Update the output limiter/compressor

This provides better characteristics for an amplitude limiter. In particular,
it utilizes the peak amplitude instead of the RMS, and the used parameters
basically guarantee no output samples exceed the given threshold... almost, due
to floating-point errors as the threshold is converted from dB to log-e for the
envelope, then is negated and converted to linear amplitude to apply to the
signal. It's quite possible for some rounding errors to creep in and not
perfectly saturate the result.

4 files changed, 431 insertions, 168 deletions

diff --git a/Alc/ALc.c b/Alc/ALc.c
index 9a1051c9..ecacf8c5 100644
--- a/Alc/ALc.c
+++ b/Alc/ALc.c
@@ -1706,8 +1706,9 @@ static void alcSetError(ALCdevice *device, ALCenum errorCode)
 
 struct Compressor *CreateDeviceLimiter(const ALCdevice *device)
 {
-    return CompressorInit(0.0f, 0.0f, AL_FALSE, AL_TRUE, 0.0f, 0.0f, 0.5f, 2.0f,
-                          0.0f, -3.0f, 3.0f, device->Frequency);
+    return CompressorInit(device->RealOut.NumChannels, device->Frequency,
+        AL_TRUE, AL_TRUE, AL_TRUE, AL_TRUE, AL_TRUE, 0.001f, 0.002f,
+        0.0f, 0.0f, -0.0003f, INFINITY, 0.0f, 0.020f, 0.200f);
 }
 
 /* UpdateClockBase
@@ -2231,7 +2232,8 @@ static ALCenum UpdateDeviceParams(ALCdevice *device, const ALCint *attrList)
      */
     if(gainLimiter != ALC_FALSE)
     {
-        if(!device->Limiter || device->Frequency != GetCompressorSampleRate(device->Limiter))
+        if(!device->Limiter || device->Frequency != GetCompressorSampleRate(device->Limiter) ||
+           device->RealOut.NumChannels != GetCompressorChannelCount(device->Limiter))
         {
             al_free(device->Limiter);
             device->Limiter = CreateDeviceLimiter(device);
diff --git a/Alc/ALu.c b/Alc/ALu.c
index e9d8aa02..03abb116 100644
--- a/Alc/ALu.c
+++ b/Alc/ALu.c
@@ -1838,8 +1838,7 @@ void aluMixData(ALCdevice *device, ALvoid *OutBuffer, ALsizei NumSamples)
                           SamplesToDo, device->RealOut.NumChannels);
 
         if(device->Limiter)
-            ApplyCompression(device->Limiter, device->RealOut.NumChannels, SamplesToDo,
-                             device->RealOut.Buffer);
+            ApplyCompression(device->Limiter, SamplesToDo, device->RealOut.Buffer);
 
         if(device->DitherDepth > 0.0f)
             ApplyDither(device->RealOut.Buffer, &device->DitherSeed, device->DitherDepth,
diff --git a/Alc/mastering.c b/Alc/mastering.c
index 1636c8d9..78f6038d 100644
--- a/Alc/mastering.c
+++ b/Alc/mastering.c
@@ -7,226 +7,439 @@
 #include "almalloc.h"
 
 
+extern inline ALsizei GetCompressorChannelCount(const Compressor *Comp);
 extern inline ALuint GetCompressorSampleRate(const Compressor *Comp);
 
-#define RMS_WINDOW_SIZE (1<<7)
-#define RMS_WINDOW_MASK (RMS_WINDOW_SIZE-1)
-#define RMS_VALUE_MAX  (1<<24)
 
-static_assert(RMS_VALUE_MAX < (UINT_MAX / RMS_WINDOW_SIZE), "RMS_VALUE_MAX is too big");
-
-
-/* Multichannel compression is linked via one of two modes:
+/* This sliding hold follows the input level with an instant attack and a
+ * fixed duration hold before an instant release to the next highest level.
+ * It is a sliding window maximum (descending maxima) implementation based on
+ * Richard Harter's ascending minima algorithm available at:
  *
- *   Summed - Absolute sum of all channels.
- *   Maxed  - Absolute maximum of any channel.
+ *   http://www.richardhartersworld.com/cri/2001/slidingmin.html
  */
-static void SumChannels(Compressor *Comp, const ALsizei NumChans, const ALsizei SamplesToDo,
-                        ALfloat (*restrict OutBuffer)[BUFFERSIZE])
+static ALfloat UpdateSlidingHold(SlidingHold *Hold, const ALsizei i, const ALfloat in)
 {
-    ALsizei c, i;
+    const ALsizei mask = BUFFERSIZE - 1;
+    const ALsizei length = Hold->Length;
+    ALfloat *restrict values = Hold->Values;
+    ALsizei *restrict expiries = Hold->Expiries;
+    ALsizei lowerIndex = Hold->LowerIndex;
+    ALsizei upperIndex = Hold->UpperIndex;
 
