diff options
author | Nathan Bailey <nathan.bailey@arm.com> | 2024-01-16 16:39:06 +0000 |
---|---|---|
committer | Nathan Bailey <nathan.bailey@arm.com> | 2024-02-09 14:31:46 +0000 |
commit | d08513a72e7fbf0626c3d69b9c4cc7056b3da4ae (patch) | |
tree | 3d332d70dbc57b6b7255bb4b429b322ad90bb2c1 /src | |
parent | be7bab89eb8ace6ab6a83687354beab156afb716 (diff) | |
download | mlia-d08513a72e7fbf0626c3d69b9c4cc7056b3da4ae.tar.gz |
feat: Integrate Vela's per-layer performance estimates
Resolves: MLIA-1055, MLIA-1056, MLIA-1057
Signed-off-by: Nathan Bailey <nathan.bailey@arm.com>
Change-Id: Id573cec94e4a69117051dcd5175f383c0955d890
Diffstat (limited to 'src')
-rw-r--r-- | src/mlia/backend/vela/compiler.py | 9 | ||||
-rw-r--r-- | src/mlia/backend/vela/performance.py | 151 | ||||
-rw-r--r-- | src/mlia/core/workflow.py | 3 | ||||
-rw-r--r-- | src/mlia/target/ethos_u/advisor.py | 3 | ||||
-rw-r--r-- | src/mlia/target/ethos_u/performance.py | 46 | ||||
-rw-r--r-- | src/mlia/target/ethos_u/reporters.py | 109 |
6 files changed, 285 insertions, 36 deletions
diff --git a/src/mlia/backend/vela/compiler.py b/src/mlia/backend/vela/compiler.py index b591056..fe9e365 100644 --- a/src/mlia/backend/vela/compiler.py +++ b/src/mlia/backend/vela/compiler.py @@ -1,4 +1,4 @@ -# SPDX-FileCopyrightText: Copyright 2022-2023, Arm Limited and/or its affiliates. +# SPDX-FileCopyrightText: Copyright 2022-2024, Arm Limited and/or its affiliates. # SPDX-License-Identifier: Apache-2.0 """Vela compiler wrapper module.""" from __future__ import annotations @@ -90,7 +90,7 @@ class VelaCompilerOptions: # pylint: disable=too-many-instance-attributes tensor_allocator: TensorAllocatorType = "HillClimb" cpu_tensor_alignment: int = Tensor.AllocationQuantum optimization_strategy: OptimizationStrategyType = "Performance" - output_dir: str = "output" + output_dir: Path = Path("output") recursion_limit: int = 1000 @@ -251,6 +251,7 @@ class VelaCompiler: # pylint: disable=too-many-instance-attributes verbose_register_command_stream=False, verbose_operators=False, verbose_weights=False, + verbose_performance=True, show_cpu_operations=False, tensor_allocator=self.tensor_allocator, timing=False, @@ -258,6 +259,10 @@ class VelaCompiler: # pylint: disable=too-many-instance-attributes cpu_tensor_alignment=self.cpu_tensor_alignment, ) + def return_compiler_options(self) -> CompilerOptions: + """Return CompilerOptions instance for test purposes.""" + return self._compiler_options() + def resolve_compiler_config( vela_compiler_options: VelaCompilerOptions, diff --git a/src/mlia/backend/vela/performance.py b/src/mlia/backend/vela/performance.py index a548b26..72a8ceb 100644 --- a/src/mlia/backend/vela/performance.py +++ b/src/mlia/backend/vela/performance.py @@ -1,11 +1,16 @@ -# SPDX-FileCopyrightText: Copyright 2022-2023, Arm Limited and/or its affiliates. +# SPDX-FileCopyrightText: Copyright 2022-2024, Arm Limited and/or its affiliates. # SPDX-License-Identifier: Apache-2.0 """Vela performance module.""" from __future__ import annotations +import csv import logging +import os +from collections import Counter from dataclasses import dataclass +from dataclasses import fields from pathlib import Path +from pydoc import locate import numpy as np from ethosu.vela.npu_performance import PassCycles @@ -37,6 +42,130 @@ class PerformanceMetrics: # pylint: disable=too-many-instance-attributes dram_memory_area_size: int on_chip_flash_memory_area_size: int off_chip_flash_memory_area_size: int + layerwise_performance_info: LayerwisePerfInfo + + +@dataclass +class LayerPerfInfo: # pylint: disable=too-many-instance-attributes + """Contains metrics from a row from the per-layer csv file from Vela.""" + + name: str + tflite_operator: str + sram_usage: int + op_cycles: int + npu_cycles: int + sram_access_cycles: int + dram_access_cycles: