🤖 AI: Follow-up to ann0see/jamulus#293, opened at @ann0see's request.
What is the current behaviour and why should it be changed?
The gain loop in DecodeReceiveData calls GetFadeInGain() twice per channel pair, once for the source channel (line 909) and once for the target (line 915): 2N²−N int-to-float conversions and float divisions per frame, 44850 at N=150. The N values behind them are constant for the frame. Both counters are written only in PutAudioData and OnNetTranspPropsReceived, reached under CServer::Mutex from PutAudioData and OnProtocolMessageReceived, and OnTimer holds that mutex for the whole decode phase, single- or multithreaded.
Describe possible approaches
Read each connected channel's fade-in gain once into a CVector<float> in the loop that builds vecChanIDsCurConChan, and multiply by vecfFadeInGains[j] and vecfFadeInGains[iChanCnt] at lines 909 and 915 — #293's one-read-per-channel pattern applied to the remaining per-pair accessor, independent of that PR. The output is unchanged: a benchmark of the two loop shapes with a lock-free gain read produces byte-identical matrices at N=50, 100 and 150, and times the fade-in term alone (Raspberry Pi 4 Model B, g++ 14.2 -O2, median of 5 batches, 64-sample frame):
| N |
per pair |
per frame |
saving |
share of a 1.33 ms frame |
| 50 |
15.4 µs |
5.4 µs |
10.0 µs |
0.75% |
| 100 |
59.4 µs |
21.3 µs |
38.1 µs |
2.9% |
| 150 |
136.3 µs |
40.6 µs |
95.7 µs |
7.2% |
At the default 128-sample frame the microseconds are the same and the shares halve.
Has this feature been discussed and generally agreed?
Requested by @ann0see on #293; no PR until the design is agreed here.
🤖 This message was written by AI and reviewed by @mcfnord.
🤖 AI: Follow-up to ann0see/jamulus#293, opened at @ann0see's request.
What is the current behaviour and why should it be changed?
The gain loop in
DecodeReceiveDatacallsGetFadeInGain()twice per channel pair, once for the source channel (line 909) and once for the target (line 915): 2N²−N int-to-float conversions and float divisions per frame, 44850 at N=150. The N values behind them are constant for the frame. Both counters are written only inPutAudioDataandOnNetTranspPropsReceived, reached underCServer::MutexfromPutAudioDataandOnProtocolMessageReceived, andOnTimerholds that mutex for the whole decode phase, single- or multithreaded.Describe possible approaches
Read each connected channel's fade-in gain once into a
CVector<float>in the loop that buildsvecChanIDsCurConChan, and multiply byvecfFadeInGains[j]andvecfFadeInGains[iChanCnt]at lines 909 and 915 — #293's one-read-per-channel pattern applied to the remaining per-pair accessor, independent of that PR. The output is unchanged: a benchmark of the two loop shapes with a lock-free gain read produces byte-identical matrices at N=50, 100 and 150, and times the fade-in term alone (Raspberry Pi 4 Model B, g++ 14.2-O2, median of 5 batches, 64-sample frame):At the default 128-sample frame the microseconds are the same and the shares halve.
Has this feature been discussed and generally agreed?
Requested by @ann0see on #293; no PR until the design is agreed here.
🤖 This message was written by AI and reviewed by @mcfnord.