-    for(i = 0;i < SamplesToDo;i++)
-        Comp->Envelope[i] = 0.0f;
+    if(i >= expiries[upperIndex])
+        upperIndex = (upperIndex + 1) & mask;
 
-    for(c = 0;c < NumChans;c++)
+    if(in >= values[upperIndex])
     {
-        for(i = 0;i < SamplesToDo;i++)
-            Comp->Envelope[i] += OutBuffer[c][i];
+        values[upperIndex] = in;
+        expiries[upperIndex] = i + length;
+        lowerIndex = upperIndex;
     }
+    else
+    {
+        while(in >= values[lowerIndex])
+            lowerIndex = (lowerIndex - 1) & mask;
 
-    for(i = 0;i < SamplesToDo;i++)
-        Comp->Envelope[i] = fabsf(Comp->Envelope[i]);
+        lowerIndex = (lowerIndex + 1) & mask;
+        values[lowerIndex] = in;
+        expiries[lowerIndex] = i + length;
+    }
+
+    Hold->LowerIndex = lowerIndex;
+    Hold->UpperIndex = upperIndex;
+
+    return values[upperIndex];
 }
 
-static void MaxChannels(Compressor *Comp, const ALsizei NumChans, const ALsizei SamplesToDo,
-                        ALfloat (*restrict OutBuffer)[BUFFERSIZE])
+static void ShiftSlidingHold(SlidingHold *Hold, const ALsizei n)
 {
+    const ALsizei mask = BUFFERSIZE - 1;
+    const ALsizei lowerIndex = Hold->LowerIndex;
+    ALsizei *restrict expiries = Hold->Expiries;
+    ALsizei i = Hold->UpperIndex;
+
+    while(i != lowerIndex)
+    {
+        expiries[i] -= n;
+        i = (i + 1) & mask;
+    }
+
+    expiries[i] -= n;
+}
+
+/* Multichannel compression is linked via the absolute maximum of all
+ * channels.
+ */
+static void LinkChannels(Compressor *Comp, const ALsizei SamplesToDo, ALfloat (*restrict OutBuffer)[BUFFERSIZE])
+{
+    const ALsizei mask = 2*BUFFERSIZE - 1;
+    const ALsizei index = Comp->SideChainIndex + Comp->LookAhead;
+    const ALsizei numChans = Comp->NumChans;
+    ALfloat *restrict sideChain = Comp->SideChain;
     ALsizei c, i;
 
     for(i = 0;i < SamplesToDo;i++)
-        Comp->Envelope[i] = 0.0f;
+        sideChain[(index + i) & mask] = 0.0f;
 
-    for(c = 0;c < NumChans;c++)
+    for(c = 0;c < numChans;c++)
     {
         for(i = 0;i < SamplesToDo;i++)
-            Comp->Envelope[i] = maxf(Comp->Envelope[i], fabsf(OutBuffer[c][i]));
+        {
+            ALsizei offset = (index + i) & mask;
+
+            sideChain[offset] = maxf(sideChain[offset], fabsf(OutBuffer[c][i]));
+        }
     }
 }
 
-/* Envelope detection/sensing can be done via:
- *
- *   RMS  - Rectangular windowed root mean square of linking stage.
- *   Peak - Implicit output from linking stage.
+/* This calculates the squared crest factor of the control signal for the
+ * basic automation of the attack/release times.  As suggested by the paper,
+ * it uses an instantaneous squared peak detector and a squared RMS detector
+ * both with 200ms release times.
  */
-static void RmsDetection(Compressor *Comp, const ALsizei SamplesToDo)
+static void CrestDetector(Compressor *Comp, const ALsizei SamplesToDo)
 {
-    ALuint sum = Comp->RmsSum;
-    ALuint *window = Comp->RmsWindow;
-    ALsizei index = Comp->RmsIndex;
+    const ALsizei mask = 2*BUFFERSIZE - 1;
+    const ALfloat a_crest = Comp->CrestCoeff;
+    const ALsizei index = Comp->SideChainIndex + Comp->LookAhead;
+    const ALfloat *restrict sideChain = Comp->SideChain;
+    ALfloat *restrict crestFactor = Comp->CrestFactor;
+    ALfloat y2_peak = Comp->LastPeakSq;
+    ALfloat y2_rms = Comp->LastRmsSq;
     ALsizei i;
 
