Make and use a processHfScale function with a separate output

3 files changed, 44 insertions, 6 deletions

diff --git a/core/filters/splitter.cpp b/core/filters/splitter.cpp
index 5ffe51e1..19f7646a 100644
--- a/core/filters/splitter.cpp
+++ b/core/filters/splitter.cpp
@@ -61,6 +61,41 @@ void BandSplitterR<Real>::process(const al::span<const Real> input, Real *hpout,
 }
 
 template<typename Real>
+void BandSplitterR<Real>::processHfScale(const al::span<const Real> input, Real *RESTRICT output,
+    const Real hfscale)
+{
+    const Real ap_coeff{mCoeff};
+    const Real lp_coeff{mCoeff*0.5f + 0.5f};
+    Real lp_z1{mLpZ1};
+    Real lp_z2{mLpZ2};
+    Real ap_z1{mApZ1};
+    auto proc_sample = [hfscale,ap_coeff,lp_coeff,&lp_z1,&lp_z2,&ap_z1](const Real in) noexcept -> Real
+    {
+        /* Low-pass sample processing. */
+        Real d{(in - lp_z1) * lp_coeff};
+        Real lp_y{lp_z1 + d};
+        lp_z1 = lp_y + d;
+
+        d = (lp_y - lp_z2) * lp_coeff;
+        lp_y = lp_z2 + d;
+        lp_z2 = lp_y + d;
+
+        /* All-pass sample processing. */
+        Real ap_y{in*ap_coeff + ap_z1};
+        ap_z1 = in - ap_y*ap_coeff;
+
+        /* High-pass generated by removing the low-passed signal, which is then
+         * scaled and added back to the low-passed signal.
+         */
+        return (ap_y-lp_y)*hfscale + lp_y;
+    };
+    std::transform(input.begin(), input.end(), output, proc_sample);
+    mLpZ1 = lp_z1;
+    mLpZ2 = lp_z2;
+    mApZ1 = ap_z1;
+}
+
+template<typename Real>
 void BandSplitterR<Real>::processHfScale(const al::span<Real> samples, const Real hfscale)
 {
     const Real ap_coeff{mCoeff};
diff --git a/core/filters/splitter.h b/core/filters/splitter.h
index e428418f..41805c70 100644
--- a/core/filters/splitter.h
+++ b/core/filters/splitter.h
@@ -24,6 +24,8 @@ public:
     void clear() noexcept { mLpZ1 = mLpZ2 = mApZ1 = 0.0f; }
     void process(const al::span<const Real> input, Real *hpout, Real *lpout);
 
+    void processHfScale(const al::span<const Real> input, Real *output, const Real hfscale);
+
     void processHfScale(const al::span<Real> samples, const Real hfscale);
     void processScale(const al::span<Real> samples, const Real hfscale, const Real lfscale);
 
diff --git a/core/mixer/hrtfbase.h b/core/mixer/hrtfbase.h
index ade6f693..606f9d4e 100644
--- a/core/mixer/hrtfbase.h
+++ b/core/mixer/hrtfbase.h
@@ -96,12 +96,11 @@ inline void MixDirectHrtfBase(const FloatBufferSpan LeftOut, const FloatBufferSp
          * the high frequency response. The band-splitter applies this scaling
          * with a consistent phase shift regardless of the scale amount.
          */
-        al::span<float> tempbuf{al::assume_aligned<16>(TempBuf), BufferSize};
-        std::copy(input.begin(), input.begin()+BufferSize, tempbuf.begin());
-
-        ChanState->mSplitter.processHfScale(tempbuf, ChanState->mHfScale);
+        ChanState->mSplitter.processHfScale({input.data(), BufferSize}, TempBuf,
+            ChanState->mHfScale);
 
         /* Now apply the HRIR coefficients to this channel. */
+        const float *RESTRICT tempbuf{al::assume_aligned<16>(TempBuf)};
         const ConstHrirSpan Coeffs{ChanState->mCoeffs};
         for(size_t i{0u};i < BufferSize;++i)
         {
@@ -113,10 +112,12 @@ inline void MixDirectHrtfBase(const FloatBufferSpan LeftOut, const FloatBufferSp
     }
 
     /* Add the HRTF signal to the existing "direct" signal. */
+    float *RESTRICT left{al::assume_aligned<16>(LeftOut.data())};
+    float *RESTRICT right{al::assume_aligned<16>(RightOut.data())};
     for(size_t i{0u};i < BufferSize;++i)
-        LeftOut[i]  += AccumSamples[i][0];
+        left[i]  += AccumSamples[i][0];
     for(size_t i{0u};i < BufferSize;++i)
-        RightOut[i] += AccumSamples[i][1];
+        right[i] += AccumSamples[i][1];
 
     /* Copy the new in-progress accumulation values to the front and clear the
      * following samples for the next mix.