Avoid holding HRTF accumulation samples per-source

It notably simplifies things to mix HRTF sources into an accumulation buffer
together, which the Dry buffer's Ambisonic-to-HRTF decode is then added to,
before being mixed to the Real output.

9 files changed, 44 insertions, 83 deletions

diff --git a/alc/alc.cpp b/alc/alc.cpp
index cd2911f5..bee42fc5 100644
--- a/alc/alc.cpp
+++ b/alc/alc.cpp
@@ -1853,6 +1853,8 @@ static ALCenum UpdateDeviceParams(ALCdevice *device, const ALCint *attrList)
     device->Limiter = nullptr;
     device->ChannelDelay.clear();
 
+    std::fill(std::begin(device->HrtfAccumData), std::end(device->HrtfAccumData), float2{});
+
     device->Dry.AmbiMap.fill(BFChannelConfig{});
     device->Dry.Buffer = {};
     std::fill(std::begin(device->NumChannelsPerOrder), std::end(device->NumChannelsPerOrder), 0u);
diff --git a/alc/alcmain.h b/alc/alcmain.h
index c26b3a28..30c5b835 100644
--- a/alc/alcmain.h
+++ b/alc/alcmain.h
@@ -277,6 +277,8 @@ struct ALCdevice : public al::intrusive_ref<ALCdevice> {
         alignas(16) ALfloat HrtfSourceData[BUFFERSIZE + HRTF_HISTORY_LENGTH];
         alignas(16) ALfloat NfcSampleData[BUFFERSIZE];
     };
+
+    /* Persistent storage for HRTF mixing. */
     alignas(16) float2 HrtfAccumData[BUFFERSIZE + HRIR_LENGTH];
 
     /* Mixing buffer used by the Dry mix and Real output. */
diff --git a/alc/hrtf.h b/alc/hrtf.h
index d133133a..a81ebfed 100644
--- a/alc/hrtf.h
+++ b/alc/hrtf.h
@@ -68,7 +68,6 @@ using HrirArray = std::array<float2,HRIR_LENGTH>;
 
 struct HrtfState {
     alignas(16) std::array<ALfloat,HRTF_HISTORY_LENGTH> History;
-    alignas(16) HrirArray Values;
 };
 
 struct HrtfFilter {
@@ -80,7 +79,6 @@ struct HrtfFilter {
 struct DirectHrtfState {
     /* HRTF filter state for dry buffer content */
     ALuint IrSize{0};
-    alignas(16) HrirArray Values{};
     al::FlexArray<HrirArray,16> Coeffs;
 
     DirectHrtfState(size_t numchans) : Coeffs{numchans} { }
diff --git a/alc/mixer/defs.h b/alc/mixer/defs.h
index b2535265..1e5b40d8 100644
--- a/alc/mixer/defs.h
+++ b/alc/mixer/defs.h
@@ -41,13 +41,11 @@ void MixRow_(const al::span<float> OutBuffer, const al::span<const float> Gains,
     const float *InSamples, const size_t InStride);
 
 template<InstSetType InstTag>
-void MixHrtf_(FloatBufferLine &LeftOut, FloatBufferLine &RightOut, const float *InSamples,
-    float2 *AccumSamples, const size_t OutPos, const ALuint IrSize, MixHrtfFilter *hrtfparams,
-    const size_t BufferSize);
+void MixHrtf_(const float *InSamples, float2 *AccumSamples, const ALuint IrSize,
+    MixHrtfFilter *hrtfparams, const size_t BufferSize);
 template<InstSetType InstTag>
-void MixHrtfBlend_(FloatBufferLine &LeftOut, FloatBufferLine &RightOut, const float *InSamples,
-    float2 *AccumSamples, const size_t OutPos, const ALuint IrSize, const HrtfFilter *oldparams,
-    MixHrtfFilter *newparams, const size_t BufferSize);
+void MixHrtfBlend_(const float *InSamples, float2 *AccumSamples, const ALuint IrSize,
+    const HrtfFilter *oldparams, MixHrtfFilter *newparams, const size_t BufferSize);
 template<InstSetType InstTag>
 void MixDirectHrtf_(FloatBufferLine &LeftOut, FloatBufferLine &RightOut,
     const al::span<const FloatBufferLine> InSamples, float2 *AccumSamples, DirectHrtfState *State,
diff --git a/alc/mixer/hrtfbase.h b/alc/mixer/hrtfbase.h
index cbc885c5..4a6eab50 100644
--- a/alc/mixer/hrtfbase.h
+++ b/alc/mixer/hrtfbase.h
@@ -13,9 +13,8 @@ using ApplyCoeffsT = void(&)(float2 *RESTRICT Values, const ALuint irSize, const
     const float left, const float right);
 
