Derived Measurements

The DerivedMeasurements API provides a consistent way of performing post processing on a provided MeasurementCsv file.

Example

The following example shows how to use an implementation of a DerivedMeasurement to obtain a list of calculated DerivedMetric’s.

# Import the relevant derived measurement module
# in this example the derived energy module is used.
In [1]: from devlib import DerivedEnergyMeasurements

# Obtain a MeasurementCsv file from an instrument or create from
# existing .csv file. In this example an existing csv file is used which was
# created with a sampling rate of 100Hz
In [2]: from devlib import MeasurementsCsv
In [3]: measurement_csv = MeasurementsCsv('/example/measurements.csv', sample_rate_hz=100)

# Process the file and obtain a list of the derived measurements
In [4]: derived_measurements = DerivedEnergyMeasurements.process(measurement_csv)

In [5]: derived_measurements
Out[5]: [device_energy: 239.1854075 joules, device_power: 5.5494089227 watts]

API

Derived Measurements

class devlib.derived.DerivedMeasurements

The DerivedMeasurements class provides an API for post-processing instrument output offline (i.e. without a connection to the target device) to generate additional metrics.

DerivedMeasurements.process(measurement_csv)

Process a MeasurementsCsv, returning a list of DerivedMetric and/or MeasurementsCsv objects that have been derived from the input. The exact nature and ordering of the list members is specific to individual ‘class’DerivedMeasurements implementations.

DerivedMeasurements.process_raw(*args)

Process raw output from an instrument, returning a list DerivedMetric and/or MeasurementsCsv objects that have been derived from the input. The exact nature and ordering of the list members is specific to individual ‘class’DerivedMeasurements implementations.

The arguments to this method should be paths to raw output files generated by an instrument. The number and order of expected arguments is specific to particular implementations.

Derived Metric

class devlib.derived.DerivedMetric

Represents a metric derived from previously collected Measurement``s. Unlike, a ``Measurement, this was not measured directly from the target.

DerivedMetric.name

The name of the derived metric. This uniquely defines a metric – two DerivedMetric objects with the same name represent to instances of the same metric (e.g. computed from two different inputs).

DerivedMetric.value

The numeric value of the metric that has been computed for a particular input.

DerivedMetric.measurement_type

The MeasurementType of the metric. This indicates which conceptual category the metric falls into, its units, and conversions to other measurement types.

DerivedMetric.units

The units in which the metric’s value is expressed.

Available Derived Measurements

Note

If a method of the API is not documented for a particular implementation, that means that it s not overridden by that implementation. It is still safe to call it – an empty list will be returned.

Energy

class devlib.derived.energy.DerivedEnergyMeasurements

The DerivedEnergyMeasurements class is used to calculate average power and cumulative energy for each site if the required data is present.

The calculation of cumulative energy can occur in 3 ways. If a site contains energy results, the first and last measurements are extracted and the delta calculated. If not, a timestamp channel will be used to calculate the energy from the power channel, failing back to using the sample rate attribute of the MeasurementCsv file if timestamps are not available. If neither timestamps or a sample rate are available then an error will be raised.

DerivedEnergyMeasurements.process(measurement_csv)

This will return total cumulative energy for each energy channel, and the average power for each power channel in the input CSV. The output will contain all energy metrics followed by power metrics. The ordering of both will match the ordering of channels in the input. The metrics will by named based on the sites of the corresponding channels according to the following patters: "<site>_total_energy" and "<site>_average_power".

FPS / Rendering

class devlib.derived.fps.DerivedGfxInfoStats(drop_threshold=5, suffix='-fps', filename=None, outdir=None)

Produces FPS (frames-per-second) and other derived statistics from GfxInfoFramesInstrument output. This takes several optional parameters in creation:

Parameters:

• drop_threshold – FPS in an application, such as a game, which this processor is primarily targeted at, cannot reasonably drop to a very low value. This is specified to this threshold. If an FPS for a frame is computed to be lower than this threshold, it will be dropped on the assumption that frame rendering was suspended by the system (e.g. when idling), or there was some sort of error, and therefore this should be used in performance calculations. defaults to 5.
• suffix – The name of the generated per-frame FPS csv file will be derived from the input frames csv file by appending this suffix. This cannot be specified at the same time as a filename.
• filename – As an alternative to the suffix, a complete file name for FPS csv can be specified. This cannot be used at the same time as the suffix.
• outdir – By default, the FPS csv file will be placed in the same directory as the input frames csv file. This can be changed by specifying an alternate directory here

Warning

Specifying both filename and oudir will mean that exactly the same file will be used for FPS output on each invocation of process() (even for different inputs) resulting in previous results being overwritten.

DerivedGfxInfoStats.process(measurement_csv)

Process the fames csv generated by GfxInfoFramesInstrument and returns a list containing exactly three entries: DerivedMetrics fps and total_frames, followed by a MeasurentCsv containing per-frame FPSs values.

DerivedGfxInfoStats.process_raw(gfxinfo_frame_raw_file)

As input, this takes a single argument, which should be the path to the raw output file of GfxInfoFramesInstrument. The returns stats accumulated by gfxinfo. At the time of writing, the stats (in order) are: janks, janks_pc (percentage of all frames), render_time_50th_ptile (50th percentile, or median, for time to render a frame), render_time_90th_ptile, render_time_95th_ptile, render_time_99th_ptile, missed_vsync, hight_input_latency, slow_ui_thread, slow_bitmap_uploads, slow_issue_draw_commands. Please see the gfxinfo documentation for details.

class devlib.derived.fps.DerivedSurfaceFlingerStats(drop_threshold=5, suffix='-fps', filename=None, outdir=None)

Produces FPS (frames-per-second) and other derived statistics from SurfaceFlingerFramesInstrument output. This takes several optional parameters in creation:

Parameters:

• drop_threshold – FPS in an application, such as a game, which this processor is primarily targeted at, cannot reasonably drop to a very low value. This is specified to this threshold. If an FPS for a frame is computed to be lower than this threshold, it will be dropped on the assumption that frame rendering was suspended by the system (e.g. when idling), or there was some sort of error, and therefore this should be used in performance calculations. defaults to 5.
• suffix – The name of the generated per-frame FPS csv file will be derived from the input frames csv file by appending this suffix. This cannot be specified at the same time as a filename.
• filename – As an alternative to the suffix, a complete file name for FPS csv can be specified. This cannot be used at the same time as the suffix.
• outdir – By default, the FPS csv file will be placed in the same directory as the input frames csv file. This can be changed by specifying an alternate directory here

Warning

Specifying both filename and oudir will mean that exactly the same file will be used for FPS output on each invocation of process() (even for different inputs) resulting in previous results being overwritten.

DerivedSurfaceFlingerStats.process(measurement_csv)

Process the fames csv generated by SurfaceFlingerFramesInstrument and returns a list containing exactly three entries: DerivedMetrics fps and total_frames, followed by a MeasurentCsv containing per-frame FPSs values, followed by janks janks_pc, and missed_vsync metrics.