Dynamically allocate the channel delay buffers

4 files changed, 66 insertions, 8 deletions

diff --git a/Alc/ALc.c b/Alc/ALc.c
index aedcd493..b46244aa 100644
--- a/Alc/ALc.c
+++ b/Alc/ALc.c
@@ -1807,6 +1807,7 @@ static ALCenum UpdateDeviceParams(ALCdevice *device, const ALCint *attrList)
     ALCuint oldFreq;
     FPUCtl oldMode;
     size_t size;
+    ALCsizei i;
 
     // Check for attributes
     if(device->Type == Loopback)
@@ -2048,6 +2049,13 @@ static ALCenum UpdateDeviceParams(ALCdevice *device, const ALCint *attrList)
     al_free(device->Bs2b);
     device->Bs2b = NULL;
 
+    al_free(device->ChannelDelay[0].Buffer);
+    for(i = 0;i < MAX_OUTPUT_CHANNELS;i++)
+    {
+        device->ChannelDelay[i].Length = 0;
+        device->ChannelDelay[i].Buffer = NULL;
+    }
+
     al_free(device->Dry.Buffer);
     device->Dry.Buffer = NULL;
     device->Dry.NumChannels = 0;
@@ -2329,6 +2337,8 @@ static ALCenum UpdateDeviceParams(ALCdevice *device, const ALCint *attrList)
  */
 static ALCvoid FreeDevice(ALCdevice *device)
 {
+    ALsizei i;
+
     TRACE("%p\n", device);
 
     V0(device->Backend,close)();
@@ -2382,6 +2392,14 @@ static ALCvoid FreeDevice(ALCdevice *device)
     ambiup_free(device->AmbiUp);
     device->AmbiUp = NULL;
 
+    al_free(device->ChannelDelay[0].Buffer);
+    for(i = 0;i < MAX_OUTPUT_CHANNELS;i++)
+    {
+        device->ChannelDelay[i].Gain   = 1.0f;
+        device->ChannelDelay[i].Length = 0;
+        device->ChannelDelay[i].Buffer = NULL;
+    }
+
     AL_STRING_DEINIT(device->DeviceName);
 
     al_free(device->Dry.Buffer);
@@ -3687,6 +3705,7 @@ ALC_API ALCdevice* ALC_APIENTRY alcOpenDevice(const ALCchar *deviceName)
     const ALCchar *fmt;
     ALCdevice *device;
     ALCenum err;
+    ALCsizei i;
 
     DO_INITCONFIG();
 
@@ -3748,6 +3767,13 @@ ALC_API ALCdevice* ALC_APIENTRY alcOpenDevice(const ALCchar *deviceName)
     InitUIntMap(&device->EffectMap, ~0);
     InitUIntMap(&device->FilterMap, ~0);
 
+    for(i = 0;i < MAX_OUTPUT_CHANNELS;i++)
+    {
+        device->ChannelDelay[i].Gain   = 1.0f;
+        device->ChannelDelay[i].Length = 0;
+        device->ChannelDelay[i].Buffer = NULL;
+    }
+
     //Set output format
     device->FmtChans = DevFmtChannelsDefault;
     device->FmtType = DevFmtTypeDefault;
@@ -3991,6 +4017,7 @@ ALC_API ALCdevice* ALC_APIENTRY alcCaptureOpenDevice(const ALCchar *deviceName,
 {
     ALCdevice *device = NULL;
     ALCenum err;
+    ALCsizei i;
 
     DO_INITCONFIG();
 
@@ -4036,6 +4063,13 @@ ALC_API ALCdevice* ALC_APIENTRY alcCaptureOpenDevice(const ALCchar *deviceName,
     InitUIntMap(&device->EffectMap, ~0);
     InitUIntMap(&device->FilterMap, ~0);
 
+    for(i = 0;i < MAX_OUTPUT_CHANNELS;i++)
+    {
+        device->ChannelDelay[i].Gain   = 1.0f;
+        device->ChannelDelay[i].Length = 0;
+        device->ChannelDelay[i].Buffer = NULL;
+    }
+
     if(!CaptureBackend.getFactory)
         device->Backend = create_backend_wrapper(device, &CaptureBackend.Funcs,
                                                  ALCbackend_Capture);
@@ -4198,6 +4232,7 @@ ALC_API ALCdevice* ALC_APIENTRY alcLoopbackOpenDeviceSOFT(const ALCchar *deviceN
 {
     ALCbackendFactory *factory;
     ALCdevice *device;
+    ALCsizei i;
 
     DO_INITCONFIG();
 
@@ -4248,6 +4283,13 @@ ALC_API ALCdevice* ALC_APIENTRY alcLoopbackOpenDeviceSOFT(const ALCchar *deviceN
     InitUIntMap(&device->EffectMap, ~0);
     InitUIntMap(&device->FilterMap, ~0);
 
