
module_hierarchy: Dict[str, Dict] = dataclasses.field(default_factory=dict)
is_delegated_op: Optional[bool] = None
delegate_backend_name: Optional[str] = None
_delegate_debug_metadatas: List[str] = dataclasses.field(default_factory=list)
debug_data: ProgramOutput = dataclasses.field(default_factory=list)
_instruction_id: Optional[int] = None
_delegate_metadata_parser: Optional[Callable[[List[str]], Dict[str, Any]]] = None
_delegate_time_scale_converter: Optional[
Callable[[Union[int, str], Union[int, float]], Union[int, float]]
] = None
_start_time: Optional[List[Union[int, float]]] = None
@cached_property
def delegate_debug_metadatas(self) -> Union[List[str], Dict[str, Any]]:
"""
Returns the parsed _delegate_debug_metadatas if a parser is available
Otherwise returns the raw _delegate_debug_metadatas
"""
if not self.is_delegated_op or self._delegate_metadata_parser is None:
return self._delegate_debug_metadatas
return self._delegate_metadata_parser(self._delegate_debug_metadatas)
@property
def raw_delegate_debug_metadatas(self) -> List[str]:
"""
Return the raw unparsed _delegate_debug_metadatas
"""
return self._delegate_debug_metadatas
@property
def start_time(self) -> Optional[List[Union[int, float]]]:
"""
Returns the start time of the event.
"""
return self._start_time
def to_dataframe(self, _units="") -> pd.DataFrame:
"""
Convert the Event into a pandas DataFrame
Args:
None
Returns:
A pandas DataFrame with the Event data
"""
event_dict = self.asdict(_units=_units)
return pd.DataFrame(event_dict)
# Override the default implementation of dataclass.asdict to handle null perf data
def asdict(self, _units="") -> dict:
"""
Convert the Event into a dict
Args:
None
Returns:
A dict with the Event data
"""
def truncated_list(long_list: List[str]) -> str:
return f"['{long_list[0]}', '{long_list[1]}' ... '{long_list[-1]}'] ({len(long_list)} total)"
return {
"event_name": self.name,
"raw": [self.perf_data.raw if self.perf_data else None],
"p10" + _units: self.perf_data.p10 if self.perf_data else None,
"p50" + _units: self.perf_data.p50 if self.perf_data else None,
"p90" + _units: self.perf_data.p90 if self.perf_data else None,
"avg" + _units: self.perf_data.avg if self.perf_data else None,
"min" + _units: self.perf_data.min if self.perf_data else None,
"max" + _units: self.perf_data.max if self.perf_data else None,
"op_types": [
(
self.op_types
if len(self.op_types) < 5
else truncated_list(self.op_types)
)
],
"delegate_debug_identifier": self.delegate_debug_identifier,
"stack_traces": [self.stack_traces],
"module_hierarchy": [self.module_hierarchy],
"is_delegated_op": self.is_delegated_op,
"delegate_backend_name": self.delegate_backend_name,
"debug_data": [self.debug_data],
"start_time": [self._start_time],
}
@staticmethod
def _gen_from_inference_events(
signature: EventSignature,
events: List[InstructionEvent],
scale_factor: float = 1.0,
output_buffer: Optional[bytes] = None,
delegate_metadata_parser: Optional[
Callable[[List[str]], Dict[str, Any]]
] = None,
delegate_time_scale_converter: Optional[
Callable[[Union[int, str], Union[int, float]], Union[int, float]]
] = None,
) -> "Event":
"""
Given an EventSignature and a list of Events with that signature,
return an Event object matching the EventSignature, with perf_data
populated from the list of ProfileEvents and debug_data populated from
the list of DebugEvents.
An optional inverse scale factor can be provided to adjust the event timestamps
An optional buffer can be provided to inflate etdump references
An optional delegate_metadata_parser can be provided to parse the delegate metadata
"""
profile_event_signature = signature.profile_event_signature
debug_event_signature = signature.debug_event_signature
# Event is gradually populated in this function
ret_event: Event = Event(
name="",
_instruction_id=signature.instruction_id,
_delegate_metadata_parser=delegate_metadata_parser,
_delegate_time_scale_converter=delegate_time_scale_converter,
)
# Populate fields from profile events
Event._populate_profiling_related_fields(
ret_event, profile_event_signature, events, scale_factor
)
# Populate fields from debug events
Event._populate_debugging_related_fields(
ret_event, debug_event_signature, events, output_buffer
)
return ret_event
@staticmethod
def _calculate_elapsed_time(start_time, end_time):
# We're assuming if there's a wraparound in the time values, then
# the time representation of that platform only contains 32 bits.
# This should be fine for now, but ideally we should source the max
# time value from the platform using etdump.
max_uint32 = 2**32 - 1
if start_time > end_time:
if (start_time > max_uint32) or (end_time > max_uint32):
raise ValueError(
f"Expected start_time ({start_time}) and end_time ({end_time}) to be less than {max_uint32} for cases where there is wrap-around of time values."
