Use a temporary buffer for HRTF filter accumulation

Similar to the history buffer, to avoid using the state buffer as a ring
buffer.

5 files changed, 122 insertions, 203 deletions

diff --git a/Alc/mixer/defs.h b/Alc/mixer/defs.h
index 1aa6ba38..cd301833 100644
--- a/Alc/mixer/defs.h
+++ b/Alc/mixer/defs.h
@@ -35,11 +35,11 @@ template<typename InstTag>
 void MixRow_(ALfloat *OutBuffer, const ALfloat *Gains, const ALfloat (*data)[BUFFERSIZE], const ALsizei InChans, const ALsizei InPos, const ALsizei BufferSize);
 
 template<typename InstTag>
-void MixHrtf_(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut, const ALfloat *data, ALsizei Offset, const ALsizei OutPos, const ALsizei IrSize, MixHrtfParams *hrtfparams, HrtfState *hrtfstate, const ALsizei BufferSize);
+void MixHrtf_(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut, const ALfloat *data, float2 *RESTRICT AccumSamples, const ALsizei OutPos, const ALsizei IrSize, MixHrtfParams *hrtfparams, const ALsizei BufferSize);
 template<typename InstTag>
-void MixHrtfBlend_(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut, const ALfloat *data, ALsizei Offset, const ALsizei OutPos, const ALsizei IrSize, const HrtfParams *oldparams, MixHrtfParams *newparams, HrtfState *hrtfstate, const ALsizei BufferSize);
+void MixHrtfBlend_(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut, const ALfloat *data, float2 *RESTRICT AccumSamples, const ALsizei OutPos, const ALsizei IrSize, const HrtfParams *oldparams, MixHrtfParams *newparams, const ALsizei BufferSize);
 template<typename InstTag>
-void MixDirectHrtf_(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut, const ALfloat (*data)[BUFFERSIZE], DirectHrtfState *State, const ALsizei NumChans, const ALsizei BufferSize);
+void MixDirectHrtf_(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut, const ALfloat (*data)[BUFFERSIZE], float2 *RESTRICT AccumSamples, DirectHrtfState *State, const ALsizei NumChans, const ALsizei BufferSize);
 
 /* Vectorized resampler helpers */
 inline void InitiatePositionArrays(ALsizei frac, ALint increment, ALsizei *RESTRICT frac_arr, ALsizei *RESTRICT pos_arr, ALsizei size)
diff --git a/Alc/mixer/hrtfbase.h b/Alc/mixer/hrtfbase.h
index 8ad4a99c..162d7289 100644
--- a/Alc/mixer/hrtfbase.h
+++ b/Alc/mixer/hrtfbase.h
@@ -1,18 +1,20 @@
 #ifndef MIXER_HRTFBASE_H
 #define MIXER_HRTFBASE_H
 
+#include <algorithm>
+
 #include "alu.h"
 #include "../hrtf.h"
 #include "opthelpers.h"
 
 
-using ApplyCoeffsT = void(ALsizei Offset, HrirArray<ALfloat> &Values, const ALsizei irSize,
+using ApplyCoeffsT = void(ALsizei Offset, float2 *RESTRICT Values, const ALsizei irSize,
     const HrirArray<ALfloat> &Coeffs, const ALfloat left, const ALfloat right);
 
 template<ApplyCoeffsT &ApplyCoeffs>
 inline void MixHrtfBase(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut, const ALfloat *data,
-    ALsizei Offset, const ALsizei OutPos, const ALsizei IrSize, MixHrtfParams *hrtfparams,
-    HrtfState *hrtfstate, const ALsizei BufferSize)
+    float2 *RESTRICT AccumSamples, const ALsizei OutPos, const ALsizei IrSize,
+    MixHrtfParams *hrtfparams, const ALsizei BufferSize)
 {
     ASSUME(OutPos >= 0);
     ASSUME(IrSize >= 4);
@@ -28,49 +30,27 @@ inline void MixHrtfBase(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut, c
         HRTF_HISTORY_LENGTH - hrtfparams->Delay[1] };
     ASSUME(Delay[0] >= 0 && Delay[1] >= 0);
 
