feat(tracker): panic synth if Inf or NaN, and handle these in detectors

Closes #210.

CHANGELOG.md

@@ -3,8 +3,12 @@ All notable changes to this project will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
 
-## [0.5.0]
+## [Unreleased]
+### Added
+- Panic the synth if it outputs NaN or Inf, and handle these more gracefully in
+  the loudness and peak detector. ([#210][i210])
 
+## [0.5.0]
 ### BREAKING CHANGES
 - BREAKING CHANGE: always first modulate delay time, then apply notetracking. In
   a delay unit, modulation adds to the delay time, while note tracking
@@ -368,3 +372,4 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
 [i192]: https://github.com/vsariola/sointu/issues/192
 [i196]: https://github.com/vsariola/sointu/issues/196
 [i200]: https://github.com/vsariola/sointu/issues/200
+[i210]: https://github.com/vsariola/sointu/issues/210

tracker/detector.go

@@ -74,6 +74,11 @@ const (
 )
 
 const MAX_INTEGRATED_DATA = 10 * 60 * 60 // 1 hour of samples at 10 Hz (100 ms per sample)
+// In the detector, we clamp the signal levels to +-MAX_SIGNAL_AMPLITUDE to
+// avoid Inf results. This is 240 dBFS. max float32 is about 3.4e38, so squaring
+// the amplitude values gives 1e24, and adding 4410 of those together (when
+// taking the mean) gives a value < 1e37, which is still < max float32.
+const MAX_SIGNAL_AMPLITUDE = 1e12
 
 const (
 	PeakMomentary PeakType = iota
@@ -232,7 +237,7 @@ func (d *loudnessDetector) update(chunk sointu.AudioBuffer) LoudnessResult {
 	for chn := range 2 {
 		// deinterleave the channels
 		for i := range chunk {
-			d.tmp[i] = chunk[i][chn]
+			d.tmp[i] = removeNaNsAndClamp(chunk[i][chn])
 		}
 		// filter the signal with the weighting filter
 		for k := range d.weighting {
@@ -287,6 +292,13 @@ func (d *loudnessDetector) reset() {
 	d.integratedPower = 0
 }
 
+func removeNaNsAndClamp(s float32) float32 {
+	if s != s { // NaN
+		return 0
+	}
+	return min(max(s, -MAX_SIGNAL_AMPLITUDE), MAX_SIGNAL_AMPLITUDE)
+}
+
 func powerToDecibel(power float32) Decibel {
 	return Decibel(float32(10 * math.Log10(float64(power))))
 }
@@ -382,7 +394,7 @@ func (d *peakDetector) update(buf sointu.AudioBuffer) (ret PeakResult) {
 	for chn := range 2 {
 		// deinterleave the channels
 		for i := range buf {
-			d.tmp[i] = buf[i][chn]
+			d.tmp[i] = removeNaNsAndClamp(buf[i][chn])
 		}
 		// 4x oversample the signal
 		var o []float32

tracker/player.go

@@ -128,7 +128,12 @@ func (p *Player) Process(buffer sointu.AudioBuffer, context PlayerProcessContext
 			rendered, timeAdvanced, err = p.synth.Render(buffer[:framesUntilEvent], timeUntilRowAdvance)
 			if err != nil {
 				p.synth = nil
-				p.send(Alert{Message: fmt.Sprintf("synth.Render: %s", err.Error()), Priority: Error, Name: "PlayerCrash"})
+				p.send(Alert{Message: fmt.Sprintf("synth.Render: %s", err.Error()), Priority: Error, Name: "PlayerCrash", Duration: defaultAlertDuration})
+			}
+			// for performance, we don't check for NaN of every sample, because typically NaNs propagate
+			if rendered > 0 && (isNaN(buffer[0][0]) || isNaN(buffer[0][1]) || isInf(buffer[0][0]) || isInf(buffer[0][1])) {
+				p.synth = nil
+				p.send(Alert{Message: "Inf or NaN detected in synth output", Priority: Error, Name: "PlayerCrash", Duration: defaultAlertDuration})
 			}
 		} else {
 			rendered = min(framesUntilEvent, timeUntilRowAdvance)
@@ -206,6 +211,14 @@ func (p NullPlayerProcessContext) BPM() (bpm float64, ok bool) {
 	return 0, false // no BPM available
 }
 
+func isNaN(f float32) bool {
+	return f != f
+}
+
+func isInf(f float32) bool {
+	return f > math.MaxFloat32 || f < -math.MaxFloat32
+}
+
 func (p *Player) processMessages(context PlayerProcessContext) {
 loop:
 	for { // process new message