     for(i = 0;i < SamplesToDo;i++)
     {
-        ALfloat sig = Comp->Envelope[i];
-
-        sum -= window[index];
-        window[index] = fastf2i(minf(sig * sig * 65536.0f, RMS_VALUE_MAX));
-        sum += window[index];
-        index = (index + 1) & RMS_WINDOW_MASK;
+        ALfloat x_abs = sideChain[(index + i) & mask];
+        ALfloat x2 = maxf(0.000001f, x_abs * x_abs);
 
-        Comp->Envelope[i] = sqrtf(sum / 65536.0f / RMS_WINDOW_SIZE);
+        y2_peak = maxf(x2, lerp(x2, y2_peak, a_crest));
+        y2_rms = lerp(x2, y2_rms, a_crest);
+        crestFactor[i] = y2_peak / y2_rms;
     }
 
-    Comp->RmsSum = sum;
-    Comp->RmsIndex = index;
+    Comp->LastPeakSq = y2_peak;
+    Comp->LastRmsSq = y2_rms;
 }
 
-/* This isn't a very sophisticated envelope follower, but it gets the job
- * done.  First, it operates at logarithmic scales to keep transitions
- * appropriate for human hearing.  Second, it can apply adaptive (automated)
- * attack/release adjustments based on the signal.
+/* The side-chain starts with a simple peak detector (based on the absolute
+ * value of the incoming signal) and performs most of its operations in the
+ * log domain.
  */
-static void FollowEnvelope(Compressor *Comp, const ALsizei SamplesToDo)
+static void PeakDetector(Compressor *Comp, const ALsizei SamplesToDo)
 {
-    ALfloat attackMin = Comp->AttackMin;
-    ALfloat attackMax = Comp->AttackMax;
-    ALfloat releaseMin = Comp->ReleaseMin;
-    ALfloat releaseMax = Comp->ReleaseMax;
-    ALfloat last = Comp->EnvLast;
+    const ALsizei mask = 2*BUFFERSIZE - 1;
+    const ALsizei index = Comp->SideChainIndex + Comp->LookAhead;
+    ALfloat *restrict sideChain = Comp->SideChain;
     ALsizei i;
 
     for(i = 0;i < SamplesToDo;i++)
     {
-        ALfloat env = log10f(maxf(Comp->Envelope[i], 0.000001f));
-        ALfloat slope = minf(1.0f, fabsf(env - last) / 4.5f);
+        ALuint offset = (index + i) & mask;
+        ALfloat x_abs = sideChain[offset];
 
-        if(env > last)
-            last = minf(env, last + lerp(attackMin, attackMax, 1.0f - (slope * slope)));
-        else
-            last = maxf(env, last + lerp(releaseMin, releaseMax, 1.0f - (slope * slope)));
+        sideChain[offset] = logf(maxf(0.000001f, x_abs));
+    }
+}
 
-        Comp->Envelope[i] = last;
+/* An optional hold can be used to extend the peak detector so it can more
+ * solidly detect fast transients.  This is best used when operating as a
+ * limiter.
+ */
+static void PeakHoldDetector(Compressor *Comp, const ALsizei SamplesToDo)
+{
+    const ALsizei mask = 2*BUFFERSIZE - 1;
+    const ALsizei index = Comp->SideChainIndex + Comp->LookAhead;
+    ALfloat *restrict sideChain = Comp->SideChain;
+    SlidingHold *hold = Comp->Hold;
+    ALsizei i;
+
+    for(i = 0;i < SamplesToDo;i++)
+    {
+        ALsizei offset = (index + i) & mask;
+        ALfloat x_abs = sideChain[offset];
+        ALfloat x_G = logf(maxf(0.000001f, x_abs));
+
+        sideChain[offset] = UpdateSlidingHold(hold, i, x_G);
     }
 