)
# Handle wraparound
elapsed_time = (max_uint32 - start_time) + end_time
else:
# Normal case
elapsed_time = end_time - start_time
return elapsed_time
@staticmethod
def _populate_event_signature_fields(
ret_event: "Event",
event_signature: Optional[Union[ProfileEventSignature, DebugEventSignature]],
) -> None:
"""
Given a partially constructed Event, populate the fields related to
the profile event signature or debug event signature
Fields Updated:
name
delegate_debug_identifier
is_delegated_op
"""
# TODO: T201347372 Push the None check to ealier in the stack.
if event_signature is not None:
if event_signature.delegate_id is not None: # 0 is a valid value
delegate_debug_identifier = event_signature.delegate_id
else:
delegate_debug_identifier = event_signature.delegate_id_str or None
# Use the delegate identifier as the event name if delegated
is_delegated_op = delegate_debug_identifier is not None
name = (
event_signature.name
if not is_delegated_op
else str(delegate_debug_identifier)
)
# Update fields
# This is for older version of etdump that doesn't have the name field for debug events, we don't update the name field
if name:
ret_event.name = name
ret_event.delegate_debug_identifier = delegate_debug_identifier
ret_event.is_delegated_op = is_delegated_op
@staticmethod
def _populate_profiling_related_fields(
ret_event: "Event",
profile_event_signature: Optional[ProfileEventSignature],
events: List[InstructionEvent],
scale_factor: float,
) -> None:
"""
Given a partially constructed Event, populate the fields related to
the profile events
Fields Updated:
name
delegate_debug_identifier
is_delegated_op
perf_data
delegate_debug_metadatas
"""
# Fill out fields from profile event signature
Event._populate_event_signature_fields(ret_event, profile_event_signature)
# Fill out fields from profile event
data = []
stime = []
delegate_debug_metadatas = []
for event in events:
if (profile_events := event.profile_events) is not None:
if len(profile_events) != 1:
raise ValueError(
f"Expected exactly one profile event per InstructionEvent when generating Inspector Event, but got {len(profile_events)}"
)
profile_event = profile_events[0]
# Scale factor should only be applied to non-delegated ops
if (
ret_event.is_delegated_op
and (convert_time_scale := ret_event._delegate_time_scale_converter)
is not None
):
scaled_time = Event._calculate_elapsed_time(
convert_time_scale(ret_event.name, profile_event.start_time),
convert_time_scale(ret_event.name, profile_event.end_time),
)
# If it's not a delegated op then we can just use the raw time values
# and then scale them according to the scale factor that was passed in.
elif not ret_event.is_delegated_op:
scaled_time = (
float(
Event._calculate_elapsed_time(
profile_event.start_time, profile_event.end_time
)
)
/ scale_factor
)
# If there was no scale factor passed in just take a difference of the
# end and start times.
else:
scaled_time = float(
Event._calculate_elapsed_time(
profile_event.start_time, profile_event.end_time
)
)
data.append(scaled_time)
stime.append(profile_event.start_time)
delegate_debug_metadatas.append(
profile_event.delegate_debug_metadata
if profile_event.delegate_debug_metadata
else ""
)
# Update fields
if len(data) > 0:
ret_event.perf_data = PerfData(data)
if any(delegate_debug_metadatas):
ret_event._delegate_debug_metadatas = delegate_debug_metadatas
# add _start_time to the event
if len(stime) > 0:
ret_event._start_time = stime
@staticmethod
def _populate_debugging_related_fields(
ret_event: "Event",
debug_event_signature: Optional[DebugEventSignature],
events: List[InstructionEvent],
output_buffer: Optional[bytes] = None,
) -> None:
"""
Given a partially constructed Event, populate the fields related to
the debug events
Fields Updated:
name
delegate_debug_identifier
is_delegated_op
debug_data
"""
# Fill out fields from debug event signature
Event._populate_event_signature_fields(ret_event, debug_event_signature)
debug_data: List[flatcc.Value] = []
for event in events:
if (debug_events := event.debug_events) is None:
continue
# Populate on the first iteration only, then verify equivalence for others
if len(debug_data) == 0:
debug_data = [debug_event.debug_entry for debug_event in debug_events]
else:
for debug_event, value in zip(debug_events, debug_data):
v1 = inflate_runtime_output(debug_event.debug_entry, output_buffer)
v2 = inflate_runtime_output(value, output_buffer)
assert is_inference_output_equal(
v1, v2
), """Corresponding debug events in multiple iterations of the model
must have the same debug entry values. This is not the case for the
intermediate data present in this ETDump and indicates potential issues
with the model/runtime."""
