Apply distance compensation when writing to the output

5 files changed, 106 insertions, 122 deletions

diff --git a/Alc/ALu.c b/Alc/ALu.c
index 2476622b..1fca0ed0 100644
--- a/Alc/ALu.c
+++ b/Alc/ALu.c
@@ -1373,15 +1373,39 @@ static inline ALubyte aluF2UB(ALfloat val)
 { return aluF2B(val)+128; }
 
 #define DECL_TEMPLATE(T, func)                                                \
-static void Write_##T(ALfloatBUFFERSIZE *InBuffer, ALvoid *OutBuffer,         \
-                      ALsizei SamplesToDo, ALsizei numchans)                  \
+static void Write_##T(const ALfloatBUFFERSIZE *InBuffer, ALvoid *OutBuffer,   \
+                      DistanceComp *distcomp, ALsizei SamplesToDo,            \
+                      ALsizei numchans)                                       \
 {                                                                             \
     ALsizei i, j;                                                             \
     for(j = 0;j < numchans;j++)                                               \
     {                                                                         \
         const ALfloat *in = InBuffer[j];                                      \
         T *restrict out = (T*)OutBuffer + j;                                  \
-        for(i = 0;i < SamplesToDo;i++)                                        \
+        const ALfloat gain = distcomp[j].Gain;                                \
+        const ALsizei base = distcomp[j].Length;                              \
+        if(base > 0 || gain != 1.0f)                                          \
+        {                                                                     \
+            if(SamplesToDo >= base)                                           \
+            {                                                                 \
+                for(i = 0;i < base;i++)                                       \
+                    out[i*numchans] = func(distcomp[j].Buffer[i]*gain);       \
+                for(;i < SamplesToDo;i++)                                     \
+                    out[i*numchans] = func(in[i-base]*gain);                  \
+                memcpy(distcomp[j].Buffer, &in[SamplesToDo-base],             \
+                        base*sizeof(ALfloat));                                \
+            }                                                                 \
+            else                                                              \
+            {                                                                 \
+                for(i = 0;i < SamplesToDo;i++)                                \
+                    out[i*numchans] = func(distcomp[j].Buffer[i]*gain);       \
+                memmove(distcomp[j].Buffer, distcomp[j].Buffer+SamplesToDo,   \
+                        (base-SamplesToDo)*sizeof(ALfloat));                  \
+                memcpy(distcomp[j].Buffer+base-SamplesToDo, in,               \
+                        SamplesToDo*sizeof(ALfloat));                         \
+            }                                                                 \
+        }                                                                     \
+        else for(i = 0;i < SamplesToDo;i++)                                   \
             out[i*numchans] = func(in[i]);                                    \
     }                                                                         \
 }
@@ -1564,33 +1588,34 @@ void aluMixData(ALCdevice *device, ALvoid *buffer, ALsizei size)
         {
             ALfloat (*OutBuffer)[BUFFERSIZE] = device->RealOut.Buffer;
             ALsizei OutChannels = device->RealOut.NumChannels;
+            DistanceComp *DistComp = device->ChannelDelay;
 