 template<ApplyCoeffsT ApplyCoeffs>
-inline void MixHrtfBase(FloatBufferLine &LeftOut, FloatBufferLine &RightOut,
-    const float *InSamples, float2 *RESTRICT AccumSamples, const size_t OutPos,
-    const ALuint IrSize, MixHrtfFilter *hrtfparams, const size_t BufferSize)
+inline void MixHrtfBase(const float *InSamples, float2 *RESTRICT AccumSamples, const ALuint IrSize,
+    MixHrtfFilter *hrtfparams, const size_t BufferSize)
 {
     ASSUME(BufferSize > 0);
 
@@ -38,17 +37,11 @@ inline void MixHrtfBase(FloatBufferLine &LeftOut, FloatBufferLine &RightOut,
         stepcount += 1.0f;
     }
 
-    for(size_t i{0u};i < BufferSize;++i)
-        LeftOut[OutPos+i]  += AccumSamples[i][0];
-    for(size_t i{0u};i < BufferSize;++i)
-        RightOut[OutPos+i] += AccumSamples[i][1];
-
     hrtfparams->Gain = gain + gainstep*stepcount;
 }
 
 template<ApplyCoeffsT ApplyCoeffs>
-inline void MixHrtfBlendBase(FloatBufferLine &LeftOut, FloatBufferLine &RightOut,
-    const float *InSamples, float2 *RESTRICT AccumSamples, const size_t OutPos,
+inline void MixHrtfBlendBase(const float *InSamples, float2 *RESTRICT AccumSamples,
     const ALuint IrSize, const HrtfFilter *oldparams, MixHrtfFilter *newparams,
     const size_t BufferSize)
 {
@@ -89,11 +82,6 @@ inline void MixHrtfBlendBase(FloatBufferLine &LeftOut, FloatBufferLine &RightOut
         stepcount += 1.0f;
     }
 
-    for(size_t i{0u};i < BufferSize;++i)
-        LeftOut[OutPos+i]  += AccumSamples[i][0];
-    for(size_t i{0u};i < BufferSize;++i)
-        RightOut[OutPos+i] += AccumSamples[i][1];
-
     newparams->Gain = newGainStep*stepcount;
 }
 
@@ -106,9 +94,6 @@ inline void MixDirectHrtfBase(FloatBufferLine &LeftOut, FloatBufferLine &RightOu
 
     const ALuint IrSize{State->IrSize};
 
-    auto accum_iter = std::copy_n(State->Values.begin(), State->Values.size(), AccumSamples);
-    std::fill_n(accum_iter, BufferSize, float2{});
-
     auto coeff_iter = State->Coeffs.begin();
     for(const FloatBufferLine &input : InSamples)
     {
@@ -124,7 +109,11 @@ inline void MixDirectHrtfBase(FloatBufferLine &LeftOut, FloatBufferLine &RightOu
     for(size_t i{0u};i < BufferSize;++i)
         RightOut[i] += AccumSamples[i][1];
 
-    std::copy_n(AccumSamples + BufferSize, State->Values.size(), State->Values.begin());
+    /* Copy the new in-progress accumulation values to the front and clear the
+     * following samples for the next mix.
+     */
+    auto accum_iter = std::copy_n(AccumSamples+BufferSize, HRIR_LENGTH, AccumSamples);
+    std::fill_n(accum_iter, BufferSize, float2{});
 }
 