ret_event.debug_data = [
inflate_runtime_output(debug_value, output_buffer)
for debug_value in debug_data
]
def _associate_with_op_graph_nodes(
self,
debug_handle_to_op_node_map: Dict[int, List[OperatorNode]],
) -> None:
"""
Helper function to populate the stack_traces, module_hierarchy and op_types attributes
based on the debug handles of this event
"""
# Framework events aren't logically associated with any nodes
if self.name in RESERVED_FRAMEWORK_EVENT_NAMES:
return
if (debug_handles := self.debug_handles) is None:
return
if isinstance(debug_handles, int):
debug_handles = [debug_handles]
for handle in debug_handles:
nodes = debug_handle_to_op_node_map.get(handle, None)
if nodes is None:
continue
for node in nodes:
# Attach node metadata including stack traces, module hierarchy and op_types to this event
if node is not None and (metadata := node.metadata) is not None:
if node.name not in self.stack_traces:
self.stack_traces[node.name] = metadata.get("stack_trace")
self.module_hierarchy[node.name] = metadata.get(
"nn_module_stack"
)
if node.op:
# TODO: consider having this as a dict from node.name -> node.op
self.op_types += [node.op]
[docs]@dataclass
class EventBlock:
r"""
An `EventBlock` contains a collection of events associated with a particular profiling/debugging block retrieved from the runtime.
Each `EventBlock` represents a pattern of execution. For example, model initiation and loading lives in a single `EventBlock`.
If there's a control flow, each branch will be represented by a separate `EventBlock`.
Args:
name: Name of the profiling/debugging block.
events: List of `Event`\ s associated with the profiling/debugging block.
bundled_input_idx: Index of the Bundled Input that this EventBlock corresponds to.
run_output: Run output extracted from the encapsulated Events
"""
name: str
events: List[Event] = dataclasses.field(default_factory=list)
source_time_scale: TimeScale = TimeScale.NS
target_time_scale: TimeScale = TimeScale.MS
bundled_input_index: Optional[int] = None
run_output: Optional[ProgramOutput] = None
reference_output: Optional[ProgramOutput] = None
def to_dataframe(
self, include_units: bool = False, include_delegate_debug_data: bool = False
) -> pd.DataFrame:
"""
Converts the EventBlock into a DataFrame with each row being an event instance
Note: Rows that have an event_name = OPERATOR_CALL correspond to the perf of the
previous operator + framework tax of making said operator call.
Args:
include_units: Whether headers should include units (default false)
include_delegate_debug_data: Whether to show the delegate debug data
Returns:
A pandas DataFrame containing the data of each Event instance in this EventBlock.
"""
units = " (" + self.target_time_scale.value + ")" if include_units else ""
df = pd.concat([e.to_dataframe(units) for e in self.events], ignore_index=True)
df.insert(
0,
"event_block_name",
np.asarray([self.name for _ in range(len(self.events))]),
allow_duplicates=True,
)
# Add Delegate Debug Metadata columns
if include_delegate_debug_data:
delegate_data = []
for event in self.events:
if (metadata := event.delegate_debug_metadatas) is not None and len(
metadata
) > 0:
if isinstance(metadata, list):
delegate_data.append(
pd.Series([metadata], index=["delegate_debug_metadata"])
)
elif isinstance(metadata, dict):
delegate_data.append(pd.Series(metadata))
else:
raise ValueError(
f"Unexpected type for delegate_debug_metadata: {type(metadata)}"
)
else:
delegate_data.append(pd.Series())
if any(not data.empty for data in delegate_data):
df = pd.concat([df, pd.DataFrame(delegate_data)], axis=1)
return df
@staticmethod
def _gen_from_etdump(
etdump: ETDumpFlatCC,
source_time_scale: TimeScale = TimeScale.NS,
target_time_scale: TimeScale = TimeScale.MS,
output_buffer: Optional[bytes] = None,
delegate_metadata_parser: Optional[
Callable[[List[str]], Dict[str, Any]]
] = None,
delegate_time_scale_converter: Optional[
Callable[[Union[int, str], Union[int, float]], Union[int, float]]
] = None,
) -> List["EventBlock"]:
"""
Given an etdump, generate a list of EventBlocks corresponding to the
contents.