-    Offset &= HRIR_MASK;
-    ALsizei HeadOffset{(Offset+IrSize-1)&HRIR_MASK};
-
-    LeftOut  += OutPos;
-    RightOut += OutPos;
-    for(ALsizei i{0};i < BufferSize;)
+    for(ALsizei i{0};i < BufferSize;++i)
     {
-        /* Calculate the number of samples we can do until one of the indices
-         * wraps on its buffer, or we reach the end.
-         */
-        const ALsizei todo_hrir{HRIR_LENGTH - maxi(HeadOffset, Offset)};
-        const ALsizei todo{mini(BufferSize-i, todo_hrir) + i};
-        ASSUME(todo > i);
-
-        for(;i < todo;++i)
-        {
-            hrtfstate->Values[HeadOffset][0] = 0.0f;
-            hrtfstate->Values[HeadOffset][1] = 0.0f;
-            ++HeadOffset;
-
-            const ALfloat g{gain + gainstep*stepcount};
-            const ALfloat left{data[Delay[0]++] * g};
-            const ALfloat right{data[Delay[1]++] * g};
-            ApplyCoeffs(Offset, hrtfstate->Values, IrSize, Coeffs, left, right);
-
-            *(LeftOut++)  += hrtfstate->Values[Offset][0];
-            *(RightOut++) += hrtfstate->Values[Offset][1];
-            ++Offset;
+        const ALfloat g{gain + gainstep*stepcount};
+        const ALfloat left{data[Delay[0]++] * g};
+        const ALfloat right{data[Delay[1]++] * g};
+        ApplyCoeffs(i, AccumSamples+i, IrSize, Coeffs, left, right);
 
-            stepcount += 1.0f;
-        }
-
-        HeadOffset &= HRIR_MASK;
-        Offset &= HRIR_MASK;
+        stepcount += 1.0f;
     }
+    for(ALsizei i{0};i < BufferSize;++i)
+        LeftOut[OutPos+i]  += AccumSamples[i][0];
+    for(ALsizei i{0};i < BufferSize;++i)
+        RightOut[OutPos+i] += AccumSamples[i][1];
+
     hrtfparams->Gain = gain + gainstep*stepcount;
 }
 
 template<ApplyCoeffsT &ApplyCoeffs>
 inline void MixHrtfBlendBase(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut,
-    const ALfloat *data, ALsizei Offset, const ALsizei OutPos, const ALsizei IrSize,
-    const HrtfParams *oldparams, MixHrtfParams *newparams, HrtfState *hrtfstate,
-    const ALsizei BufferSize)
+    const ALfloat *data, float2 *RESTRICT AccumSamples, const ALsizei OutPos, const ALsizei IrSize,
+    const HrtfParams *oldparams, MixHrtfParams *newparams, const ALsizei BufferSize)
 {
     const auto &OldCoeffs = oldparams->Coeffs;
     const ALfloat oldGain{oldparams->Gain};
@@ -92,50 +72,32 @@ inline void MixHrtfBlendBase(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightO
         HRTF_HISTORY_LENGTH - newparams->Delay[1] };
     ASSUME(NewDelay[0] >= 0 && NewDelay[1] >= 0);
 
-    Offset &= HRIR_MASK;
-    ALsizei HeadOffset{(Offset+IrSize-1)&HRIR_MASK};
-
-    LeftOut  += OutPos;
-    RightOut += OutPos;
-    for(ALsizei i{0};i < BufferSize;)
+    for(ALsizei i{0};i < BufferSize;++i)
     {
-        const ALsizei todo_hrir{HRIR_LENGTH - maxi(HeadOffset, Offset)};
-        const ALsizei todo{mini(BufferSize-i, todo_hrir) + i};
-        ASSUME(todo > i);
+        ALfloat g{oldGain + oldGainStep*stepcount};
+        ALfloat left{data[OldDelay[0]++] * g};
+        ALfloat right{data[OldDelay[1]++] * g};
+        ApplyCoeffs(i, AccumSamples+i, IrSize, OldCoeffs, left, right);
 
