Adjust OpenSL ES buffer queue length based on capacity

[philburk]

Change kBufferQueueLength constant to mBufferQueueLength variable. Set mBufferQueueLength based on capacity and callbacksize.

See experiment on the effect of kBufferQueueLength on latency in: Why Oboe only enqueue one buffer? #464

Understanding FramesPerCallback / kBufferQueueLength / Buffer Size #736

According to this internal bug:

b/27819623 | Allow OpenSL ES to specify a total buffer size for a fast track

starting in N, adding more buffers would improve glitch protection. So this change would help devices running 7.0, 7.1 and 8.0. That would be about 1/3 of Android devices. For 8.1 and later, AAudio is available.

[Update: as of August 2022, 17.5% of devices are running < 8.1]

[klw111]

Allow OpenSL ES to specify a total buffer size for a fast track Has this feature implemented in branch 1.3-stable？

[philburk]

No. This issue is still open.

[philburk]

Measure the effect of changing kBufferQueueLength on Sargo running SP2A with OpenSL ES with Low Latency. Burst = 192.

kBufferQueueLength	RT latency	Capacity
2	22.2
3	24.1	576
4	25.9	768

Latency increases by 2 msec (96 frames) per buffer. Not 192!

Used main Oboetester screen to set frames per callback to 96.

kBufferQueueLength	RT latency	Capacity
2	17.8	192
3	19.9	288
4	22.2	384

The latency still increases by 2 msec/buffer but the latency is lower!

I then set kBufferQueueLength and added some code to the beginning of processBufferCallback(). Burst = 192. if (getBufferDepth(bq) >= 3) { return false; }

maxDepth	RT latency	Capacity
3	34.2	1536
4	34.0	1536
5	34.27	1536

This does not seem to affect latency! I expected it to limit the amount of data in the buffer. I added some logs and the depth was always zero.

[philburk]

I then added a loop in processBufferCallback:

while (depth < 4 && countdown-- > 0 && !shouldStopStream) {

In the first callback I could see the depth grow to four then stabilize.

maxDepth	RT latency	Capacity
1	34.4	1536
2	37.8	1536
3	42.2	1536
4	45.8	1536
5	50.6	1536
6	54.3	1536
7	58.0	1536
8	62.3	1536

At 9 it failed to start because there were only 8 buffers allocated.

[philburk]

Hypothesis: in Android S the kBufferQueueLength affects the size of the internal buffer, which increases latency and should afford some glitch protection. Enqueuing multiple buffers increases the queue depth, adds latency. But I suspect these buffers are stored on the client side and will not help with glitch protection.

[philburk]

Testing on a Nexus 6 running SDK 6.0.1, burst = 192. Use format = PCM_I16 for input and output, analog headset jack.

kBufferQueueLength	RT latency	Capacity
2	50.0	384
3	49.5	576
4	49.7	768

Latency, and glitch protection, does not increase with the number of buffers. That was added in 7.0. ag/941250 SDK 6.0 has problems that we cannot fix in Oboe.

[philburk]

Testing on a Nexus 6 running SDK 7 MR2, burst = 192. Use format = PCM_I16 for input and output, analog headset jack.

kBufferQueueLength	RT latency	Capacity
2	53.3	384
3	57.5	576
4	61.5	768

Latency, and glitch protection, increases with the number of buffers.