An optional (inverse) scale factor can be provided to adjust the
etdump timestamps associated with each EventBlocks
An optional buffer to inflate etdump references
An optional delegate metadata parser function to parse delegate profiling metadata
"""
# Map each RunSignatures to instances of its constituent events.
# The value of the map is a GroupedRunInstance which contains:
# (1) a map from each EventSignature to InstructionEvents with the signature
# (2) the run output for this RunSignature
@dataclass
class GroupedRunInstances:
events: OrderedDict[EventSignature, List[InstructionEvent]]
run_output: ProgramOutput
run_groups: Mapping[RunSignature, GroupedRunInstances] = defaultdict(
lambda: GroupedRunInstances(OrderedDict(), [])
)
# Collect all the run data
for run in etdump.run_data:
if (run_events := run.events) is None:
continue
# Collate the run_events into InstructionEvents
instruction_events: List[InstructionEvent] = (
InstructionEvent.gen_from_events(run_events)
)
# Map EventSignatures to the InstructionEvents
event_signatures: Dict[EventSignature, InstructionEvent] = OrderedDict()
for instruction_event in instruction_events:
if (
instruction_event.debug_events is None
and instruction_event.profile_events is None
):
# Currently corresponds to run output
continue
generated_event_signatures: List[
Tuple[EventSignature, InstructionEvent]
] = EventSignature.gen_from_instruction_event(instruction_event)
for (
event_signature,
filtered_instruction_event,
) in generated_event_signatures:
event_signatures[event_signature] = filtered_instruction_event
# Create a RunSignature from the EventSignatures
run_signature = RunSignature(
name=run.name,
events=tuple(event_signatures.keys()),
bundled_input_index=run.bundled_input_index,
)
# Update the Run Groups, indexed on the RunSignature
run_signature_events: OrderedDict[
EventSignature, List[InstructionEvent]
] = run_groups[run_signature].events
for event_signature, event in event_signatures.items():
run_signature_events.setdefault(event_signature, []).append(event)
# Populate (or Verify if already populated) Run Outputs
run_outputs: ProgramOutput = EventBlock._collect_run_outputs(
run_events, output_buffer
)
if len(existing_run_outputs := run_groups[run_signature].run_output) == 0:
existing_run_outputs.extend(run_outputs)
else:
verify_debug_data_equivalence(existing_run_outputs, run_outputs)
# Construct the EventBlocks
event_blocks = []
scale_factor = calculate_time_scale_factor(source_time_scale, target_time_scale)
for run_signature, grouped_run_instance in run_groups.items():
run_group: OrderedDict[EventSignature, List[InstructionEvent]] = (
grouped_run_instance.events
)
run_outputs: ProgramOutput = grouped_run_instance.run_output
# Construct the Events
events: List[Event] = [
Event._gen_from_inference_events(
signature,
instruction_events,
scale_factor,
output_buffer,
delegate_metadata_parser,
delegate_time_scale_converter,
)
for signature, instruction_events in run_group.items()
]
# Add the EventBlock to the return list
event_blocks.append(
EventBlock(
name=run_signature.name,
events=events,
source_time_scale=source_time_scale,
target_time_scale=target_time_scale,
bundled_input_index=run_signature.bundled_input_index,
run_output=run_outputs,
)
)
return event_blocks
@staticmethod
def _collect_run_outputs(
events: List[flatcc.Event], output_buffer: Optional[bytes] = None
) -> ProgramOutput:
"""
Given a list of events, search the events for ProgramOutputs (aka lists of InferenceOutputs) marked
as run outputs
"""
output_events = []
for event in events:
if event.debug_event is None:
continue
if event.debug_event.debug_entry is None:
raise RuntimeError(
"Debug entry inside debug event should not be empty!"