 #endif /* MIXER_HRTFBASE_H */
diff --git a/alc/mixer/mixer_c.cpp b/alc/mixer/mixer_c.cpp
index 8d375de7..fad33746 100644
--- a/alc/mixer/mixer_c.cpp
+++ b/alc/mixer/mixer_c.cpp
@@ -136,21 +136,16 @@ const ALfloat *Resample_<FastBSincTag,CTag>(const InterpState *state, const ALfl
 
 
 template<>
-void MixHrtf_<CTag>(FloatBufferLine &LeftOut, FloatBufferLine &RightOut,
-    const float *InSamples, float2 *AccumSamples, const size_t OutPos, const ALuint IrSize,
+void MixHrtf_<CTag>(const float *InSamples, float2 *AccumSamples, const ALuint IrSize,
     MixHrtfFilter *hrtfparams, const size_t BufferSize)
-{
-    MixHrtfBase<ApplyCoeffs>(LeftOut, RightOut, InSamples, AccumSamples, OutPos, IrSize,
-        hrtfparams, BufferSize);
-}
+{ MixHrtfBase<ApplyCoeffs>(InSamples, AccumSamples, IrSize, hrtfparams, BufferSize); }
 
 template<>
-void MixHrtfBlend_<CTag>(FloatBufferLine &LeftOut, FloatBufferLine &RightOut,
-    const float *InSamples, float2 *AccumSamples, const size_t OutPos, const ALuint IrSize,
+void MixHrtfBlend_<CTag>(const float *InSamples, float2 *AccumSamples, const ALuint IrSize,
     const HrtfFilter *oldparams, MixHrtfFilter *newparams, const size_t BufferSize)
 {
-    MixHrtfBlendBase<ApplyCoeffs>(LeftOut, RightOut, InSamples, AccumSamples, OutPos, IrSize,
-        oldparams, newparams, BufferSize);
+    MixHrtfBlendBase<ApplyCoeffs>(InSamples, AccumSamples, IrSize, oldparams, newparams,
+        BufferSize);
 }
 
 template<>
diff --git a/alc/mixer/mixer_neon.cpp b/alc/mixer/mixer_neon.cpp
index ae782897..67bf9c71 100644
--- a/alc/mixer/mixer_neon.cpp
+++ b/alc/mixer/mixer_neon.cpp
@@ -190,21 +190,16 @@ const ALfloat *Resample_<FastBSincTag,NEONTag>(const InterpState *state,
 
 
 template<>
-void MixHrtf_<NEONTag>(FloatBufferLine &LeftOut, FloatBufferLine &RightOut,
-    const float *InSamples, float2 *AccumSamples, const size_t OutPos, const ALuint IrSize,
+void MixHrtf_<NEONTag>(const float *InSamples, float2 *AccumSamples, const ALuint IrSize,
     MixHrtfFilter *hrtfparams, const size_t BufferSize)
-{
-    MixHrtfBase<ApplyCoeffs>(LeftOut, RightOut, InSamples, AccumSamples, OutPos, IrSize,
-        hrtfparams, BufferSize);
-}
+{ MixHrtfBase<ApplyCoeffs>(InSamples, AccumSamples, IrSize, hrtfparams, BufferSize); }
 
 template<>
-void MixHrtfBlend_<NEONTag>(FloatBufferLine &LeftOut, FloatBufferLine &RightOut,
-    const float *InSamples, float2 *AccumSamples, const size_t OutPos, const ALuint IrSize,
+void MixHrtfBlend_<NEONTag>(const float *InSamples, float2 *AccumSamples, const ALuint IrSize,
     const HrtfFilter *oldparams, MixHrtfFilter *newparams, const size_t BufferSize)
 {
-    MixHrtfBlendBase<ApplyCoeffs>(LeftOut, RightOut, InSamples, AccumSamples, OutPos, IrSize,
-        oldparams, newparams, BufferSize);
+    MixHrtfBlendBase<ApplyCoeffs>(InSamples, AccumSamples, IrSize, oldparams, newparams,
+        BufferSize);
 }
 
 template<>
diff --git a/alc/mixer/mixer_sse.cpp b/alc/mixer/mixer_sse.cpp
index 62ce5eab..aaf37df6 100644
--- a/alc/mixer/mixer_sse.cpp
+++ b/alc/mixer/mixer_sse.cpp
@@ -165,21 +165,16 @@ const ALfloat *Resample_<FastBSincTag,SSETag>(const InterpState *state,
 