-#define WRITE(T, a, b, c, d) do {               \
-    Write_##T((a), (b), (c), (d));              \
-    buffer = (T*)buffer + (c)*(d);              \
+#define WRITE(T, a, b, c, d, e) do {                                          \
+    Write_##T(SAFE_CONST(ALfloatBUFFERSIZE*,(a)), (b), (c), (d), (e));        \
+    buffer = (T*)buffer + (d)*(e);                                            \
 } while(0)
             switch(device->FmtType)
             {
                 case DevFmtByte:
-                    WRITE(ALbyte, OutBuffer, buffer, SamplesToDo, OutChannels);
+                    WRITE(ALbyte, OutBuffer, buffer, DistComp, SamplesToDo, OutChannels);
                     break;
                 case DevFmtUByte:
-                    WRITE(ALubyte, OutBuffer, buffer, SamplesToDo, OutChannels);
+                    WRITE(ALubyte, OutBuffer, buffer, DistComp, SamplesToDo, OutChannels);
                     break;
                 case DevFmtShort:
-                    WRITE(ALshort, OutBuffer, buffer, SamplesToDo, OutChannels);
+                    WRITE(ALshort, OutBuffer, buffer, DistComp, SamplesToDo, OutChannels);
                     break;
                 case DevFmtUShort:
-                    WRITE(ALushort, OutBuffer, buffer, SamplesToDo, OutChannels);
+                    WRITE(ALushort, OutBuffer, buffer, DistComp, SamplesToDo, OutChannels);
                     break;
                 case DevFmtInt:
-                    WRITE(ALint, OutBuffer, buffer, SamplesToDo, OutChannels);
+                    WRITE(ALint, OutBuffer, buffer, DistComp, SamplesToDo, OutChannels);
                     break;
                 case DevFmtUInt:
-                    WRITE(ALuint, OutBuffer, buffer, SamplesToDo, OutChannels);
+                    WRITE(ALuint, OutBuffer, buffer, DistComp, SamplesToDo, OutChannels);
                     break;
                 case DevFmtFloat:
-                    WRITE(ALfloat, OutBuffer, buffer, SamplesToDo, OutChannels);
+                    WRITE(ALfloat, OutBuffer, buffer, DistComp, SamplesToDo, OutChannels);
                     break;
             }
 #undef WRITE
diff --git a/Alc/bformatdec.c b/Alc/bformatdec.c
index 549cd42c..5b848ee0 100644
--- a/Alc/bformatdec.c
+++ b/Alc/bformatdec.c
@@ -156,10 +156,6 @@ static void init_bformatdec(void)
 }
 
 
-#define MAX_DELAY_LENGTH 128
-
-#define INVALID_UPSAMPLE_INDEX  INT_MAX
-
 /* NOTE: BandSplitter filters are unused with single-band decoding */
 typedef struct BFormatDec {
     ALboolean Enabled[MAX_OUTPUT_CHANNELS];
@@ -179,11 +175,6 @@ typedef struct BFormatDec {
     alignas(16) ALfloat ChannelMix[BUFFERSIZE];
 
     struct {
-        alignas(16) ALfloat Buffer[MAX_DELAY_LENGTH];
-        ALsizei Length; /* Valid range is [0...MAX_DELAY_LENGTH). */
-    } Delay[MAX_OUTPUT_CHANNELS];
-
-    struct {
         BandSplitter XOver;
         ALfloat Gains[FB_Max];
     } UpSampler[4];
@@ -235,14 +226,13 @@ int bformatdec_isPeriphonic(const struct BFormatDec *dec)
     return dec->Periphonic;
 }
 
-void bformatdec_reset(BFormatDec *dec, const AmbDecConf *conf, ALsizei chancount, ALuint srate, const ALsizei chanmap[MAX_OUTPUT_CHANNELS], int flags)
+void bformatdec_reset(BFormatDec *dec, const AmbDecConf *conf, ALsizei chancount, ALuint srate, const ALsizei chanmap[MAX_OUTPUT_CHANNELS])
 {
     static const ALsizei map2DTo3D[MAX_AMBI2D_COEFFS] = {
         0,  1, 3,  4, 8,  9, 15
     };
     const ALfloat *coeff_scale = UnitScale;
-    ALfloat distgain[MAX_OUTPUT_CHANNELS];
-    ALfloat maxdist, ratio;
+    ALfloat ratio;
     ALsizei i;
 
     al_free(dec->Samples);
@@ -300,41 +290,6 @@ void bformatdec_reset(BFormatDec *dec, const AmbDecConf *conf, ALsizei chancount
         dec->UpSampler[3].Gains[FB_LowFreq] = 0.0f;
     }
 