-    Comp->EnvLast = last;
+    ShiftSlidingHold(hold, SamplesToDo);
 }
 
-/* The envelope is converted to control gain with an optional soft knee. */
-static void EnvelopeGain(Compressor *Comp, const ALsizei SamplesToDo, const ALfloat Slope)
+/* This is the heart of the feed-forward compressor.  It operates in the log
+ * domain (to better match human hearing) and can apply some basic automation
+ * to knee width, attack/release times, make-up/post gain, and clipping
+ * reduction.
+ */
+static void GainCompressor(Compressor *Comp, const ALsizei SamplesToDo)
 {
+    const ALsizei mask = 2*BUFFERSIZE - 1;
+    const bool autoKnee = Comp->Auto.Knee;
+    const bool autoAttack = Comp->Auto.Attack;
+    const bool autoRelease = Comp->Auto.Release;
+    const bool autoPostGain = Comp->Auto.PostGain;
+    const bool autoDeclip = Comp->Auto.Declip;
+    const ALsizei lookAhead = Comp->LookAhead;
     const ALfloat threshold = Comp->Threshold;
-    const ALfloat knee = Comp->Knee;
+    const ALfloat slope = Comp->Slope;
+    const ALfloat attack = Comp->Attack;
+    const ALfloat release = Comp->Release;
+    const ALsizei index = Comp->SideChainIndex;
+    const ALfloat *restrict crestFactor = Comp->CrestFactor;
+    ALfloat *restrict sideChain = Comp->SideChain;
+    ALfloat postGain = Comp->PostGain;
+    ALfloat knee = Comp->Knee;
+    ALfloat c_est = Comp->GainEstimate;
+    ALfloat a_adp = Comp->AdaptCoeff;
+    ALfloat t_att = attack;
+    ALfloat t_rel = release - attack;
+    ALfloat a_att = expf(-1.0f / t_att);
+    ALfloat a_rel = expf(-1.0f / t_rel);
+    ALfloat y_1 = Comp->LastRelease;
+    ALfloat y_L = Comp->LastAttack;
+    ALfloat c_dev = Comp->LastGainDev;
     ALsizei i;
 
-    if(!(knee > 0.0f))
+    for(i = 0;i < SamplesToDo;i++)
     {
-        for(i = 0;i < SamplesToDo;i++)
+        const ALfloat y2_crest = crestFactor[i];
+        ALfloat x_G = sideChain[(index + lookAhead + i) & mask];
+        ALfloat x_over = x_G - threshold;
+        ALfloat knee_h;
+        ALfloat y_G;
+        ALfloat x_L;
+
+        if(autoKnee)
+            knee = maxf(0.0f, 2.5f * (c_dev + c_est));
+        knee_h = 0.5f * knee;
+
+        /* This is the gain computer.  It applies a static compression curve
+         * to the control signal.
+         */
+        if(x_over <= -knee_h)
+            y_G = 0.0f;
+        else if(fabsf(x_over) < knee_h)
+            y_G = (x_over + knee_h) * (x_over + knee_h) / (2.0f * knee);
+        else
+            y_G = x_over;
+
+        x_L = -slope * y_G;
+
+        if(autoAttack)
+        {
+            t_att = 2.0f * attack / y2_crest;
+            a_att = expf(-1.0f / t_att);
+        }
+
+        if(autoRelease)
         {
-            ALfloat gain = Slope * (threshold - Comp->Envelope[i]);
-            Comp->Envelope[i] = powf(10.0f, minf(0.0f, gain));
+            t_rel = 2.0f * release / y2_crest - t_att;
+            a_rel = expf(-1.0f / t_rel);
         }
+
+        /* Gain smoothing (ballistics) is done via a smooth decoupled peak
+         * detector.  The attack time is subtracted from the release time
+         * above to compensate for the chained operating mode.
+         */
+        y_1 = maxf(x_L, lerp(x_L, y_1, a_rel));
+        y_L = lerp(y_1, y_L, a_att);
+
+        /* Knee width and make-up gain automation make use of a smoothed
+         * measurement of deviation between the control signal and estimate.
+         * The estimate is also used to bias the measurement to hot-start its
+         * average.
+         */
+        c_dev = lerp(-y_L - c_est, c_dev, a_adp);
+
+        if(autoPostGain)
+        {
+            /* Clipping reduction is only viable when make-up gain is being
+             * automated.  It modifies the deviation to further attenuate the
+             * control signal when clipping is detected.  The adaptation
+             * time is sufficiently long enough to suppress further clipping
+             * at the same output level.
+             */
+            if(autoDeclip)
+            {
+                x_G = sideChain[(index + i) & mask];
+                if((x_G - c_dev - c_est - y_L) > threshold)
+                    c_dev = x_G - c_est - y_L - threshold;
+            }
+
+            postGain = -(c_dev + c_est);
+        }
+
+        sideChain[(index + i) & mask] = expf(postGain - y_L);
     }
-    else
-    {
-        const ALfloat lower = threshold - (0.5f * knee);
-        const ALfloat upper = threshold + (0.5f * knee);
-        const ALfloat m = 0.5f * Slope / knee;
 