 
 template<>
-void MixHrtf_<SSETag>(FloatBufferLine &LeftOut, FloatBufferLine &RightOut,
-    const float *InSamples, float2 *AccumSamples, const size_t OutPos, const ALuint IrSize,
+void MixHrtf_<SSETag>(const float *InSamples, float2 *AccumSamples, const ALuint IrSize,
     MixHrtfFilter *hrtfparams, const size_t BufferSize)
-{
-    MixHrtfBase<ApplyCoeffs>(LeftOut, RightOut, InSamples, AccumSamples, OutPos, IrSize,
-        hrtfparams, BufferSize);
-}
+{ MixHrtfBase<ApplyCoeffs>(InSamples, AccumSamples, IrSize, hrtfparams, BufferSize); }
 
 template<>
-void MixHrtfBlend_<SSETag>(FloatBufferLine &LeftOut, FloatBufferLine &RightOut,
-    const float *InSamples, float2 *AccumSamples, const size_t OutPos, const ALuint IrSize,
+void MixHrtfBlend_<SSETag>(const float *InSamples, float2 *AccumSamples, const ALuint IrSize,
     const HrtfFilter *oldparams, MixHrtfFilter *newparams, const size_t BufferSize)
 {
-    MixHrtfBlendBase<ApplyCoeffs>(LeftOut, RightOut, InSamples, AccumSamples, OutPos, IrSize,
-        oldparams, newparams, BufferSize);
+    MixHrtfBlendBase<ApplyCoeffs>(InSamples, AccumSamples, IrSize, oldparams, newparams,
+        BufferSize);
 }
 
 template<>
diff --git a/alc/voice.cpp b/alc/voice.cpp
index 0aac0258..1c38f36f 100644
--- a/alc/voice.cpp
+++ b/alc/voice.cpp
@@ -72,12 +72,11 @@ Resampler ResamplerDefault{Resampler::Linear};
 
 namespace {
 
-using HrtfMixerFunc = void(*)(FloatBufferLine &LeftOut, FloatBufferLine &RightOut,
-    const ALfloat *InSamples, float2 *AccumSamples, const size_t OutPos, const ALuint IrSize,
+using HrtfMixerFunc = void(*)(const ALfloat *InSamples, float2 *AccumSamples, const ALuint IrSize,
     MixHrtfFilter *hrtfparams, const size_t BufferSize);
-using HrtfMixerBlendFunc = void(*)(FloatBufferLine &LeftOut, FloatBufferLine &RightOut,
-    const ALfloat *InSamples, float2 *AccumSamples, const size_t OutPos, const ALuint IrSize,
-    const HrtfFilter *oldparams, MixHrtfFilter *newparams, const size_t BufferSize);
+using HrtfMixerBlendFunc = void(*)(const ALfloat *InSamples, float2 *AccumSamples,
+    const ALuint IrSize, const HrtfFilter *oldparams, MixHrtfFilter *newparams,
+    const size_t BufferSize);
 
 HrtfMixerFunc MixHrtfSamples = MixHrtf_<CTag>;
 HrtfMixerBlendFunc MixHrtfBlendSamples = MixHrtfBlend_<CTag>;
@@ -423,12 +422,10 @@ ALfloat *LoadBufferQueue(ALbufferlistitem *BufferListItem, ALbufferlistitem *Buf
 }
 
 
-void DoHrtfMix(ALvoice::TargetData &Direct, const float TargetGain, DirectParams &parms,
-    const float *samples, const ALuint DstBufferSize, const ALuint Counter, const ALuint OutPos,
-    const ALuint IrSize, ALCdevice *Device)
+void DoHrtfMix(const float TargetGain, DirectParams &parms, const float *samples,
+    const ALuint DstBufferSize, const ALuint Counter, ALuint OutPos, const ALuint IrSize,
+    ALCdevice *Device)
 {
-    const ALuint OutLIdx{GetChannelIdxByName(Device->RealOut, FrontLeft)};
-    const ALuint OutRIdx{GetChannelIdxByName(Device->RealOut, FrontRight)};
     auto &HrtfSamples = Device->HrtfSourceData;
     auto &AccumSamples = Device->HrtfAccumData;
 