int + on_chip_flash_access_cycles: int + off_chip_flash_access_cycles: int + mac_count: int + util_mac_percentage: float + + def __repr__(self) -> str: + """Return String Representation of LayerPerfInfo object.""" + header_values = {key: value for key, value, _ in layer_metrics} + string_to_check = "" + for field in fields(self): + string_to_check += ( + f"{header_values[field.name]}: {getattr(self, field.name)}, " + ) + return string_to_check + + +@dataclass +class LayerwisePerfInfo: + """Contains all the per-layer metrics from the per-layer csv file from Vela.""" + + layerwise_info: list[LayerPerfInfo] + + def __repr__(self) -> str: + """Return String Representation of LayerwisePerfInfo object.""" + strings_to_check_layerwise_object = "" + for layer in self.layerwise_info: + string_to_check = repr(layer) + strings_to_check_layerwise_object += string_to_check + return strings_to_check_layerwise_object + + +complete_layer_metrics = [ + ("tflite_operator", "TFLite_operator", "TFLite Operator"), + ("nng_operator", "NNG Operator", "NNG Operator"), + ("sram_usage", "SRAM Usage", "SRAM Usage"), + ("peak_percentage", "Peak%", "Peak SRAM Usage (%)"), + ("op_cycles", "Op Cycles", "OP Cycles"), + ("network_percentage_1", "Network%", "OP Cycles in Network (%)"), + ("npu_cycles", "NPU", "NPU Cycles"), + ("sram_access_cycles", "SRAM AC", "SRAM AC"), + ("dram_access_cycles", "DRAM AC", "DRAM AC"), + ("on_chip_flash_access_cycles", "OnFlash AC", "OnFlash AC"), + ("off_chip_flash_access_cycles", "OffFlash AC", "OffFlash AC"), + ("mac_count", "MAC Count", "MAC Count"), + ("network_percentage_2", "Network% (1)", "MAC Count in Network (%)"), + ("util_mac_percentage", "Util%", "MAC Util (%)"), + ("name", "Name", "Layer Name"), +] + +OUTPUT_METRICS = [field.name for field in fields(LayerPerfInfo)] + +layer_metrics = [ + layer_metric + for layer_metric in complete_layer_metrics + if layer_metric[0] in OUTPUT_METRICS +] +layer_metrics.sort(key=lambda e: OUTPUT_METRICS.index(e[0])) + + +def parse_layerwise_perf_csv( # pylint: disable=too-many-locals + vela_csv_file: Path, metrics: list +) -> LayerwisePerfInfo: + """Parse the per-layer csv file from backend vela.""" + if not vela_csv_file.is_file(): + raise FileNotFoundError(f"CSV File not found at {vela_csv_file}\n") + layerwise_info = [] # type: list[LayerPerfInfo] + with open(vela_csv_file, encoding="UTF-8") as csv_file: + layerwise_reader = csv.reader(csv_file, delimiter=",") + try: + headers = list(next(layerwise_reader)) + except StopIteration: + return LayerwisePerfInfo(layerwise_info=layerwise_info) + headers_to_check_cpu_ops = headers.copy() + multiple_header_count = Counter(headers) + # Deal with multiple of the same values in CSV header. + for idx, header in enumerate(reversed(headers)): + if multiple_header_count[header] > 1: + headers[len(headers) - idx - 1] = ( + headers[len(headers) - idx - 1] + + " (" + + str(multiple_header_count[header] - 1) + + ")" + ) + multiple_header_count[header] -= 1 + for row in layerwise_reader: + row_as_dict = dict(zip(headers, row)) + if row == headers_to_check_cpu_ops: + continue + try: + key_types = { + field.name: locate(str(field.type)) + for field in fields(LayerPerfInfo) + } + ids_to_metrics = {} + for key, title, _ in metrics: + try: + ids_to_metrics[key] = key_types[key]( # type: ignore + row_as_dict[title] + ) + except ValueError as err: + if "invalid literal for int() with base 10" in str(err): + ids_to_metrics[key] = key_types[key]( # type: ignore + float(row_as_dict[title]) + ) + else: + raise + layerwise_info.append(LayerPerfInfo(**ids_to_metrics)) + except KeyError as err: + raise KeyError("Generated CSV missing expected headers") from err + return LayerwisePerfInfo(layerwise_info=layerwise_info) def estimate_performance( @@ -61,11 +190,26 @@ def estimate_performance( ) optimized_model = vela_compiler.compile_model(initial_model) + output_dir = optimized_model.compiler_options.output_dir + csv_paths = [entry for entry in os.listdir(output_dir) if "per-layer.csv" in entry] + model_name = str(model_path.stem) + csv_file_found = None + for path in csv_paths: + if model_name in path: + csv_file_found = path + if csv_file_found is None: + raise FileNotFoundError("Vela per-layer CSV file not found") + csv_path = Path(output_dir) / csv_file_found + layerwise_performance_info = parse_layerwise_perf_csv( + vela_csv_file=csv_path, metrics=layer_metrics + ) - return _performance_metrics(optimized_model) + return _performance_metrics(layerwise_performance_info, optimized_model) -def _performance_metrics(optimized_model: OptimizedModel) -> PerformanceMetrics: +def _performance_metrics( + layerwise_performance_info: LayerwisePerfInfo, optimized_model: OptimizedModel +) -> PerformanceMetrics: """Return performance metrics for optimized model.""" cycles = optimized_model.nng.cycles @@ -96,4 +240,5 @@ def _performance_metrics(optimized_model: OptimizedModel) -> PerformanceMetrics: dram_memory_area_size=memory_usage(MemArea.Dram), on_chip_flash_memory_area_size=memory_usage(MemArea.OnChipFlash), off_chip_flash_memory_area_size=memory_usage(MemArea.OffChipFlash), + layerwise_performance_info=layerwise_performance_info, ) diff --git a/src/mlia/core/workflow.py b/src/mlia/core/workflow.py index 9f8ac83..c645857 100644 --- a/src/mlia/core/workflow.py +++ b/src/mlia/core/workflow.py @@ -1,4 +1,4 @@ -# SPDX-FileCopyrightText: Copyright 2022-2023, Arm Limited and/or its affiliates. +# SPDX-FileCopyrightText: Copyright 2022-2024, Arm Limited and/or its affiliates. # SPDX-License-Identifier: Apache-2.0 """Module for executors. @@ -114,6 +114,7 @@ class DefaultWorkflowExecutor(WorkflowExecutor): self.before_start() collected_data = self.collect_data() + analyzed_data = self.analyze_data(collected_data) self.produce_advice(analyzed_data) diff --git a/src/mlia/target/ethos_u/advisor.py b/src/mlia/target/ethos_u/advisor.py index 9f5b3a6..b5932d0 100644 --- a/src/mlia/target/ethos_u/advisor.py +++ b/src/mlia/target/ethos_u/advisor.py @@ -1,4 +1,4 @@ -# SPDX-FileCopyrightText: Copyright 2022-2023, Arm Limited and/or its affiliates. +# SPDX-FileCopyrightText: Copyright 2022-2024, Arm Limited and/or its affiliates. # SPDX-License-Identifier: Apache-2.0 """Ethos-U MLIA module.""" from __future__ import annotations @@ -44,6 +44,7 @@ class EthosUInferenceAdvisor(DefaultInferenceAdvisor): """Return list of the data collectors.""" model = self.get_model(context) target_config = self._get_target_config(context) + target_config.compiler_options.output_dir = context.output_dir # type: ignore backends = self._get_backends(context) collectors: list[DataCollector] = [] diff --git a/src/mlia/target/ethos_u/performance.py b/src/mlia/target/ethos_u/performance.py index a0526e4..8decb75 100644 --- a/src/mlia/target/ethos_u/performance.py +++ b/src/mlia/target/ethos_u/performance.py @@ -1,4 +1,4 @@ -# SPDX-FileCopyrightText: Copyright 2022-2023, Arm Limited and/or its affiliates. +# SPDX-FileCopyrightText: Copyright 2022-2024, Arm Limited and/or its affiliates. # SPDX-License-Identifier: Apache-2.0 """Performance estimation.""" from __future__ import annotations @@ -13,6 +13,7 @@ import mlia.backend.vela.compiler as vela_comp import mlia.backend.vela.performance as vela_perf from mlia.backend.corstone import is_corstone_backend from mlia.backend.corstone.performance import estimate_performance +from mlia.backend.vela.performance import LayerwisePerfInfo from mlia.core.context import Context from mlia.core.performance import PerformanceEstimator from mlia.nn.select import OptimizationSettings @@ -95,16 +96,23 @@ class PerformanceMetrics: target_config: EthosUConfiguration npu_cycles: NPUCycles | None memory_usage: MemoryUsage | None + layerwise_perf_info: LayerwisePerfInfo | None def in_kilobytes(self) -> PerformanceMetrics: """Return metrics with memory usage in KiB.""" if self.memory_usage is None: return PerformanceMetrics( - self.target_config, self.npu_cycles, self.memory_usage + self.target_config, + self.npu_cycles, + self.memory_usage, + self.layerwise_perf_info, ) return PerformanceMetrics( - self.target_config, self.npu_cycles, self.memory_usage.in_kilobytes() + self.target_config, + self.npu_cycles, + self.memory_usage.in_kilobytes(), + self.layerwise_perf_info, ) @@ -119,7 +127,9 @@ class OptimizationPerformanceMetrics: class VelaPerformanceEstimator( - PerformanceEstimator[Union[Path, ModelConfiguration], MemoryUsage] + PerformanceEstimator[ + Union[Path, ModelConfiguration], tuple[MemoryUsage, LayerwisePerfInfo] + ] ): """Vela based performance estimator.""" @@ -128,7 +138,9 @@ class VelaPerformanceEstimator( self.context = context self.target = target_config - def estimate(self, model: Path | ModelConfiguration) -> MemoryUsage: + def estimate( + self, model: Path | ModelConfiguration + ) -> tuple[MemoryUsage, LayerwisePerfInfo]: """Estimate performance.""" with log_action("Getting the memory usage metrics ..."): model_path = ( @@ -141,12 +153,15 @@ class VelaPerformanceEstimator( model_path, self.target.compiler_options ) - return MemoryUsage( - vela_perf_metrics.sram_memory_area_size, - vela_perf_metrics.dram_memory_area_size, - vela_perf_metrics.unknown_memory_area_size, - vela_perf_metrics.on_chip_flash_memory_area_size, - vela_perf_metrics.off_chip_flash_memory_area_size, + return ( + MemoryUsage( + vela_perf_metrics.sram_memory_area_size, + vela_perf_metrics.dram_memory_area_size, + vela_perf_metrics.unknown_memory_area_size, + vela_perf_metrics.on_chip_flash_memory_area_size, + vela_perf_metrics.off_chip_flash_memory_area_size, + ), + vela_perf_metrics.layerwise_performance_info, ) @@ -238,12 +253,15 @@ class EthosUPerformanceEstimator( memory_usage = None npu_cycles = None + layerwise_perf_info = None for backend in self.backends: if backend == "vela": vela_estimator = VelaPerformanceEstimator( self.context, self.target_config ) - memory_usage = vela_estimator.estimate(tflite_model) + memory_usage, layerwise_perf_info = vela_estimator.estimate( + tflite_model + ) elif is_corstone_backend(backend): corstone_estimator = CorstonePerformanceEstimator( self.context, self.target_config, backend @@ -256,4 +274,6 @@ class EthosUPerformanceEstimator( backend, ) - return PerformanceMetrics(self.target_config, npu_cycles, memory_usage) + return PerformanceMetrics( + self.target_config, npu_cycles, memory_usage, layerwise_perf_info + ) diff --git a/src/mlia/target/ethos_u/reporters.py b/src/mlia/target/ethos_u/reporters.py index 711f036..b747ce5 100644 --- a/src/mlia/target/ethos_u/reporters.py +++ b/src/mlia/target/ethos_u/reporters.py @@ -1,14 +1,16 @@ -# SPDX-FileCopyrightText: Copyright 2022-2023, Arm Limited and/or its affiliates. +# SPDX-FileCopyrightText: Copyright 2022-2024, Arm Limited and/or its affiliates. # SPDX-License-Identifier: Apache-2.0 """Reports module.""" from __future__ import annotations from collections import defaultdict +from dataclasses import fields from typing import Any from typing import Callable from mlia.backend.vela.compat import Operator from mlia.backend.vela.compat import Operators +from mlia.backend.vela.performance import layer_metrics from mlia.core.advice_generation import Advice from mlia.core.reporters import report_advice from mlia.core.reporting import BytesCell @@ -16,6 +18,7 @@ from mlia.core.reporting import Cell from mlia.core.reporting import ClockCell from mlia.core.reporting import Column from mlia.core.reporting import CompoundFormatter +from mlia.core.reporting import