-    maxdist = 0.0f;
-    for(i = 0;i < conf->NumSpeakers;i++)
-    {
-        maxdist = maxf(maxdist, conf->Speakers[i].Distance);
-        distgain[i] = 1.0f;
-    }
-
-    memset(dec->Delay, 0, sizeof(dec->Delay));
-    if((flags&BFDF_DistanceComp) && maxdist > 0.0f)
-    {
-        for(i = 0;i < conf->NumSpeakers;i++)
-        {
-            ALsizei chan = chanmap[i];
-            ALfloat delay;
-
-            /* Distance compensation only delays in steps of the sample rate.
-             * This is a bit less accurate since the delay time falls to the
-             * nearest sample time, but it's far simpler as it doesn't have to
-             * deal with phase offsets. This means at 48khz, for instance, the
-             * distance delay will be in steps of about 7 millimeters.
-             */
-            delay = floorf((maxdist-conf->Speakers[i].Distance) / SPEEDOFSOUNDMETRESPERSEC *
-                           (ALfloat)srate + 0.5f);
-            if(delay >= (ALfloat)MAX_DELAY_LENGTH)
-                ERR("Delay for speaker \"%s\" exceeds buffer length (%f >= %u)\n",
-                    al_string_get_cstr(conf->Speakers[i].Name), delay, MAX_DELAY_LENGTH);
-
-            dec->Delay[chan].Length = (ALuint)clampf(delay, 0.0f, (ALfloat)(MAX_DELAY_LENGTH-1));
-            distgain[i] = conf->Speakers[i].Distance / maxdist;
-            TRACE("Channel %u \"%s\" distance compensation: %d samples, %f gain\n", chan,
-                al_string_get_cstr(conf->Speakers[i].Name), dec->Delay[chan].Length, distgain[i]
-            );
-        }
-    }
-
     memset(&dec->Matrix, 0, sizeof(dec->Matrix));
     if(conf->FreqBands == 1)
     {
@@ -356,7 +311,7 @@ void bformatdec_reset(BFormatDec *dec, const AmbDecConf *conf, ALsizei chancount
                     else if(j == 5) gain = conf->HFOrderGain[3];
                     if((conf->ChanMask&(1<<l)))
                         dec->Matrix.Single[chan][j] = conf->HFMatrix[i][k++] / coeff_scale[l] *
-                                                      gain * distgain[i];
+                                                      gain;
                 }
             }
             else
@@ -369,7 +324,7 @@ void bformatdec_reset(BFormatDec *dec, const AmbDecConf *conf, ALsizei chancount
                     else if(j == 9) gain = conf->HFOrderGain[3];
                     if((conf->ChanMask&(1<<j)))
                         dec->Matrix.Single[chan][j] = conf->HFMatrix[i][k++] / coeff_scale[j] *
-                                                      gain * distgain[i];
+                                                      gain;
                 }
             }
         }
@@ -400,8 +355,7 @@ void bformatdec_reset(BFormatDec *dec, const AmbDecConf *conf, ALsizei chancount
                     else if(j == 5) gain = conf->HFOrderGain[3] * ratio;
                     if((conf->ChanMask&(1<<l)))
                         dec->Matrix.Dual[chan][FB_HighFreq][j] = conf->HFMatrix[i][k++] /
-                                                                 coeff_scale[l] * gain *
-                                                                 distgain[i];
+                                                                 coeff_scale[l] * gain;
                 }
                 for(j = 0,k = 0;j < MAX_AMBI2D_COEFFS;j++)
                 {
@@ -412,8 +366,7 @@ void bformatdec_reset(BFormatDec *dec, const AmbDecConf *conf, ALsizei chancount
                     else if(j == 5) gain = conf->LFOrderGain[3] / ratio;
                     if((conf->ChanMask&(1<<l)))
                         dec->Matrix.Dual[chan][FB_LowFreq][j] = conf->LFMatrix[i][k++] /
-                                                                coeff_scale[l] * gain *
-                                                                distgain[i];
+                                                                coeff_scale[l] * gain;
                 }
             }
             else
@@ -426,8 +379,7 @@ void bformatdec_reset(BFormatDec *dec, const AmbDecConf *conf, ALsizei chancount
                     else if(j == 9) gain = conf->HFOrderGain[3] * ratio;
                     if((conf->ChanMask&(1<<j)))
                         dec->Matrix.Dual[chan][FB_HighFreq][j] = conf->HFMatrix[i][k++] /
-                                                                 coeff_scale[j] * gain *
-                                                                 distgain[i];
+                                                                 coeff_scale[j] * gain;
                 }
                 for(j = 0,k = 0;j < MAX_AMBI_COEFFS;j++)
                 {
@@ -437,8 +389,7 @@ void bformatdec_reset(BFormatDec *dec, const AmbDecConf *conf, ALsizei chancount
                     else if(j == 9) gain = conf->LFOrderGain[3] / ratio;
                     if((conf->ChanMask&(1<<j)))
                         dec->Matrix.Dual[chan][FB_LowFreq][j] = conf->LFMatrix[i][k++] /
-                                                                coeff_scale[j] * gain *
-                                                                distgain[i];
+                                                                coeff_scale[j] * gain;
                 }
             }
         }
@@ -471,29 +422,7 @@ void bformatdec_process(struct BFormatDec *dec, ALfloat (*restrict OutBuffer)[BU
                 SamplesToDo
             );
 