-        for(;i < todo;++i)
-        {
-            hrtfstate->Values[HeadOffset][0] = 0.0f;
-            hrtfstate->Values[HeadOffset][1] = 0.0f;
-            ++HeadOffset;
-
-            ALfloat g{oldGain + oldGainStep*stepcount};
-            ALfloat left{data[OldDelay[0]++] * g};
-            ALfloat right{data[OldDelay[1]++] * g};
-            ApplyCoeffs(Offset, hrtfstate->Values, IrSize, OldCoeffs, left, right);
-
-            g = newGainStep*stepcount;
-            left = data[NewDelay[0]++] * g;
-            right = data[NewDelay[1]++] * g;
-            ApplyCoeffs(Offset, hrtfstate->Values, IrSize, NewCoeffs, left, right);
+        g = newGainStep*stepcount;
+        left = data[NewDelay[0]++] * g;
+        right = data[NewDelay[1]++] * g;
+        ApplyCoeffs(i, AccumSamples+i, IrSize, NewCoeffs, left, right);
 
-            *(LeftOut++)  += hrtfstate->Values[Offset][0];
-            *(RightOut++) += hrtfstate->Values[Offset][1];
-            ++Offset;
-
-            stepcount += 1.0f;
-        }
-
-        HeadOffset &= HRIR_MASK;
-        Offset &= HRIR_MASK;
+        stepcount += 1.0f;
     }
+    for(ALsizei i{0};i < BufferSize;++i)
+        LeftOut[OutPos+i]  += AccumSamples[i][0];
+    for(ALsizei i{0};i < BufferSize;++i)
+        RightOut[OutPos+i] += AccumSamples[i][1];
+
     newparams->Gain = newGainStep*stepcount;
 }
 
 template<ApplyCoeffsT &ApplyCoeffs>
 inline void MixDirectHrtfBase(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut,
-    const ALfloat (*data)[BUFFERSIZE], DirectHrtfState *State, const ALsizei NumChans,
-    const ALsizei BufferSize)
+    const ALfloat (*data)[BUFFERSIZE], float2 *RESTRICT AccumSamples, DirectHrtfState *State,
+    const ALsizei NumChans, const ALsizei BufferSize)
 {
     ASSUME(NumChans > 0);
     ASSUME(BufferSize > 0);
@@ -147,32 +109,23 @@ inline void MixDirectHrtfBase(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT Right
     {
         const ALfloat (&input)[BUFFERSIZE] = data[c];
         const auto &Coeffs = State->Chan[c].Coeffs;
-        auto &Values = State->Chan[c].Values;
-        ALsizei Offset{State->Offset&HRIR_MASK};
 
-        ALsizei HeadOffset{(Offset+IrSize-1)&HRIR_MASK};
-        for(ALsizei i{0};i < BufferSize;)
+        auto accum_iter = std::copy_n(State->Chan[c].Values.begin(),
+            State->Chan[c].Values.size(), AccumSamples);
+        std::fill_n(accum_iter, BufferSize, float2{});
+
+        for(ALsizei i{0};i < BufferSize;++i)
         {
-            const ALsizei todo_hrir{HRIR_LENGTH - maxi(HeadOffset, Offset)};
-            const ALsizei todo{mini(BufferSize-i, todo_hrir) + i};
-            ASSUME(todo > i);
-
-            for(;i < todo;++i)
-            {
-                Values[HeadOffset][0] = 0.0f;
-                Values[HeadOffset][1] = 0.0f;
-                ++HeadOffset;
-
-                const ALfloat insample{input[i]};
-                ApplyCoeffs(Offset, Values, IrSize, Coeffs, insample, insample);
-
-                LeftOut[i]  += Values[Offset][0];
-                RightOut[i] += Values[Offset][1];
-                ++Offset;
-            }
-            HeadOffset &= HRIR_MASK;
-            Offset &= HRIR_MASK;
+            const ALfloat insample{input[i]};
+            ApplyCoeffs(i, AccumSamples+i, IrSize, Coeffs, insample, insample);
         }
+        for(ALsizei i{0};i < BufferSize;++i)
+            LeftOut[i]  += AccumSamples[i][0];
+        for(ALsizei i{0};i < BufferSize;++i)
+            RightOut[i] += AccumSamples[i][1];
+
+        std::copy_n(AccumSamples + BufferSize, State->Chan[c].Values.size(),
+            State->Chan[c].Values.begin());
     }
 }
 