)
if is_debug_output(event.debug_event.debug_entry):
output_events += [event]
return [
inflate_runtime_output(debug_event.debug_entry, output_buffer)
for output_event in output_events
if (debug_event := output_event.debug_event) is not None
]
# TODO: Considering changing ETRecord deserialization logic to cast the ints in string format to actual ints
def _gen_resolve_debug_handles(
self,
handle_map: Dict[str, List[int]],
delegate_map: Optional[Dict[str, DelegateMetadata]] = None,
instruction_id_to_num_outs_map: Dict[int, int] = None,
):
"""
Given mappings from instruction id to debug handles, populate the
debug_handles field of all underlying events
If the event is delegated, index with the instruction_id and delegate_debug_identifier
to obtain the debug_handle via the delegate map
"""
for event in self.events:
# Check if instruction_id is present in the event
if event._instruction_id is None:
continue
# Check for the instruction_id in handle map
if (instruction_id := str(event._instruction_id)) not in handle_map:
continue
num_outputs = 1
if instruction_id_to_num_outs_map is not None:
num_outputs = instruction_id_to_num_outs_map.get(instruction_id, 1)
event.num_outputs = num_outputs
# For non-delegated event, handles are found in handle_map
if (delegate_debug_id := event.delegate_debug_identifier) is None:
event.debug_handles = handle_map[instruction_id]
# DELEGATE_CALL is a special non-delegated event and benefits from having the name populated
if (
event.name == "DELEGATE_CALL"
and delegate_map is not None
and (delegate_metadata := delegate_map.get(instruction_id))
is not None
):
event.delegate_backend_name = delegate_metadata.get("name", "")
continue
# Check that the delegated event has a corresponding mapping
if (
delegate_map is None
or (delegate_metadata := delegate_map.get(instruction_id)) is None
):
event.debug_handles = handle_map[instruction_id]
log.warning(
f" No delegate mapping found for delegate with instruction id {event._instruction_id}"
)
continue
# For delegated events, handles are found via delegateMetadata
event.delegate_backend_name = delegate_metadata.get("name", "")
delegate_metadata_delegate_map = delegate_metadata.get("delegate_map") or {}
# delegate_debug_id can be either int based or string based, therefore we need to check both
debug_handles = delegate_metadata_delegate_map.get(
delegate_debug_id # pyre-ignore
)
if debug_handles is not None:
event.debug_handles = debug_handles
else:
event.debug_handles = delegate_metadata_delegate_map.get(
str(delegate_debug_id) # pyre-ignore
)
for key, value in delegate_metadata_delegate_map.items():
if key in str(delegate_debug_id):
event.debug_handles = value
class Inspector:
"""
APIs for examining model architecture and performance stats.
Public Attributes:
event_blocks: List["EventBlocks"]. Structured data from ETDump (correlated with ETRecord if provided).
Private Attributes:
_etrecord: Optional[ETRecord]. File under etrecord_path deserialized into an object.
"""
def __init__(
self,
etdump_path: Optional[str] = None,
etdump_data: Optional[bytes] = None,
etrecord: Optional[Union[ETRecord, str]] = None,
source_time_scale: TimeScale = TimeScale.NS,
target_time_scale: TimeScale = TimeScale.MS,
debug_buffer_path: Optional[str] = None,
delegate_metadata_parser: Optional[
Callable[[List[str]], Dict[str, Any]]
] = None,
delegate_time_scale_converter: Optional[
Callable[[Union[int, str], Union[int, float]], Union[int, float]]
] = None,
enable_module_hierarchy: bool = False,
) -> None:
r"""
Initialize an `Inspector` instance with the underlying `EventBlock`\ s populated with data from the provided ETDump path or binary,
and optional ETRecord path.
Args:
etdump_path: Path to the ETDump file. Either this parameter or etdump_data should be provided.
etdump_data: ETDump binary. Either this parameter or etdump_path should be provided.
etrecord: Optional ETRecord object or path to the ETRecord file.
source_time_scale: The time scale of the performance data retrieved from the runtime. The default time hook implentation in the runtime returns NS.
target_time_scale: The target time scale to which the users want their performance data converted to. Defaults to MS.
debug_buffer_path: Debug buffer file path that contains the debug data referenced by ETDump for intermediate and program outputs.
delegate_metadata_parser: Optional function to parse delegate metadata from an Profiling Event. Expected signature of the function is:
(delegate_metadata_list: List[bytes]) -> Union[List[str], Dict[str, Any]]
delegate_time_scale_converter: Optional function to convert the time scale of delegate profiling data. If not given, use the conversion ratio of
target_time_scale/source_time_scale.
enable_module_hierarchy: Enable submodules in the operator graph. Defaults to False.
Returns:
None
"""
if (source_time_scale == TimeScale.CYCLES) ^ (
target_time_scale == TimeScale.CYCLES
):
raise RuntimeError(
"For TimeScale in cycles both the source and target time scale have to be in cycles."
)
self._source_time_scale = source_time_scale
self._target_time_scale = target_time_scale
if delegate_time_scale_converter is None:
scale_factor = calculate_time_scale_factor(
source_time_scale, target_time_scale
)
delegate_time_scale_converter = (
lambda event_name, input_time: input_time / scale_factor
)
if etrecord is None:
self._etrecord = None
elif isinstance(etrecord, ETRecord):
self._etrecord = etrecord
elif isinstance(etrecord, str):
self._etrecord = parse_etrecord(etrecord_path=etrecord)
else:
raise TypeError("Unsupported ETRecord type")
if (etdump_path is None) == (etdump_data is None):
raise ValueError(
"Expecting exactly one of etdump_path or etdump_data to be specified."
