From afc9d72c8968f4069f8b123fe3fb4ea94f79a3b9 Mon Sep 17 00:00:00 2001
From: Paul Ferrand <paul@ferrand.cc>
Date: Fri, 21 Aug 2020 11:17:44 +0200
Subject: [PATCH] Further changes to the envelope follower- Apply on a single
 channel- Track attack and release separately

---
 src/sfizz/Config.h          |  9 +++++----
 src/sfizz/Voice.cpp         | 32 +++++++++++++++++++++-----------
 src/sfizz/Voice.h           |  5 +++--
 src/sfizz/VoiceStealing.cpp |  2 +-
 4 files changed, 30 insertions(+), 18 deletions(-)

diff --git a/src/sfizz/Config.h b/src/sfizz/Config.h
index d3c301ba..fc48aab5 100644
--- a/src/sfizz/Config.h
+++ b/src/sfizz/Config.h
@@ -57,7 +57,8 @@ namespace config {
     constexpr Oversampling defaultOversamplingFactor { Oversampling::x1 };
     constexpr float A440 { 440.0 };
     constexpr size_t powerHistoryLength { 16 };
-    constexpr float powerFollowerFactor { 10 };
+    constexpr float powerFollowerAttackFactor { 100 };
+    constexpr float powerFollowerReleaseFactor { 10 };
     constexpr uint16_t numCCs { 512 };
     constexpr int maxCurves { 256 };
     constexpr int chunkSize { 1024 };
@@ -72,10 +73,10 @@ namespace config {
      */
     constexpr float stealingAgeCoeff { 0.5f };
     /**
-     * @brief The threshold for envelope stealing.
-     *        In percentage of the sum of all envelopes.
+     * @brief The threshold for power stealing.
+     *        In percentage of the sum of all powers.
      */
-    constexpr float stealingEnvelopeCoeff { 0.5f };
+    constexpr float stealingPowerCoeff { 0.5f };
     constexpr int filtersPerVoice { 2 };
     constexpr int eqsPerVoice { 3 };
     constexpr int oscillatorsPerVoice { 9 };
diff --git a/src/sfizz/Voice.cpp b/src/sfizz/Voice.cpp
index 8b1a0bed..54c5ebc2 100644
--- a/src/sfizz/Voice.cpp
+++ b/src/sfizz/Voice.cpp
@@ -249,13 +249,13 @@ void sfz::Voice::setSampleRate(float sampleRate) noexcept
     for (auto& lfo : lfos)
         lfo->setSampleRate(sampleRate);
 
-    trackingFactor = samplesPerBlock / sampleRate * config::powerFollowerFactor;
+    attackTrackingFactor =  config::powerFollowerAttackFactor / sampleRate;
+    releaseTrackingFactor =  config::powerFollowerReleaseFactor / sampleRate;
 }
 
 void sfz::Voice::setSamplesPerBlock(int samplesPerBlock) noexcept
 {
     this->samplesPerBlock = samplesPerBlock;
-    this->trackingFactor = samplesPerBlock / sampleRate * config::powerFollowerFactor;
 }
 
 void sfz::Voice::renderBlock(AudioSpan<float> buffer) noexcept
@@ -732,8 +732,7 @@ void sfz::Voice::reset() noexcept
     floatPositionOffset = 0.0f;
     noteIsOff = false;
 
-    for (auto& p : meanChannelPowers)
-        p = 0.0f;
+    meanChannelPower = 0.0f;
 
     filters.clear();
     equalizers.clear();
@@ -765,7 +764,7 @@ void sfz::Voice::removeVoiceFromRing() noexcept
 
 float sfz::Voice::getAveragePower() const noexcept
 {
-    return max(meanChannelPowers[0], meanChannelPowers[1]);
+    return meanChannelPower;
 }
 
 bool sfz::Voice::releasedOrFree() const noexcept
@@ -865,12 +864,23 @@ void sfz::Voice::updateChannelPowers(AudioSpan<float> buffer)
     if (buffer.getNumFrames() == 0)
         return;
 
-    const float factor = buffer.getNumFrames() / samplesPerBlock * trackingFactor;
-    for (unsigned i = 0; i < meanChannelPowers.size(); ++i) {
-        const auto input = buffer.getConstSpan(i);
-        const float meanPower = sfz::meanSquared(input);
-        meanChannelPowers[i] = meanChannelPowers[i] * (1 - factor) + meanPower * factor;
-    }
+    auto tempBuffer = resources.bufferPool.getBuffer(buffer.getNumFrames());
+    if (!tempBuffer)
+        return;
+
+    sfz::copy(buffer.getConstSpan(0), *tempBuffer);
+    for (unsigned i = 1; i < buffer.getNumChannels(); ++i)
+        sfz::add(buffer.getConstSpan(i), *tempBuffer);
+
+    const float meanPower = sfz::meanSquared<float>(*tempBuffer);
+
+    const float attackFactor = static_cast<float>(buffer.getNumFrames()) * attackTrackingFactor;
+    const float releaseFactor = static_cast<float>(buffer.getNumFrames()) * releaseTrackingFactor;
+
+    meanChannelPower = max(
+        meanChannelPower * (1 - attackFactor) + meanPower * attackFactor,
+        meanChannelPower * (1 - releaseFactor) + meanPower * releaseFactor
+    );
 }
 
 
diff --git a/src/sfizz/Voice.h b/src/sfizz/Voice.h
index facd3c1d..c2b3d1c4 100644
--- a/src/sfizz/Voice.h
+++ b/src/sfizz/Voice.h
@@ -485,8 +485,9 @@ private:
     Smoother xfadeSmoother;
     void resetSmoothers() noexcept;
 
-    float trackingFactor { config::defaultSamplesPerBlock / config::defaultSampleRate * config::powerFollowerFactor };
-    std::array<float, 2> meanChannelPowers;
+    float attackTrackingFactor { config::powerFollowerAttackFactor / config::defaultSampleRate };
+    float releaseTrackingFactor { config::powerFollowerReleaseFactor / config::defaultSampleRate };
+    float meanChannelPower;
     LEAK_DETECTOR(Voice);
 };
 
diff --git a/src/sfizz/VoiceStealing.cpp b/src/sfizz/VoiceStealing.cpp
index 0c935df5..0fa4f1a7 100644
--- a/src/sfizz/VoiceStealing.cpp
+++ b/src/sfizz/VoiceStealing.cpp
@@ -19,7 +19,7 @@ sfz::Voice* sfz::VoiceStealing::steal(absl::Span<sfz::Voice*> voices) noexcept
     // We are checking the power to try and kill voices with relative low contribution
     // to the output compared to the rest.
     const auto powerThreshold = sumPower
-        / static_cast<float>(voices.size()) * config::stealingEnvelopeCoeff;
+        / static_cast<float>(voices.size()) * config::stealingPowerCoeff;
     // We are checking the age so that voices have the time to build up attack
     // This is not perfect because pad-type voices will take a long time to output
     // their sound, but it's reasonable for sounds with a quick attack and longer