Run the default VoiceMark benchmark. On Linux:
./synthmark.app
On Android, run as root so that SynthMark can request SCHED_FIFO:
adb root
adb shell synthmark
If you run with the -h option you will get some “help” information:
SynthMark version 1.26
synthmark -t{test} -n{numVoices} -d{noteOnDelay} -p{percentCPU} -r{sampleRate} -s{seconds} -b{burstSize} -c{cpuAffinity}
-t{test}, v=voice, l=latency, j=jitter, u=utilization, s=series_util, c=clock_ramp, a=automated, default is v
-a{audioLevel} 0 = normal thread, 1 = audio callback (default), 2 = audio output
-b{burstSize} frames read by virtual hardware at one time, default = 96
-B{bursts} initial buffer size in bursts, default = 1
-c{cpuAffinity} index of CPU to run on, default = UNSPECIFIED
-d{noteOnDelay} seconds to delay the first NoteOn, default = 0
-f{enable} use SCHED_FIFO for normal thread, 0 = off, 1 = on (default)
-n{numVoices} to render, default = 8
-N{numVoices} to render for toggling high load, only for -t{l|j|c|s}
-m{voicesMode} algorithm to choose the number of voices in the range
[-n, -N]. This value can be 'l' for a linear increment, 'r' for a
random choice, or 's' to switch between -n and -N. default = s
-p{percentCPU} target load, default = 50
-r{sampleRate} should be typical, 44100, 48000, etc. default is 48000
-s{seconds} to run the test, latencyMark may take longer, default is 10
-u{utilClampLevel} 0 = off (default), 1 = on, 2 = on verbose, >2 = fixed
Using utilClamp helps the scheduler adapt to dynamic workloads.
-w{workloadHintsEnabled} 0 = no (default), 1 = give workload hints to scheduler
-z{enable} use ADPF for performance hints, 0 = off (default), 1 = on
VoiceMark measures CPU performance and the behavior of the CPU governor.
Run VoiceMark with the percent CPU set to 50% for 20 seconds. Then run it again at 80%. Note that the number of voices may not be linear because of the CPU frequency governor.
synthmark -tv -s20 -p50
synthmark -tv -s20 -p80
JitterMark measures thread scheduling, preemption and the behavior of the CPU governor.
Run the JitterMark with 20 voices, and 96 frame burst size.
adb shell synthmark -tj -n20 -b96
You can copy and paste the CSV output into a spreadsheet and plot a histogram. In the chart below, for example, we plotted the LOG base 2 of the counts against milliseconds. Jitter Histogram for SCHED_FIFO Thread on Pixel 7 Pro.
LatencyMark measures the output latency on the virtual audio device that is required to avoid glitches.
Run the LatencyMark with 4 voices.
adb shell synthmark -tl -n4
These tests are designed to give an overall measure of the real-time performance of the device.
Step #1 - Prepare the device
Ensure the device is stationary and the screen is off (both screen and movement can influence governor behaviour)
Run as root so that SynthMark can request SCHED_FIFO.
adb root
Step #2 : Measure maximum voices, M
Run VoiceMark90 for 30 seconds (any less than this results in unreliable data caused by varying CPU frequency) to obtain M voices at 90% CPU load
adb shell synthmark -tv -s30 -p90
Look for a report of the number of voices. Here is an example:
VoiceMark_90 = 192.191, normalized to 100% = 213.546
Step #3: JitterMark
The theoretical maximum number of voices in this example is 213. We calculate N = M / 2
N = 213 / 2 = 106
Now run JitterMark with N.
adb shell synthmark -tj -n106
Step #4: LatencyMark with light steady Load
In this test we measure the minimum latency required to generate a modest number voices without audible glitches.
adb shell synthmark -tl -n16
Step #4: LatencyMark with variable load
In this test we increase and decrease the load to see how the CPU governor responds.
adb shell synthmark -tl -n16 -N64