+    for(i = 0;i < MAX_OUTPUT_CHANNELS;i++)
+    {
+        device->ChannelDelay[i].Gain   = 1.0f;
+        device->ChannelDelay[i].Length = 0;
+        device->ChannelDelay[i].Buffer = NULL;
+    }
+
     factory = ALCloopbackFactory_getFactory();
     device->Backend = V(factory,createBackend)(device, ALCbackend_Loopback);
     if(!device->Backend)
diff --git a/Alc/ALu.c b/Alc/ALu.c
index b4938152..9fe5712d 100644
--- a/Alc/ALu.c
+++ b/Alc/ALu.c
@@ -1332,24 +1332,24 @@ static void Write_##T(const ALfloatBUFFERSIZE *InBuffer, ALvoid *OutBuffer,   \
         T *restrict out = (T*)OutBuffer + j;                                  \
         const ALfloat gain = distcomp[j].Gain;                                \
         const ALsizei base = distcomp[j].Length;                              \
+        ALfloat *restrict distbuf = ASSUME_ALIGNED(distcomp[j].Buffer, 16);   \
         if(base > 0 || gain != 1.0f)                                          \
         {                                                                     \
             if(SamplesToDo >= base)                                           \
             {                                                                 \
                 for(i = 0;i < base;i++)                                       \
-                    out[i*numchans] = func(distcomp[j].Buffer[i]*gain);       \
+                    out[i*numchans] = func(distbuf[i]*gain);                  \
                 for(;i < SamplesToDo;i++)                                     \
                     out[i*numchans] = func(in[i-base]*gain);                  \
-                memcpy(distcomp[j].Buffer, &in[SamplesToDo-base],             \
-                        base*sizeof(ALfloat));                                \
+                memcpy(distbuf, &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,   \
+                    out[i*numchans] = func(distbuf[i]*gain);                  \
+                memmove(distbuf, distbuf+SamplesToDo,                         \
                         (base-SamplesToDo)*sizeof(ALfloat));                  \
-                memcpy(distcomp[j].Buffer+base-SamplesToDo, in,               \
+                memcpy(distbuf+base-SamplesToDo, in,                          \
                         SamplesToDo*sizeof(ALfloat));                         \
             }                                                                 \
         }                                                                     \
diff --git a/Alc/panning.c b/Alc/panning.c
index 84e1dbed..c4d3e43f 100644
--- a/Alc/panning.c
+++ b/Alc/panning.c
@@ -616,6 +616,7 @@ static void InitDistanceComp(ALCdevice *device, const AmbDecConf *conf, const AL
 {
     const char *devname = al_string_get_cstr(device->DeviceName);
     ALfloat maxdist = 0.0f;
+    ALsizei total = 0;
     ALsizei i;
 
     for(i = 0;i < conf->NumSpeakers;i++)
@@ -649,6 +650,21 @@ static void InitDistanceComp(ALCdevice *device, const AmbDecConf *conf, const AL
                 al_string_get_cstr(conf->Speakers[i].Name), device->ChannelDelay[chan].Length,
                 device->ChannelDelay[chan].Gain
             );
+
+            /* Round up to the next 4th sample, so each channel buffer starts
+             * 16-byte aligned.
+             */
+            total += RoundUp(device->ChannelDelay[chan].Length, 4);
+        }
+    }
+
+    if(total > 0)
+    {
+        device->ChannelDelay[0].Buffer = al_calloc(16, total * sizeof(ALfloat));
+        for(i = 1;i < MAX_OUTPUT_CHANNELS;i++)
+        {
+            size_t len = RoundUp(device->ChannelDelay[i-1].Length, 4);
+            device->ChannelDelay[i].Buffer = device->ChannelDelay[i-1].Buffer + len;
         }
     }
 }
diff --git a/OpenAL32/Include/alMain.h b/OpenAL32/Include/alMain.h
index 9407cdce..cc30dcef 100644
--- a/OpenAL32/Include/alMain.h
+++ b/OpenAL32/Include/alMain.h
@@ -631,12 +631,12 @@ TYPEDEF_VECTOR(HrtfEntry, vector_HrtfEntry)
 
 
 /* Maximum delay in samples for speaker distance compensation. */
-#define MAX_DELAY_LENGTH 128
+#define MAX_DELAY_LENGTH 1024
 
 typedef struct DistanceComp {
     ALfloat Gain;
     ALsizei Length; /* Valid range is [0...MAX_DELAY_LENGTH). */
-    alignas(16) ALfloat Buffer[MAX_DELAY_LENGTH];
+    ALfloat *Buffer;
 } DistanceComp;
 
 /* Size for temporary storage of buffer data, in ALfloats. Larger values need