| #!/usr/bin/env fuchsia-vendored-python | |
| # Copyright 2025 The Fuchsia Authors. All rights reserved. | |
| # Use of this source code is governed by a BSD-style license that can be | |
| # found in the LICENSE file. | |
| import itertools | |
| import logging | |
| import typing | |
| from collections.abc import MutableSequence | |
| from reporting import metrics | |
| from trace_processing import trace_metrics, trace_model, trace_utils | |
| _LOGGER: logging.Logger = logging.getLogger(__name__) | |
| _GPU_USAGE_EVENT_NAME = "GPU Utilization" | |
| class GpuMetricsProcessor(trace_metrics.MetricsProcessor): | |
| """Computes GPU utilization metrics. | |
| GPU utilization metrics mark the percentage of time the GPU is busy performing | |
| work. It is a value from 0 to 1. | |
| """ | |
| def __init__( | |
| self, | |
| aggregates_only: bool = True, | |
| ): | |
| """Constructor. | |
| Args: | |
| aggregates_only: When True, generates GpuMin, GpuMax, GpuAverage and GpuP* (%iles). | |
| Otherwise generates GpuLoad metric with all gpu values. | |
| """ | |
| self.aggregates_only: bool = aggregates_only | |
| @property | |
| @typing.override | |
| def event_patterns(self) -> set[str]: | |
| return {_GPU_USAGE_EVENT_NAME} | |
| @typing.override | |
| def process_metrics( | |
| self, model: trace_model.Model | |
| ) -> MutableSequence[metrics.TestCaseResult]: | |
| gpu_starts: list[float] = [] | |
| gpu_percentages: list[float] = [] | |
| # Parse the start time and percentage for each `Event`. | |
| for event in trace_utils.filter_events( | |
| model.all_events(), | |
| name=_GPU_USAGE_EVENT_NAME, | |
| type=trace_model.CounterEvent, | |
| ): | |
| gpu_starts.append(event.start.to_epoch_delta().to_nanoseconds()) | |
| gpu_percentages.append(event.args.get("utilization") or 0) | |
| # The duration for each utilization is the time between it and the next one. | |
| # We don't know the duration for the last event so add one '0'. | |
| gpu_durations: list[float] = [ | |
| curr - prev for prev, curr in itertools.pairwise(gpu_starts) | |
| ] + [0] | |
| if len(gpu_percentages) == 0: | |
| _LOGGER.info( | |
| "No gpu usage measurements are present. Perhaps the trace " | |
| "duration is too short to provide gpu usage information" | |
| ) | |
| return [] | |
| if len(gpu_durations) != len(gpu_percentages): | |
| _LOGGER.warning( | |
| "The number of GPU duration segments don't match the number " | |
| f"of percentages. ({len(gpu_durations)} != {len(gpu_percentages)}) Data may be truncated." | |
| ) | |
| if self.aggregates_only: | |
| return trace_utils.standard_metrics_set( | |
| values=gpu_percentages, | |
| label_prefix="Gpu", | |
| unit=metrics.Unit.percent, | |
| durations=gpu_durations, | |
| ) | |
| return [ | |
| metrics.TestCaseResult( | |
| "GpuUtilization", metrics.Unit.percent, gpu_percentages | |
| ) | |
| ] |