+    Comp->LastRelease = y_1;
+    Comp->LastAttack = y_L;
+    Comp->LastGainDev = c_dev;
+}
+
+/* Combined with the hold time, a look-ahead delay can improve handling of
+ * fast transients by allowing the envelope time to converge prior to
+ * reaching the offending impulse.  This is best used when operating as a
+ * limiter.
+ */
+static void SignalDelay(Compressor *Comp, const ALsizei SamplesToDo, ALfloat (*restrict OutBuffer)[BUFFERSIZE])
+{
+    const ALsizei mask = BUFFERSIZE - 1;
+    const ALsizei numChans = Comp->NumChans;
+    const ALsizei indexIn = Comp->DelayIndex;
+    const ALsizei indexOut = Comp->DelayIndex - Comp->LookAhead;
+    ALfloat (*restrict delay)[BUFFERSIZE] = Comp->Delay;
+    ALsizei c, i;
+
+    for(c = 0;c < numChans;c++)
+    {
         for(i = 0;i < SamplesToDo;i++)
         {
-            ALfloat env = Comp->Envelope[i];
-            ALfloat gain;
-
-            if(env > lower && env < upper)
-                gain = m * (env - lower) * (lower - env);
-            else
-                gain = Slope * (threshold - env);
+            ALfloat sig = OutBuffer[c][i];
 
-            Comp->Envelope[i] = powf(10.0f, minf(0.0f, gain));
+            OutBuffer[c][i] = delay[c][(indexOut + i) & mask];
+            delay[c][(indexIn + i) & mask] = sig;
         }
     }
-}
 
+    Comp->DelayIndex = (indexIn + SamplesToDo) & mask;
+}
 
-Compressor *CompressorInit(const ALfloat PreGainDb, const ALfloat PostGainDb,
-                           const ALboolean SummedLink, const ALboolean RmsSensing,
-                           const ALfloat AttackTimeMin, const ALfloat AttackTimeMax,
-                           const ALfloat ReleaseTimeMin, const ALfloat ReleaseTimeMax,
-                           const ALfloat Ratio, const ALfloat ThresholdDb,
-                           const ALfloat KneeDb, const ALuint SampleRate)
+/* The compressor is initialized with the following settings:
+ *
+ *   NumChans       - Number of channels to process.
+ *   SampleRate     - Sample rate to process.
+ *   AutoKnee       - Whether to automate the knee width parameter.
+ *   AutoAttack     - Whether to automate the attack time parameter.
+ *   AutoRelease    - Whether to automate the release time parameter.
+ *   AutoPostGain   - Whether to automate the make-up (post) gain parameter.
+ *   AutoDeclip     - Whether to automate clipping reduction.  Ignored when
+ *                    not automating make-up gain.
+ *   LookAheadTime  - Look-ahead time (in seconds).
+ *   HoldTime       - Peak hold-time (in seconds).
+ *   PreGainDb      - Gain applied before detection (in dB).
+ *   PostGainDb     - Make-up gain applied after compression (in dB).
+ *   ThresholdDb    - Triggering threshold (in dB).
+ *   Ratio          - Compression ratio (x:1).  Set to INFINITY for true
+ *                    limiting.  Ignored when automating knee width.
+ *   KneeDb         - Knee width (in dB).  Ignored when automating knee
+ *                    width.
+ *   AttackTimeMin  - Attack time (in seconds).  Acts as a maximum when
+ *                    automating attack time.
+ *   ReleaseTimeMin - Release time (in seconds).  Acts as a maximum when
+ *                    automating release time.
+ */
+Compressor* CompressorInit(const ALuint NumChans, const ALuint SampleRate,
+                           const ALboolean AutoKnee, const ALboolean AutoAttack,
+                           const ALboolean AutoRelease, const ALboolean AutoPostGain,
+                           const ALboolean AutoDeclip, const ALfloat LookAheadTime,
+                           const ALfloat HoldTime, const ALfloat PreGainDb,
+                           const ALfloat PostGainDb, const ALfloat ThresholdDb,
+                           const ALfloat Ratio, const ALfloat KneeDb,
+                           const ALfloat AttackTime, const ALfloat ReleaseTime)
 {
     Compressor *Comp;
+    ALsizei lookAhead;
+    ALsizei hold;
     size_t size;
-    ALsizei i;
 