CompoundReport from mlia.core.reporting import CyclesCell from mlia.core.reporting import Format from mlia.core.reporting import NestedReport @@ -237,10 +240,59 @@ def report_target_details(target_config: EthosUConfiguration) -> Report: ) -def metrics_as_records(perf_metrics: list[PerformanceMetrics]) -> list[tuple]: +def metrics_as_records( + perf_metrics: list[PerformanceMetrics], +) -> tuple[list[tuple], list[tuple]]: """Convert perf metrics object into list of records.""" perf_metrics = [item.in_kilobytes() for item in perf_metrics] + def _layerwise_as_metrics( + perf_metrics: list[PerformanceMetrics], + ) -> list[tuple]: + metric_map = defaultdict(list) # type: dict[str, list] + format_types = {int: "12,d", str: "", float: "12.2f"} + rows = [] + for perf_metric in perf_metrics: + if perf_metric.layerwise_perf_info: + for layerwise_metric in perf_metric.layerwise_perf_info.layerwise_info: + field_names = [ + field.name + for field in fields(layerwise_metric) + if field.name != "name" + ] + duplicate_idx = 1 + dict_key = getattr(layerwise_metric, "name") + while dict_key in metric_map: + dict_key = ( + getattr(layerwise_metric, "name") + + " (" + + str(duplicate_idx) + + ")" + ) + duplicate_idx += 1 + for field_name in field_names: + metric_map[dict_key].append( + getattr(layerwise_metric, field_name) + ) + rows = [ + ( + name, + *( + Cell( + value, + Format( + str_fmt=format_types[type(value)] + if type(value) in format_types + else "" + ), + ) + for value in values + ), + ) + for name, values in metric_map.items() + ] + return rows + def _cycles_as_records(perf_metrics: list[PerformanceMetrics]) -> list[tuple]: metric_map = defaultdict(list) for metrics in perf_metrics: @@ -306,7 +358,7 @@ def metrics_as_records(perf_metrics: list[PerformanceMetrics]) -> list[tuple]: _data_beats_as_records, ) for metrics in metrics_func(perf_metrics) - ] + ], _layerwise_as_metrics(perf_metrics) def report_perf_metrics( @@ -315,9 +367,9 @@ def report_perf_metrics( """Return comparison table for the performance metrics.""" if isinstance(perf_metrics, PerformanceMetrics): perf_metrics = [perf_metrics] + rows, layerwise_rows = metrics_as_records(perf_metrics) - rows = metrics_as_records(perf_metrics) - + # Create a seperate table for layerwise data if len(perf_metrics) == 2: return Table( columns=[ @@ -349,17 +401,42 @@ def report_perf_metrics( alias="performance_metrics", notes="IMPORTANT: The performance figures above refer to NPU only", ) - - return Table( - columns=[ - Column("Metric", alias="metric", fmt=Format(wrap_width=30)), - Column("Value", alias="value", fmt=Format(wrap_width=15)), - Column("Unit", alias="unit", fmt=Format(wrap_width=15)), - ], - rows=rows, - name="Performance metrics", - alias="performance_metrics", - notes="IMPORTANT: The performance figures above refer to NPU only", + if layerwise_rows == []: + return Table( + columns=[ + Column("Metric", alias="metric", fmt=Format(wrap_width=30)), + Column("Value", alias="value", fmt=Format(wrap_width=15)), + Column("Unit", alias="unit", fmt=Format(wrap_width=15)), + ], + rows=rows, + name="Performance metrics", + alias="performance_metrics", + notes="IMPORTANT: The performance figures above refer to NPU only", + ) + return CompoundReport( + [ + Table( + columns=[ + Column("Metric", alias="metric", fmt=Format(wrap_width=30)), + Column("Value", alias="value", fmt=Format(wrap_width=15)), + Column("Unit", alias="unit", fmt=Format(wrap_width=15)), + ], + rows=rows, + name="Performance metrics", + alias="performance_metrics", + notes="IMPORTANT: The performance figures above refer to NPU only", + ), + Table( + columns=[ + Column(name, alias=alias, fmt=Format(wrap_width=30)) + for alias, _, name in layer_metrics + ], + rows=layerwise_rows, + name="Layer-Wise Metrics", + alias="layerwise_metrics", + notes="", + ), + ] ) |