-            if(dec->Delay[chan].Length > 0)
-            {
-                const ALsizei base = dec->Delay[chan].Length;
-                if(SamplesToDo >= base)
-                {
-                    for(i = 0;i < base;i++)
-                        OutBuffer[chan][i] += dec->Delay[chan].Buffer[i];
-                    for(;i < SamplesToDo;i++)
-                        OutBuffer[chan][i] += dec->ChannelMix[i-base];
-                    memcpy(dec->Delay[chan].Buffer, &dec->ChannelMix[SamplesToDo-base],
-                           base*sizeof(ALfloat));
-                }
-                else
-                {
-                    for(i = 0;i < SamplesToDo;i++)
-                        OutBuffer[chan][i] += dec->Delay[chan].Buffer[i];
-                    memmove(dec->Delay[chan].Buffer, dec->Delay[chan].Buffer+SamplesToDo,
-                            base - SamplesToDo);
-                    memcpy(dec->Delay[chan].Buffer+base-SamplesToDo, dec->ChannelMix,
-                           SamplesToDo*sizeof(ALfloat));
-                }
-            }
-            else for(i = 0;i < SamplesToDo;i++)
+            for(i = 0;i < SamplesToDo;i++)
                 OutBuffer[chan][i] += dec->ChannelMix[i];
         }
     }
@@ -508,29 +437,7 @@ void bformatdec_process(struct BFormatDec *dec, ALfloat (*restrict OutBuffer)[BU
             MixMatrixRow(dec->ChannelMix, dec->Matrix.Single[chan], InSamples,
                          dec->NumChannels, 0, SamplesToDo);
 
-            if(dec->Delay[chan].Length > 0)
-            {
-                const ALsizei base = dec->Delay[chan].Length;
-                if(SamplesToDo >= base)
-                {
-                    for(i = 0;i < base;i++)
-                        OutBuffer[chan][i] += dec->Delay[chan].Buffer[i];
-                    for(;i < SamplesToDo;i++)
-                        OutBuffer[chan][i] += dec->ChannelMix[i-base];
-                    memcpy(dec->Delay[chan].Buffer, &dec->ChannelMix[SamplesToDo-base],
-                           base*sizeof(ALfloat));
-                }
-                else
-                {
-                    for(i = 0;i < SamplesToDo;i++)
-                        OutBuffer[chan][i] += dec->Delay[chan].Buffer[i];
-                    memmove(dec->Delay[chan].Buffer, dec->Delay[chan].Buffer+SamplesToDo,
-                            base - SamplesToDo);
-                    memcpy(dec->Delay[chan].Buffer+base-SamplesToDo, dec->ChannelMix,
-                           SamplesToDo*sizeof(ALfloat));
-                }
-            }
-            else for(i = 0;i < SamplesToDo;i++)
+            for(i = 0;i < SamplesToDo;i++)
                 OutBuffer[chan][i] += dec->ChannelMix[i];
         }
     }
@@ -570,6 +477,8 @@ void bformatdec_upSample(struct BFormatDec *dec, ALfloat (*restrict OutBuffer)[B
 }
 