+    lookAhead = (ALsizei)minf(BUFFERSIZE, roundf(maxf(0.0f, LookAheadTime) * SampleRate));
+    hold = (ALsizei)minf(BUFFERSIZE, roundf(maxf(0.0f, HoldTime) * SampleRate));
     size = sizeof(*Comp);
-    if(RmsSensing)
-        size += sizeof(Comp->RmsWindow[0]) * RMS_WINDOW_SIZE;
-    Comp = al_calloc(16, size);
 
-    Comp->PreGain = powf(10.0f, PreGainDb / 20.0f);
-    Comp->PostGain = powf(10.0f, PostGainDb / 20.0f);
-    Comp->SummedLink = SummedLink;
-    Comp->AttackMin = 1.0f / maxf(0.000001f, AttackTimeMin * SampleRate * logf(10.0f));
-    Comp->AttackMax = 1.0f / maxf(0.000001f, AttackTimeMax * SampleRate * logf(10.0f));
-    Comp->ReleaseMin = -1.0f / maxf(0.000001f, ReleaseTimeMin * SampleRate * logf(10.0f));
-    Comp->ReleaseMax = -1.0f / maxf(0.000001f, ReleaseTimeMax * SampleRate * logf(10.0f));
-    Comp->Ratio = Ratio;
-    Comp->Threshold = ThresholdDb / 20.0f;
-    Comp->Knee = maxf(0.0f, KneeDb / 20.0f);
-    Comp->SampleRate = SampleRate;
+    if(lookAhead > 0)
+    {
+        size += sizeof(*Comp->Delay) * NumChans;
+        if(hold > 0)
+            size += sizeof(*Comp->Hold);
+    }
 
-    Comp->RmsSum = 0;
-    if(RmsSensing)
-        Comp->RmsWindow = (ALuint*)(Comp+1);
-    else
-        Comp->RmsWindow = NULL;
-    Comp->RmsIndex = 0;
+    Comp = al_calloc(16, size);
+    Comp->NumChans = NumChans;
+    Comp->SampleRate = SampleRate;
+    Comp->Auto.Knee = AutoKnee;
+    Comp->Auto.Attack = AutoAttack;
+    Comp->Auto.Release = AutoRelease;
+    Comp->Auto.PostGain = AutoPostGain;
+    Comp->Auto.Declip = AutoPostGain && AutoDeclip;
+    Comp->LookAhead = lookAhead;
+    Comp->PreGain = powf(10.0f, PreGainDb / 20.0f);
+    Comp->PostGain = PostGainDb * logf(10.0f) / 20.0f;
+    Comp->Threshold = ThresholdDb * logf(10.0f) / 20.0f;
+    Comp->Slope = 1.0f / maxf(1.0f, Ratio) - 1.0f;
+    Comp->Knee = maxf(0.0f, KneeDb * logf(10.0f) / 20.0f);
+    Comp->Attack = maxf(1.0f, AttackTime * SampleRate);
+    Comp->Release = maxf(1.0f, ReleaseTime * SampleRate);
+
+    /* Knee width automation actually treats the compressor as a limiter.  By
+     * varying the knee width, it can effectively be seen as applying
+     * compression over a wide range of ratios.
+     */
+    if(AutoKnee)
+        Comp->Slope = -1.0f;
+
+    if(lookAhead > 0)
+    {
+        if(hold > 0)
+        {
+            Comp->Hold = (SlidingHold*)(Comp + 1);
+            Comp->Hold->Values[0] = -INFINITY;
+            Comp->Hold->Expiries[0] = hold;
+            Comp->Hold->Length = hold;
+            Comp->Delay = (ALfloat(*)[])(Comp->Hold + 1);
+        }
+        else
+        {
+            Comp->Delay = (ALfloat(*)[])(Comp + 1);
+        }
+    }
 
-    for(i = 0;i < BUFFERSIZE;i++)
-        Comp->Envelope[i] = 0.0f;
-    Comp->EnvLast = -6.0f;
+    Comp->CrestCoeff = expf(-1.0f / (0.200f * SampleRate)); // 200ms
+    Comp->GainEstimate = Comp->Threshold * -0.5f * Comp->Slope;
+    Comp->AdaptCoeff = expf(-1.0f / (2.0f * SampleRate)); // 2s
 
     return Comp;
 }
 
-void ApplyCompression(Compressor *Comp, const ALsizei NumChans, const ALsizei SamplesToDo,
-                      ALfloat (*restrict OutBuffer)[BUFFERSIZE])
+void ApplyCompression(Compressor *Comp, const ALsizei SamplesToDo, ALfloat (*restrict OutBuffer)[BUFFERSIZE])
 {
+    const ALsizei mask = 2*BUFFERSIZE - 1;
+    const ALsizei numChans = Comp->NumChans;
+    const ALfloat preGain = Comp->PreGain;
+    const ALsizei index = Comp->SideChainIndex;
+    ALfloat *restrict sideChain = Comp->SideChain;
     ALsizei c, i;
 
