further convolving optimisaions

This commit is contained in:
minjaesong
2023-11-28 19:25:42 +09:00
parent af48a171eb
commit 6059ba92ac
2 changed files with 35 additions and 37 deletions

View File

@@ -49,6 +49,20 @@ private val IM1 = 0
*/
object FFT: Disposable {
private val ffts = hashMapOf(
128 to FloatFFT_1D(128),
256 to FloatFFT_1D(256),
512 to FloatFFT_1D(512),
1024 to FloatFFT_1D(1024),
2048 to FloatFFT_1D(2048),
4096 to FloatFFT_1D(4096),
8192 to FloatFFT_1D(8192),
16384 to FloatFFT_1D(16384),
32768 to FloatFFT_1D(32768),
65536 to FloatFFT_1D(65536),
)
init {
// Loader.load(org.bytedeco.fftw.global.fftw3::class.java)
@@ -102,8 +116,7 @@ object FFT: Disposable {
// USING JTRANSFORMS //
val signal = FloatArray(signal0.size * 2) { if (it % 2 == 0) signal0[it / 2] else 0f }
val fft = FloatFFT_1D(signal0.size.toLong())
fft.complexForward(signal)
ffts[signal0.size]!!.complexForward(signal)
return ComplexArray(signal)
}
@@ -137,12 +150,17 @@ object FFT: Disposable {
// USING JTRANSFORMS //
val signal = signal0.reim
val fft = FloatFFT_1D(signal0.size.toLong())
fft.complexInverse(signal, true)
ffts[signal0.size]!!.complexInverse(signal0.reim, true)
return signal0.getReal()
}
fun ifftAndGetReal(signal0: ComplexArray, output: FloatArray) {
ffts[signal0.size]!!.complexInverse(signal0.reim, true)
for (i in 0 until signal0.size) {
output[i] = signal0.reim[i * 2]
}
}
// org.apache.commons.math3.transform.FastFouriesTransformer.java:214
/**
* Computes the standard transform of the specified complex data. The

View File

@@ -2,6 +2,7 @@ package net.torvald.terrarum.audio.dsp
import com.jme3.math.FastMath
import net.torvald.terrarum.App.measureDebugTime
import net.torvald.terrarum.App.setDebugTime
import net.torvald.terrarum.audio.*
import net.torvald.terrarum.audio.TerrarumAudioMixerTrack.Companion.BUFFER_SIZE
import java.io.File
@@ -107,47 +108,26 @@ class Convolv(ir: File, val gain: Float = 1f / 256f): TerrarumAudioFilter() {
}
}
private val targetY = ComplexArray(FloatArray(fftLen * 2))
private val realtime = (BLOCKSIZE / TerrarumAudioMixerTrack.SAMPLING_RATEF * 1000000000L)
/**
* https://thewolfsound.com/fast-convolution-fft-based-overlap-add-overlap-save-partitioned/
*/
override fun thru(inbuf: List<FloatArray>, outbuf: List<FloatArray>) {
// println("Convolv thru")
val t1 = System.nanoTime()
for (ch in outbuf.indices) {
push(inbuf[ch].applyGain(gain), this.inbuf[ch])
lateinit var u: FloatArray
measureDebugTime("audio.convolve") {
val inputFFT = FFT.fft(this.inbuf[ch])
val Y = inputFFT * convFFT[ch]
val y = FFT.ifftAndGetReal(Y)
u = y.takeLast(BLOCKSIZE).toFloatArray()
}
// doesn't work AND slightly slower than the lines above
/*measureDebugTime("audio.convolve") {
// orthodox uneven-partitioning
fillUnevenly(this.inbuf[ch], inputPartd[ch])
val partY = inputPartd[ch].mapIndexed { i, inputSamples ->
FFT.fft(inputSamples) * convFFTpartd[ch][i]
}
concatParts(partY, targetY)
val y = FFT.ifftAndGetReal(targetY)
u = y.takeLast(BLOCKSIZE).toFloatArray()
}*/
System.arraycopy(u, 0, outbuf[ch], 0, BLOCKSIZE)
val inputFFT = FFT.fft(this.inbuf[ch])
val Y = inputFFT * convFFT[ch]
val y = FFT.ifftAndGetReal(Y)
System.arraycopy(y, fftLen - BLOCKSIZE, outbuf[ch], 0, BLOCKSIZE)
}
val t2 = System.nanoTime()
val ptime = (t2 - t1).toFloat()
val realtime = BLOCKSIZE / TerrarumAudioMixerTrack.SAMPLING_RATEF * 1000000000L
val ptime = System.nanoTime() - t1
setDebugTime("audio.convolve", ptime)
processingSpeed = realtime / ptime
}