@@ -440,12 +437,6 @@ void DoHrtfMix(ALvoice::TargetData &Direct, const float TargetGain, DirectParams
     std::copy_n(std::begin(HrtfSamples) + DstBufferSize, parms.Hrtf.State.History.size(),
         parms.Hrtf.State.History.begin());
 
-    /* Copy the current filtered values being accumulated into the temp buffer. */
-    auto accum_iter = std::copy_n(parms.Hrtf.State.Values.begin(), parms.Hrtf.State.Values.size(),
-        std::begin(AccumSamples));
-    /* Clear the accumulation buffer that will start getting filled in. */
-    std::fill_n(accum_iter, DstBufferSize, float2{});
-
     /* If fading, the old gain is not silence, and this is the first mixing
      * pass, fade between the IRs.
      */
@@ -473,14 +464,15 @@ void DoHrtfMix(ALvoice::TargetData &Direct, const float TargetGain, DirectParams
         hrtfparams.Gain = 0.0f;
         hrtfparams.GainStep = gain / static_cast<float>(fademix);
 
-        MixHrtfBlendSamples(Direct.Buffer[OutLIdx], Direct.Buffer[OutRIdx], HrtfSamples,
-            AccumSamples, OutPos, IrSize, &parms.Hrtf.Old, &hrtfparams, fademix);
+        MixHrtfBlendSamples(HrtfSamples, AccumSamples+OutPos, IrSize, &parms.Hrtf.Old, &hrtfparams,
+            fademix);
         /* Update the old parameters with the result. */
         parms.Hrtf.Old = parms.Hrtf.Target;
         if(fademix < Counter)
             parms.Hrtf.Old.Gain = hrtfparams.Gain;
         else
             parms.Hrtf.Old.Gain = TargetGain;
+        OutPos += fademix;
     }
 
     if LIKELY(fademix < DstBufferSize)
@@ -503,8 +495,7 @@ void DoHrtfMix(ALvoice::TargetData &Direct, const float TargetGain, DirectParams
         hrtfparams.Delay[1] = parms.Hrtf.Target.Delay[1];
         hrtfparams.Gain = parms.Hrtf.Old.Gain;
         hrtfparams.GainStep = (gain - parms.Hrtf.Old.Gain) / static_cast<float>(todo);
-        MixHrtfSamples(Direct.Buffer[OutLIdx], Direct.Buffer[OutRIdx], HrtfSamples+fademix,
-            AccumSamples+fademix, OutPos+fademix, IrSize, &hrtfparams, todo);
+        MixHrtfSamples(HrtfSamples+fademix, AccumSamples+OutPos, IrSize, &hrtfparams, todo);
         /* Store the interpolated gain or the final target gain depending if
          * the fade is done.
          */
@@ -513,10 +504,6 @@ void DoHrtfMix(ALvoice::TargetData &Direct, const float TargetGain, DirectParams
         else
             parms.Hrtf.Old.Gain = TargetGain;
     }
-
-    /* Copy the new in-progress accumulation values back for the next mix. */
-    std::copy_n(std::begin(AccumSamples) + DstBufferSize, parms.Hrtf.State.Values.size(),
-        parms.Hrtf.State.Values.begin());
 }
 
 void DoNfcMix(ALvoice::TargetData &Direct, const float *TargetGains, DirectParams &parms,
@@ -713,8 +700,8 @@ void ALvoice::mix(const State vstate, ALCcontext *Context, const ALuint SamplesT
                 {
                     const ALfloat TargetGain{UNLIKELY(vstate == ALvoice::Stopping) ? 0.0f :
                         parms.Hrtf.Target.Gain};
-                    DoHrtfMix(mDirect, TargetGain, parms, samples, DstBufferSize, Counter, OutPos,
-                        IrSize, Device);
+                    DoHrtfMix(TargetGain, parms, samples, DstBufferSize, Counter, OutPos, IrSize,
+                        Device);
                 }
                 else if((mFlags&VOICE_HAS_NFC))
                 {