-    if(Comp->PreGain != 1.0f)
+    if(preGain != 1.0f)
     {
-        for(c = 0;c < NumChans;c++)
+        for(c = 0;c < numChans;c++)
         {
             for(i = 0;i < SamplesToDo;i++)
-                OutBuffer[c][i] *= Comp->PreGain;
+                OutBuffer[c][i] *= preGain;
         }
     }
 
-    if(Comp->SummedLink)
-        SumChannels(Comp, NumChans, SamplesToDo, OutBuffer);
-    else
-        MaxChannels(Comp, NumChans, SamplesToDo, OutBuffer);
+    LinkChannels(Comp, SamplesToDo, OutBuffer);
 
-    if(Comp->RmsWindow)
-        RmsDetection(Comp, SamplesToDo);
-    FollowEnvelope(Comp, SamplesToDo);
+    if(Comp->Auto.Attack || Comp->Auto.Release)
+        CrestDetector(Comp, SamplesToDo);
 
-    if(Comp->Ratio > 0.0f)
-        EnvelopeGain(Comp, SamplesToDo, 1.0f - (1.0f / Comp->Ratio));
+    if(Comp->Hold)
+        PeakHoldDetector(Comp, SamplesToDo);
     else
-        EnvelopeGain(Comp, SamplesToDo, 1.0f);
+        PeakDetector(Comp, SamplesToDo);
 
-    if(Comp->PostGain != 1.0f)
-    {
-        for(i = 0;i < SamplesToDo;i++)
-            Comp->Envelope[i] *= Comp->PostGain;
-    }
-    for(c = 0;c < NumChans;c++)
+    GainCompressor(Comp, SamplesToDo);
+
+    if(Comp->Delay)
+        SignalDelay(Comp, SamplesToDo, OutBuffer);
+
+    for(c = 0;c < numChans;c++)
     {
         for(i = 0;i < SamplesToDo;i++)
-            OutBuffer[c][i] *= Comp->Envelope[i];
+            OutBuffer[c][i] *= sideChain[(index + i) & mask];
     }
+
+    Comp->SideChainIndex = (index + SamplesToDo) & mask;
 }
diff --git a/Alc/mastering.h b/Alc/mastering.h
index 0a7b4901..206a6c91 100644
--- a/Alc/mastering.h
+++ b/Alc/mastering.h
@@ -6,51 +6,100 @@
 /* For BUFFERSIZE. */
 #include "alMain.h"
 
+/* These structures assume BUFFERSIZE is a power of 2.
+ */
+typedef struct SlidingHold
+{
+    ALfloat Values[BUFFERSIZE];
+    ALsizei Expiries[BUFFERSIZE];
+    ALsizei LowerIndex;
+    ALsizei UpperIndex;
+    ALsizei Length;
+} SlidingHold;
+
+/* General topology and basic automation was based on the following paper:
+ *
+ *   D. Giannoulis, M. Massberg and J. D. Reiss,
+ *   "Parameter Automation in a Dynamic Range Compressor,"
+ *   Journal of the Audio Engineering Society, v61 (10), Oct. 2013
+ *
+ * Available (along with supplemental reading) at:
+ *
+ *   http://c4dm.eecs.qmul.ac.uk/audioengineering/compressors/
+ */
 typedef struct Compressor {
+    ALsizei NumChans;
+    ALuint SampleRate;
+    struct {
+        ALuint Knee:1;
+        ALuint Attack:1;
+        ALuint Release:1;
+        ALuint PostGain:1;
+        ALuint Declip:1;
+    } Auto;
+    ALsizei LookAhead;
     ALfloat PreGain;
     ALfloat PostGain;
-    ALboolean SummedLink;
-    ALfloat AttackMin;
-    ALfloat AttackMax;
-    ALfloat ReleaseMin;
-    ALfloat ReleaseMax;
-    ALfloat Ratio;
     ALfloat Threshold;
+    ALfloat Slope;
     ALfloat Knee;
-    ALuint SampleRate;
-
-    ALuint RmsSum;
-    ALuint *RmsWindow;
-    ALsizei RmsIndex;
-    ALfloat Envelope[BUFFERSIZE];
-    ALfloat EnvLast;
+    ALfloat Attack;
+    ALfloat Release;
+    ALfloat SideChain[2*BUFFERSIZE];
+    ALsizei SideChainIndex;
+    ALfloat CrestFactor[BUFFERSIZE];
+    SlidingHold *Hold;
+    ALfloat (*Delay)[BUFFERSIZE];
+    ALsizei DelayIndex;
+    ALfloat CrestCoeff;
+    ALfloat GainEstimate;
+    ALfloat AdaptCoeff;
+    ALfloat LastPeakSq;
+    ALfloat LastRmsSq;
+    ALfloat LastRelease;
+    ALfloat LastAttack;
+    ALfloat LastGainDev;
 } Compressor;
 
