Optimize CELT inverse transform

decoder.go

@@ -42,6 +42,8 @@ type Decoder struct {
 	silkResamplerChannels  int
 	hybridSilkResampler    [2]silkresample.Resampler
 	hybridSilkChannels     int
+	hybridSilkBuffer       []float32
+	hybridSilkPCM          []float32
 	silkRedundancyFades    []silkRedundancyFade
 	silkCeltAdditions      []silkCeltAddition
 	floatBuffer            []float32
@@ -755,7 +757,7 @@ func (d *Decoder) decodeHybridFrame(
 	d.rangeDecoder.Init(encodedFrame)
 
 	silkOutputChannelCount := min(streamChannelCount, outputChannelCount)
-	silkInternal := make([]float32, silkSamplesPerChannel*silkOutputChannelCount)
+	silkInternal := resizeFloat32Buffer(&d.hybridSilkBuffer, silkSamplesPerChannel*silkOutputChannelCount)
 	if err := d.silkDecoder.DecodeWithRangeToChannels(
 		&d.rangeDecoder,
 		silkInternal,
@@ -792,7 +794,7 @@ func (d *Decoder) decodeHybridFrame(
 		return err
 	}
 
-	silkPCM := make([]float32, outputFrameSampleCount*silkOutputChannelCount)
+	silkPCM := resizeFloat32Buffer(&d.hybridSilkPCM, outputFrameSampleCount*silkOutputChannelCount)
 	if err = d.resampleHybridSilk(silkInternal, silkPCM, silkOutputChannelCount); err != nil {
 		return err
 	}
@@ -1363,6 +1365,15 @@ func float32ToInt16(in []float32, out []int16, sampleCount int) {
 	}
 }
 
+func resizeFloat32Buffer(buffer *[]float32, sampleCount int) []float32 {
+	if cap(*buffer) < sampleCount {
+		*buffer = make([]float32, sampleCount)
+	}
+	*buffer = (*buffer)[:sampleCount]
+
+	return *buffer
+}
+
 // Decode decodes the Opus bitstream into S16LE PCM.
 func (d *Decoder) Decode(in, out []byte) (bandwidth Bandwidth, isStereo bool, err error) {
 	if cap(d.floatBuffer) < len(out)/2 {

internal/celt/celt.go

@@ -10,6 +10,7 @@ const (
 	sampleRate            = 48000
 	shortBlockSampleCount = 120
 	maxLM                 = 3
+	maxFrameSampleCount   = shortBlockSampleCount << maxLM
 	maxBands              = 21
 	hybridStartBand       = 17
 )

internal/celt/decoder.go

@@ -19,6 +19,7 @@ type Decoder struct {
 	preemphasisMem [2]float32
 	rng            uint32
 	lossCount      int
+	scratch        *decoderScratch
 }
 
 // NewDecoder creates a CELT decoder with the static Opus 48 kHz mode.
@@ -149,6 +150,7 @@ func (d *Decoder) decode(
 	outputSampleRate int,
 	rangeDecoder *rangecoding.Decoder,
 ) error {
+	scratch := d.scratchBuffer()
 	channelCount := 1
 	if isStereo {
 		channelCount = 2
@@ -189,10 +191,12 @@ func (d *Decoder) decode(
 		return err
 	}
 	if info.silence {
-		x := make([]float32, frameSampleCount)
+		x := scratch.x[:frameSampleCount]
+		clear(x)
 		var y []float32
 		if isStereo {
-			y = make([]float32, frameSampleCount)
+			y = scratch.y[:frameSampleCount]
+			clear(y)
 		}
 		for channel := range info.channelCount {
 			for band := info.startBand; band < info.endBand; band++ {
@@ -213,10 +217,12 @@ func (d *Decoder) decode(
 
 	// RFC 6716 Sections 4.3.4 through 4.3.7 decode the normalized residual,
 	// optionally repair collapsed transient blocks, then synthesize PCM.
-	x := make([]float32, frameSampleCount)
+	x := scratch.x[:frameSampleCount]
+	clear(x)
 	var y []float32
 	if isStereo {
-		y = make([]float32, frameSampleCount)
+		y = scratch.y[:frameSampleCount]
+		clear(y)
 	}
 	state := bandDecodeState{
 		rangeDecoder: &d.rangeDecoder,
@@ -284,3 +290,11 @@ func (d *Decoder) Mode() *Mode {
 func (d *Decoder) FinalRange() uint32 {
 	return d.rangeDecoder.FinalRange()
 }
+
+func (d *Decoder) scratchBuffer() *decoderScratch {
+	if d.scratch == nil {
+		d.scratch = &decoderScratch{}
+	}
+
+	return d.scratch
+}