 
+#define INVALID_UPSAMPLE_INDEX INT_MAX
+
 static ALsizei GetACNIndex(const BFChannelConfig *chans, ALsizei numchans, ALsizei acn)
 {
     ALsizei i;
diff --git a/Alc/bformatdec.h b/Alc/bformatdec.h
index 5f6e230d..97f36d0e 100644
--- a/Alc/bformatdec.h
+++ b/Alc/bformatdec.h
@@ -33,7 +33,7 @@ struct BFormatDec *bformatdec_alloc();
 void bformatdec_free(struct BFormatDec *dec);
 int bformatdec_getOrder(const struct BFormatDec *dec);
 int bformatdec_isPeriphonic(const struct BFormatDec *dec);
-void bformatdec_reset(struct BFormatDec *dec, const struct AmbDecConf *conf, ALsizei chancount, ALuint srate, const ALsizei chanmap[MAX_OUTPUT_CHANNELS], int flags);
+void bformatdec_reset(struct BFormatDec *dec, const struct AmbDecConf *conf, ALsizei chancount, ALuint srate, const ALsizei chanmap[MAX_OUTPUT_CHANNELS]);
 
 /* Decodes the ambisonic input to the given output channels. */
 void bformatdec_process(struct BFormatDec *dec, ALfloat (*restrict OutBuffer)[BUFFERSIZE], ALsizei OutChannels, const ALfloat (*restrict InSamples)[BUFFERSIZE], ALsizei SamplesToDo);
diff --git a/Alc/panning.c b/Alc/panning.c
index eb2ec0c8..80532bc4 100644
--- a/Alc/panning.c
+++ b/Alc/panning.c
@@ -693,13 +693,10 @@ static void InitCustomPanning(ALCdevice *device, const AmbDecConf *conf, const A
 static void InitHQPanning(ALCdevice *device, const AmbDecConf *conf, const ALsizei speakermap[MAX_OUTPUT_CHANNELS])
 {
     const char *devname;
-    int decflags = 0;
     size_t count;
     size_t i;
 
     devname = al_string_get_cstr(device->DeviceName);
-    if(GetConfigValueBool(devname, "decoder", "distance-comp", 1))
-        decflags |= BFDF_DistanceComp;
 
     if((conf->ChanMask&AMBI_PERIPHONIC_MASK))
     {
@@ -731,8 +728,7 @@ static void InitHQPanning(ALCdevice *device, const AmbDecConf *conf, const ALsiz
         (conf->ChanMask > 0xf) ? (conf->ChanMask > 0x1ff) ? "third" : "second" : "first",
         (conf->ChanMask&AMBI_PERIPHONIC_MASK) ? " periphonic" : ""
     );
-    bformatdec_reset(device->AmbiDecoder, conf, count, device->Frequency,
-                     speakermap, decflags);
+    bformatdec_reset(device->AmbiDecoder, conf, count, device->Frequency, speakermap);
 
