Improve HRTF delay handling

Ensures source-level HRTF, the dry mix, and direct output all align properly,
and simplifies adding the delay in the direct mix output.

4 files changed, 19 insertions, 32 deletions

diff --git a/alc/alcmain.h b/alc/alcmain.h
index 3a14e68f..c577551b 100644
--- a/alc/alcmain.h
+++ b/alc/alcmain.h
@@ -269,7 +269,7 @@ struct ALCdevice : public al::intrusive_ref<ALCdevice> {
     };
 
     /* Persistent storage for HRTF mixing. */
-    alignas(16) float2 HrtfAccumData[BUFFERSIZE + HRIR_LENGTH];
+    alignas(16) float2 HrtfAccumData[BUFFERSIZE + HRIR_LENGTH + HRTF_DIRECT_DELAY];
 
     /* Mixing buffer used by the Dry mix and Real output. */
     al::vector<FloatBufferLine, 16> MixBuffer;
diff --git a/alc/hrtf.h b/alc/hrtf.h
index 7876265b..29a0864c 100644
--- a/alc/hrtf.h
+++ b/alc/hrtf.h
@@ -88,8 +88,6 @@ struct DirectHrtfState {
         alignas(16) HrirArray mCoeffs{};
     };
 
-    std::array<float,HRTF_DIRECT_DELAY> mLeftDelay{};
-    std::array<float,HRTF_DIRECT_DELAY> mRightDelay{};
     std::array<float,HRTF_DIRECT_DELAY+BUFFERSIZE> mTemp;
 
     /* HRTF filter state for dry buffer content */
diff --git a/alc/mixer/hrtfbase.h b/alc/mixer/hrtfbase.h
index 59ff7633..8db26d70 100644
--- a/alc/mixer/hrtfbase.h
+++ b/alc/mixer/hrtfbase.h
@@ -85,8 +85,16 @@ inline void MixDirectHrtfBase(FloatBufferLine &LeftOut, FloatBufferLine &RightOu
 {
     ASSUME(BufferSize > 0);
 
-    const uint_fast32_t IrSize{State->mIrSize};
+    /* Add the existing signal directly to the accumulation buffer, unfiltered,
+     * and with a delay to align with the input delay.
+     */
+    for(size_t i{0};i < BufferSize;++i)
+    {
+        AccumSamples[HRTF_DIRECT_DELAY+i][0] += LeftOut[i];
+        AccumSamples[HRTF_DIRECT_DELAY+i][1] += RightOut[i];
+    }
 
+    const uint_fast32_t IrSize{State->mIrSize};
     auto chan_iter = State->mChannels.begin();
     for(const FloatBufferLine &input : InSamples)
     {
@@ -123,47 +131,25 @@ inline void MixDirectHrtfBase(FloatBufferLine &LeftOut, FloatBufferLine &RightOu
 
         /* Now apply the HRIR coefficients to this channel. */
         const auto &Coeffs = chan_iter->mCoeffs;
-        ++chan_iter;
-
         for(size_t i{0u};i < BufferSize;++i)
         {
             const float insample{tempbuf[i]};
             ApplyCoeffs(AccumSamples+i, IrSize, Coeffs, insample, insample);
         }
-    }
-
-    /* Apply a delay to the existing signal to align with the input delay. */
-    auto &ldelay = State->mLeftDelay;
-    auto &rdelay = State->mRightDelay;
-    if LIKELY(BufferSize >= HRTF_DIRECT_DELAY)
-    {
-        auto buffer_end = LeftOut.begin() + BufferSize;
-        auto delay_end = std::rotate(LeftOut.begin(), buffer_end - HRTF_DIRECT_DELAY, buffer_end);
-        std::swap_ranges(LeftOut.begin(), delay_end, ldelay.begin());
 
-        buffer_end = RightOut.begin() + BufferSize;
-        delay_end = std::rotate(RightOut.begin(), buffer_end - HRTF_DIRECT_DELAY, buffer_end);
-        std::swap_ranges(RightOut.begin(), delay_end, rdelay.begin());
+        ++chan_iter;
     }
-    else
-    {
-        auto buffer_end = LeftOut.begin() + BufferSize;
-        auto delay_start = std::swap_ranges(LeftOut.begin(), buffer_end, ldelay.begin());
-        std::rotate(ldelay.begin(), delay_start, ldelay.end());
 
-        buffer_end = RightOut.begin() + BufferSize;
-        delay_start = std::swap_ranges(RightOut.begin(), buffer_end, rdelay.begin());
-        std::rotate(rdelay.begin(), delay_start, rdelay.end());
-    }
     for(size_t i{0u};i < BufferSize;++i)
-        LeftOut[i]  += AccumSamples[i][0];
+        LeftOut[i]  = AccumSamples[i][0];
     for(size_t i{0u};i < BufferSize;++i)
-        RightOut[i] += AccumSamples[i][1];
+        RightOut[i] = AccumSamples[i][1];
 
     /* 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);
+    auto accum_iter = std::copy_n(AccumSamples+BufferSize, HRIR_LENGTH+HRTF_DIRECT_DELAY,
+        AccumSamples);
     std::fill_n(accum_iter, BufferSize, float2{});
 }
 
diff --git a/alc/voice.cpp b/alc/voice.cpp
index 0e92e423..7d4f638c 100644
--- a/alc/voice.cpp
+++ b/alc/voice.cpp
@@ -468,7 +468,10 @@ void DoHrtfMix(const float *samples, const ALuint DstBufferSize, DirectParams &p
     ALCdevice *Device)
 {
     auto &HrtfSamples = Device->HrtfSourceData;
-    auto &AccumSamples = Device->HrtfAccumData;
+    /* Source HRTF mixing needs to include the direct delay so it remains
+     * aligned with the direct mix's HRTF filtering.
+     */
+    float2 *AccumSamples{Device->HrtfAccumData + HRTF_DIRECT_DELAY};
 
     /* Copy the HRTF history and new input samples into a temp buffer. */
     auto src_iter = std::copy(parms.Hrtf.History.begin(), parms.Hrtf.History.end(),