)
# Create EventBlocks from ETDump
etdump = gen_etdump_object(etdump_path=etdump_path, etdump_data=etdump_data)
if debug_buffer_path is not None:
with open(debug_buffer_path, "rb") as f:
output_buffer = f.read()
else:
output_buffer = None
warnings.warn(
"Output Buffer not found. Tensor Debug Data will not be available.",
stacklevel=1,
)
self.event_blocks = EventBlock._gen_from_etdump(
etdump=etdump,
source_time_scale=self._source_time_scale,
target_time_scale=self._target_time_scale,
output_buffer=output_buffer,
delegate_metadata_parser=delegate_metadata_parser,
delegate_time_scale_converter=delegate_time_scale_converter,
)
# Connect ETRecord to EventBlocks
self.op_graph_dict: Optional[Mapping[str, OperatorGraph]] = None
# _consume_etrecord() will populate the _reference_outputs dict
# Key str is method name; value is list of ProgramOutputs because of list of test cases
self._reference_outputs: Dict[str, List[ProgramOutput]] = {}
self._enable_module_hierarchy = enable_module_hierarchy
self._consume_etrecord()
def _consume_etrecord(self) -> None:
"""
If an ETRecord is provided, connect it to the EventBlocks and populate the Event metadata.
Steps:
1. Debug Handle Symbolification:
For each Event, find the debug_handle counterparts using
ETRecord's debug_handle_map and delegate_map
2. Event Metadata Association:
For each Event, populate its metadata from OperatorGraph Nodes,
generated from ETRecord. The debug_handle is used to
identify the corresponding OperatorGraph Nodes.
3. Reference Outputs Extraction:
If there're reference outputs saved in ETRecord, assign each reference output to the corresponding
EventBlock based on the method name (currently assumes only "forward") and the
bundled_input_index of the EventBlock.
"""
if self._etrecord is None:
return
# (1) Debug Handle Symbolification
for event_block in self.event_blocks:
event_block._gen_resolve_debug_handles(
self._etrecord._debug_handle_map[FORWARD],
(
self._etrecord._delegate_map[FORWARD]
if self._etrecord._delegate_map is not None
else None
),
self._etrecord._instruction_id_to_num_outs_map[FORWARD],
)
# (2) Event Metadata Association
self.op_graph_dict = gen_graphs_from_etrecord(
etrecord=self._etrecord,
enable_module_hierarchy=self._enable_module_hierarchy,
)
debug_handle_to_op_node_map = create_debug_handle_to_op_node_mapping(
self.op_graph_dict[EDGE_DIALECT_GRAPH_KEY],
)
for event_block in self.event_blocks:
for event in event_block.events:
event._associate_with_op_graph_nodes(
debug_handle_to_op_node_map=debug_handle_to_op_node_map,
)
# (3) Reference Outputs Extraction
if self._etrecord._reference_outputs is not None:
self._reference_outputs = self._etrecord._reference_outputs
# Associate each reference output to the corresponding event block
for event_block in self.event_blocks:
index = event_block.bundled_input_index
if index is not None:
event_block.reference_output = self._reference_outputs[FORWARD][
index
]
def _get_aot_intermediate_outputs_and_op_names(
self,
disable_debug_handle_valdiation: bool = False,
) -> Tuple[Dict[DebugHandle, Any], Dict[DebugHandle, List[str]]]:
"""
Capture intermediate outputs only if _representative_inputs are provided
when using bundled program to create the etrecord
"""
if self._etrecord._representative_inputs is None:
return {}, {}
export_program = None
# Will use the exported program to extract intermediate output if and only if exported_program has been provided, and it is one of the ancestors of the edge_dialect_program
if self._etrecord.exported_program and propagate_back_debug_handle(
self._etrecord.exported_program,
self._etrecord.export_graph_id,
self._etrecord.edge_dialect_program,
disable_debug_handle_valdiation,
):
export_program = self._etrecord.exported_program
else:
log.warning(
"Either aten dialect exported program is not in ETRecord, or it is not one of the ancestors of current edge dialect program."
"Will fall back to use edge dialect program to extract intermediate output",
)
export_program = self._etrecord.edge_dialect_program
graph_module = export_program.module()
aot_debug_handle_to_op_name = get_aot_debug_handle_to_op_name_mapping(
graph_module
)
capturer = IntermediateOutputCapturer(graph_module)
aot_intermediate_outputs = capturer.run_and_capture(
self._etrecord._representative_inputs
)
return aot_intermediate_outputs, aot_debug_handle_to_op_name
# TODO: Make it more extensible to further merge overlapping debug handles
def _get_runtime_intermediate_outputs_and_op_names(
self,
) -> Tuple[Dict[DebugHandle, Tuple[Any, int]], Dict[DebugHandle, List[str]]]:
"""
Retrieve the runtime intermediate outputs(debug handles and intermediate values mappings)
from the event blocks, along with the corresponding debug handles and op names mapping.
