November 11, 2015, ARM TechCon, Santa Clara, CA--Mike Assenti from Dolby Laboratories talked about the energy consumed in audio processing tasks. He provided definitions and basic techniques for taking power measurements in smaller devices.
The growing number of apps for audio functions in mobile devices is a clear indicator of user desire for better audio. The drivers for better audio processing are the expectations from users for better audio for phone calls and for entertainment. At the same time, users want longer battery life, for longer listening times.
The definition for power consumption and measurements can include a direct measurement of the rate at which energy flows from battery to handset: milliWatts (mW), total energy consumed over a fixed period: joules (J), or more indirect measurement like battery run down time: minutes of battery life, software tooling to report power consumption of specific process, or simulation tools to calculate the power.
The indirect measurements are simpler and more likely to be repeatable at the cost of accuracy and precision. The direct testing solves the accuracy and repeatable issues but can require complex setup and human interpretation. All of the testing methods are influenced by the operating processes, and a single app may invoke many functions.
![]( "dolby measurement")
A direct measurement of power requires some additional fixturing.
Dolby uses direct physical testing with as many standard devices and measurement procedures as possible to try for greater repeatability and uniformity across brands and apps. A standard device is used for A-B comparisons and as a relative standard and the tests use some standard player with specified brightness level and loudness. The units under test are put into airplane mode and all unnecessary apps and processes are killed. The intent is to place the units under test as close as possible into the same conditions.
Some challenges in the testing include the need to have the whole audio chain running even though you are only trying to test one particular aspect, such as keeping the audio running while testing a video codec. Testing uses a number of different inputs to allow for averaging and discarding of the outliers due to noisy signals or power spikes during testing.
Unfortunately, their test results show that there is no such thing as a single power efficiency for a piece of code as the power varies with different parts of the algorithms. Although simulation can be more precise, it cannot address all of the possible situations and interactions that can occur in audio and video processing. as a result, human analysis is crucial to understand the context and interactions. It is possible to do more and different types of measurements, but the tradeoff is between detail and efficiency.