     if(bformatdec_getOrder(device->AmbiDecoder) < 2)
     {
@@ -759,6 +755,41 @@ static void InitHQPanning(ALCdevice *device, const AmbDecConf *conf, const ALsiz
         }
         device->FOAOut.CoeffCount = 0;
     }
+
+    ALfloat maxdist = 0.0f;
+    for(i = 0;i < (size_t)conf->NumSpeakers;i++)
+        maxdist = maxf(maxdist, conf->Speakers[i].Distance);
+
+    if(GetConfigValueBool(devname, "decoder", "distance-comp", 1) && maxdist > 0.0f)
+    {
+        ALfloat srate = (ALfloat)device->Frequency;
+        for(i = 0;i < (size_t)conf->NumSpeakers;i++)
+        {
+            ALsizei chan = speakermap[i];
+            ALfloat delay;
+
+            /* Distance compensation only delays in steps of the sample rate.
+             * This is a bit less accurate since the delay time falls to the
+             * nearest sample time, but it's far simpler as it doesn't have to
+             * deal with phase offsets. This means at 48khz, for instance, the
+             * distance delay will be in steps of about 7 millimeters.
+             */
+            delay = floorf((maxdist-conf->Speakers[i].Distance) / SPEEDOFSOUNDMETRESPERSEC *
+                           srate + 0.5f);
+            if(delay >= (ALfloat)MAX_DELAY_LENGTH)
+                ERR("Delay for speaker \"%s\" exceeds buffer length (%f >= %u)\n",
+                    al_string_get_cstr(conf->Speakers[i].Name), delay, MAX_DELAY_LENGTH);
+
+            device->ChannelDelay[chan].Length = (ALsizei)clampf(
+                delay, 0.0f, (ALfloat)(MAX_DELAY_LENGTH-1)
+            );
+            device->ChannelDelay[chan].Gain = conf->Speakers[i].Distance / maxdist;
+            TRACE("Channel %u \"%s\" distance compensation: %d samples, %f gain\n", chan,
+                al_string_get_cstr(conf->Speakers[i].Name), device->ChannelDelay[chan].Length,
+                device->ChannelDelay[chan].Gain
+            );
+        }
+    }
 }
 
 static void InitHrtfPanning(ALCdevice *device, bool hoa_mode)
@@ -887,6 +918,13 @@ void aluInitRenderer(ALCdevice *device, ALint hrtf_id, enum HrtfRequestMode hrtf
     device->Dry.CoeffCount = 0;
     device->Dry.NumChannels = 0;
 
+    memset(device->ChannelDelay, 0, sizeof(device->ChannelDelay));
+    for(i = 0;i < MAX_OUTPUT_CHANNELS;i++)
+    {
+        device->ChannelDelay[i].Gain = 1.0f;
+        device->ChannelDelay[i].Length = 0;
+    }
+
     if(device->FmtChans != DevFmtStereo)
     {
         ALsizei speakermap[MAX_OUTPUT_CHANNELS];
diff --git a/OpenAL32/Include/alMain.h b/OpenAL32/Include/alMain.h
index 947a16ba..bb8a57ce 100644
--- a/OpenAL32/Include/alMain.h
+++ b/OpenAL32/Include/alMain.h
@@ -603,6 +603,15 @@ typedef struct HrtfEntry {
 TYPEDEF_VECTOR(HrtfEntry, vector_HrtfEntry)
 
 
+/* Maximum delay in samples for speaker distance compensation. */
+#define MAX_DELAY_LENGTH 128
+
+typedef struct DistanceComp {
+    ALfloat Gain;
+    ALsizei Length; /* Valid range is [0...MAX_DELAY_LENGTH). */
+    alignas(16) ALfloat Buffer[MAX_DELAY_LENGTH];
+} DistanceComp;
+
 /* Size for temporary storage of buffer data, in ALfloats. Larger values need
  * more memory, while smaller values may need more iterations. The value needs
  * to be a sensible size, however, as it constrains the max stepping value used
@@ -723,6 +732,9 @@ struct ALCdevice_struct
         ALsizei NumChannels;
     } RealOut;
 
+    /* Delay buffers used to compensate for speaker distances. */
+    DistanceComp ChannelDelay[MAX_OUTPUT_CHANNELS];
+
     /* Running count of the mixer invocations, in 31.1 fixed point. This
      * actually increments *twice* when mixing, first at the start and then at
      * the end, so the bottom bit indicates if the device is currently mixing