"""
debug_handle_to_output = {}
debug_handle_to_op_names = {}
for event_block in self.event_blocks:
for event in event_block.events:
# Skip OPERATOR_CALL events to avoid double-counting and exclude framework tax
if (
event.name in EXCLUDED_EVENTS_FOR_INTERMEDIATE_OUTPUT
or not event.op_types
):
continue
# Normalize debug_handle to a tuple
debug_handle = event.debug_handles
if isinstance(debug_handle, int):
debug_handle = (debug_handle,)
else:
debug_handle = tuple(debug_handle)
current_entry = debug_handle_to_output.get(
debug_handle, (-1, None, event.num_outputs)
)
# When event has same debug_handle, only keep the one with the largest instruction id
if event._instruction_id > current_entry[0]:
debug_handle_to_output[debug_handle] = (
event._instruction_id,
event.debug_data,
event.num_outputs,
)
# TODO: One debug handle can be associated with multiple op names
debug_handle_to_op_names[debug_handle] = [event.name]
debug_handle_to_output = merge_runtime_overlapping_debug_handles(
debug_handle_to_output
)
return {
k: (v[1], v[2]) for k, v in debug_handle_to_output.items()
}, debug_handle_to_op_names
def to_dataframe(
self,
include_units: bool = True,
include_delegate_debug_data: bool = False,
) -> pd.DataFrame:
"""
Args:
include_units: Whether headers should include units (default true)
include_delegate_debug_data: Whether to include delegate debug metadata (default false)
Returns:
Returns a pandas DataFrame of the Events in each EventBlock in the inspector, with each row representing an Event.
"""
df_list = [
event_block.to_dataframe(
include_units=include_units,
include_delegate_debug_data=include_delegate_debug_data,
)
for event_block in self.event_blocks
]
return pd.concat(df_list, ignore_index=True)
def _prepare_dataframe(
self,
include_units: bool = True,
include_delegate_debug_data: bool = False,
) -> pd.DataFrame:
"""
Args:
include_units: Whether headers should include units (default true)
include_delegate_debug_data: Whether to include delegate debug metadata (default false)
Returns:
Returns a pandas DataFrame of the Events in each EventBlock in the inspector, with additional filtering.
"""
combined_df = self.to_dataframe(include_units, include_delegate_debug_data)
# Filter out some columns and rows for better readability when printing
filtered_column_df = combined_df.drop(columns=EXCLUDED_COLUMNS_WHEN_PRINTING)
for filter_name in EXCLUDED_EVENTS_WHEN_PRINTING:
filtered_column_df = filtered_column_df[
~filtered_column_df["event_name"].str.contains(filter_name)
]
filtered_column_df.reset_index(drop=True, inplace=True)
return filtered_column_df
def print_data_tabular(
self,
file: IO[str] = sys.stdout,
include_units: bool = True,
include_delegate_debug_data: bool = False,
) -> None:
"""
Displays the underlying EventBlocks in a structured tabular format, with each row representing an Event.
Args:
file: Which IO stream to print to. Defaults to stdout.
Not used if this is in an IPython environment such as a Jupyter notebook.
include_units: Whether headers should include units (default true)
include_delegate_debug_data: Whether to include delegate debug metadata (default false)
Returns:
None
"""
df = self._prepare_dataframe(include_units, include_delegate_debug_data)
display_or_print_df(df, file)
def save_data_to_tsv(
self,
file: IO[str],
include_units: bool = True,
include_delegate_debug_data: bool = False,
) -> None:
"""
Stores the underlying EventBlocks in tsv format to facilitate copy-paste into spreadsheets.
Args:
file: Which IO stream to print to. Do not use stdout, as tab separator is not preserved.
include_units: Whether headers should include units (default true)
include_delegate_debug_data: Whether to include delegate debug metadata (default false)
Returns:
None
"""
df = self._prepare_dataframe(include_units, include_delegate_debug_data)
df.to_csv(file, sep="\t")
# TODO: write unit test
def find_total_for_module(self, module_name: str) -> float:
"""
Returns the total average compute time of all operators within the specified module.
Args:
module_name: Name of the module to be aggregated against.
Returns:
Sum of the average compute time (in seconds) of all operators within the module with "module_name".