-/* The compressor requires the following information for proper
- * initialization:
+/* The compressor is initialized with the following settings:
  *
+ *   NumChans       - Number of channels to process.
+ *   SampleRate     - Sample rate to process.
+ *   AutoKnee       - Whether to automate the knee width parameter.
+ *   AutoAttack     - Whether to automate the attack time parameter.
+ *   AutoRelease    - Whether to automate the release time parameter.
+ *   AutoPostGain   - Whether to automate the make-up (post) gain parameter.
+ *   AutoDeclip     - Whether to automate clipping reduction.  Ignored when
+ *                    not automating make-up gain.
+ *   LookAheadTime  - Look-ahead time (in seconds).
+ *   HoldTime       - Peak hold-time (in seconds).
  *   PreGainDb      - Gain applied before detection (in dB).
- *   PostGainDb     - Gain applied after compression (in dB).
- *   SummedLink     - Whether to use summed (true) or maxed (false) linking.
- *   RmsSensing     - Whether to use RMS (true) or Peak (false) sensing.
- *   AttackTimeMin  - Minimum attack time (in seconds).
- *   AttackTimeMax  - Maximum attack time.  Automates when min != max.
- *   ReleaseTimeMin - Minimum release time (in seconds).
- *   ReleaseTimeMax - Maximum release time.  Automates when min != max.
- *   Ratio          - Compression ratio (x:1).  Set to 0 for true limiter.
+ *   PostGainDb     - Make-up gain applied after compression (in dB).
  *   ThresholdDb    - Triggering threshold (in dB).
- *   KneeDb         - Knee width (below threshold; in dB).
- *   SampleRate     - Sample rate to process.
+ *   Ratio          - Compression ratio (x:1).  Set to INFINIFTY for true
+ *                    limiting.  Ignored when automating knee width.
+ *   KneeDb         - Knee width (in dB).  Ignored when automating knee
+ *                    width.
+ *   AttackTimeMin  - Attack time (in seconds).  Acts as a maximum when
+ *                    automating attack time.
+ *   ReleaseTimeMin - Release time (in seconds).  Acts as a maximum when
+ *                    automating release time.
  */
-Compressor *CompressorInit(const ALfloat PreGainDb, const ALfloat PostGainDb,
-    const ALboolean SummedLink, const ALboolean RmsSensing, const ALfloat AttackTimeMin,
-    const ALfloat AttackTimeMax, const ALfloat ReleaseTimeMin, const ALfloat ReleaseTimeMax,
-    const ALfloat Ratio, const ALfloat ThresholdDb, const ALfloat KneeDb,
-    const ALuint SampleRate);
+Compressor* CompressorInit(const ALuint NumChans, const ALuint SampleRate,
+    const ALboolean AutoKnee, const ALboolean AutoAttack,
+    const ALboolean AutoRelease, const ALboolean AutoPostGain,
+    const ALboolean AutoDeclip, const ALfloat LookAheadTime,
+    const ALfloat HoldTime, const ALfloat PreGainDb,
+    const ALfloat PostGainDb, const ALfloat ThresholdDb,
+    const ALfloat Ratio, const ALfloat KneeDb,
+    const ALfloat AttackTime, const ALfloat ReleaseTime);
 
-void ApplyCompression(struct Compressor *Comp, const ALsizei NumChans, const ALsizei SamplesToDo,
+void ApplyCompression(struct Compressor *Comp, const ALsizei SamplesToDo,
                       ALfloat (*restrict OutBuffer)[BUFFERSIZE]);
 
+inline ALsizei GetCompressorChannelCount(const Compressor *Comp)
+{ return Comp->NumChans; }
+
 inline ALuint GetCompressorSampleRate(const Compressor *Comp)
 { return Comp->SampleRate; }