"""
total = 0.0
for block in self.event_blocks:
for event in block.events:
# Skip OPERATOR_CALL events to avoid double-counting and exclude framework tax
if event.name == "OPERATOR_CALL":
continue
module_hierarchy = event.module_hierarchy.values()
for hierarchy in module_hierarchy:
if not hierarchy:
continue
found = any(module_name in key for key in hierarchy.keys())
if found:
if event.perf_data is not None:
total += event.perf_data.avg
break
return total
def get_op_list(
self, event_block: str, show_delegated_ops: Optional[bool] = True
) -> Dict[str, List[Event]]:
"""
Return a map of op_types to Events of that op_type
"""
# TODO: implement
return {}
def write_tensorboard_artifact(self, path: str) -> None:
"""
Write to the provided path, the artifacts required for visualization in TensorBoard
"""
# TODO: implement
pass
def get_exported_program(
self, graph: Optional[str] = None
) -> Optional[ExportedProgram]:
"""
Access helper for ETRecord, defaults to returning the Edge Dialect program.
Args:
graph: Optional name of the graph to access. If None, returns the Edge Dialect program.
Returns:
The ExportedProgram object of "graph".
"""
if self._etrecord is None:
log.warning(
"Exported program is only available when a valid etrecord_path was provided at the time of Inspector construction"
)
return None
return (
self._etrecord.edge_dialect_program
if graph is None
else self._etrecord.graph_map.get(graph)
)
def calculate_numeric_gap(
self, distance: str = "MSE", disable_debug_handle_valdiation: bool = False
):
"""
Compares logged intermediate outputs from the exported graph (in ETRecord)
with runtime outputs (in ETDump) using a user-specific numerical comparator.
If the exported graph is not supported, the function will fall back to use edge dialect graph.
To use this function, you must first generate the ETRecord with representative inputs,
and then create the Inspector instance with the ETRecord and ETDump. The Inspector can then
compare the intermediate outputs from the AOT and the runtime.
Args:
distance: the metrics the inspector will use for gap calculation. Should be one of "MSE", "L1" and "SNR".
disable_debug_handle_validation: Often when aten graph has symbolic shape nodes, and inbuilt ops like gt/lt etc.,
during re-export of such a graph 'from_node' information is lost from node.meta. As a result we loose connection
between edge IR nodes and aten nodes for such ops. By default we validate that every edge IR node has corresponding
node in aten IR, and when such validation fails numeric debugger falls back to edge IR as reference graph. This
flag allows one to override such behavior and make best effort comparison.
Returns:
pd.DataFrame: A DataFrame listing corresponding operator intermediate outputs from both stages and their computed numerical gaps.
"""
aot_intermediate_outputs, aot_debug_handle_to_op_names = (
self._get_aot_intermediate_outputs_and_op_names(
disable_debug_handle_valdiation
)
)
if len(aot_intermediate_outputs) == 0 or len(aot_debug_handle_to_op_names) == 0:
raise ValueError(
"Missing etrecord or missing representative inputs within etrecord, both of which are required for calculating numerical gap"
)
# The runtime_op_names will be used later to map runtime debug_handle to op_name
runtime_intermediate_outputs, runtime_debug_handle_to_op_names = (
self._get_runtime_intermediate_outputs_and_op_names()
)
mapping = map_runtime_aot_intermediate_outputs(
aot_intermediate_outputs, runtime_intermediate_outputs
)
metric = distance.strip().upper()
if metric == "MSE":
comparator = MSEComparator()
elif metric == "L1":
comparator = L1Comparator()
elif metric == "SNR":
comparator = SNRComparator()
else:
raise ValueError(f"Unsupported distance metric {distance!r}")
rows = []
for (aot_debug_handle, aot_intermediate_output), (
runtime_debug_handle,
runtime_intermediate_output,
) in mapping.items():
if aot_intermediate_output is None or runtime_intermediate_output is None:
continue
# If aot outputs length is > 1 then comparison fails since we dont really have
# any instances where runtime intermediate output is a tuple or list
# This does not happen when edge dialect program is reference for comparison
# but happens in aten graph where ops like unbind remain undecomposed
if (
isinstance(aot_intermediate_output, Sequence)
and len(aot_intermediate_output) > 1
):
continue
rows.append(
{
"aot_ops": find_op_names(
aot_debug_handle, aot_debug_handle_to_op_names
),
"aot_intermediate_output": aot_intermediate_output,
"runtime_ops": find_op_names(
runtime_debug_handle, runtime_debug_handle_to_op_names
),
"runtime_intermediate_output": runtime_intermediate_output,
"gap": compare_intermediate_outputs(
aot_intermediate_output, runtime_intermediate_output, comparator
),
}
)
return pd.DataFrame(rows)