diff --git a/Alc/mixer/mixer_c.cpp b/Alc/mixer/mixer_c.cpp
index ba094999..1c22115d 100644
--- a/Alc/mixer/mixer_c.cpp
+++ b/Alc/mixer/mixer_c.cpp
@@ -103,54 +103,43 @@ const ALfloat *Resample_<BSincTag,CTag>(const InterpState *state, const ALfloat
 { return DoResample<do_bsinc>(state, src-state->bsinc.l, frac, increment, dst, dstlen); }
 
 
-static inline void ApplyCoeffs(ALsizei Offset, HrirArray<ALfloat> &Values, const ALsizei IrSize,
+static inline void ApplyCoeffs(ALsizei /*Offset*/, float2 *RESTRICT Values, const ALsizei IrSize,
     const HrirArray<ALfloat> &Coeffs, const ALfloat left, const ALfloat right)
 {
-    ASSUME(Offset >= 0 && Offset < HRIR_LENGTH);
     ASSUME(IrSize >= 2);
-    ASSUME(&Values != &Coeffs);
-
-    ALsizei count{mini(IrSize, HRIR_LENGTH - Offset)};
-    ASSUME(count > 0);
-    for(ALsizei c{0};;)
+    for(ALsizei c{0};c < IrSize;++c)
     {
-        for(;c < count;++c)
-        {
-            Values[Offset][0] += Coeffs[c][0] * left;
-            Values[Offset][1] += Coeffs[c][1] * right;
-            ++Offset;
-        }
-        if(c >= IrSize)
-            break;
-        Offset = 0;
-        count = IrSize;
+        Values[c][0] += Coeffs[c][0] * left;
+        Values[c][1] += Coeffs[c][1] * right;
     }
 }
 
 template<>
 void MixHrtf_<CTag>(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut, const ALfloat *data,
-    ALsizei Offset, const ALsizei OutPos, const ALsizei IrSize, MixHrtfParams *hrtfparams,
-    HrtfState *hrtfstate, const ALsizei BufferSize)
+    float2 *RESTRICT AccumSamples, const ALsizei OutPos, const ALsizei IrSize,
+    MixHrtfParams *hrtfparams, const ALsizei BufferSize)
 {
-    MixHrtfBase<ApplyCoeffs>(LeftOut, RightOut, data, Offset, OutPos, IrSize, hrtfparams,
-        hrtfstate, BufferSize);
+    MixHrtfBase<ApplyCoeffs>(LeftOut, RightOut, data, AccumSamples, OutPos, IrSize, hrtfparams,
+        BufferSize);
 }
 
 template<>
 void MixHrtfBlend_<CTag>(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut,
-    const ALfloat *data, ALsizei Offset, const ALsizei OutPos, const ALsizei IrSize,
-    const HrtfParams *oldparams, MixHrtfParams *newparams, HrtfState *hrtfstate,
-    const ALsizei BufferSize)
+    const ALfloat *data, float2 *RESTRICT AccumSamples, const ALsizei OutPos, const ALsizei IrSize,
+    const HrtfParams *oldparams, MixHrtfParams *newparams, const ALsizei BufferSize)
 {
-    MixHrtfBlendBase<ApplyCoeffs>(LeftOut, RightOut, data, Offset, OutPos, IrSize, oldparams,
-        newparams, hrtfstate, BufferSize);
+    MixHrtfBlendBase<ApplyCoeffs>(LeftOut, RightOut, data, AccumSamples, OutPos, IrSize, oldparams,
+        newparams, BufferSize);
 }
 
 template<>
 void MixDirectHrtf_<CTag>(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut,
-    const ALfloat (*data)[BUFFERSIZE], DirectHrtfState *State, const ALsizei NumChans,
-    const ALsizei BufferSize)
-{ MixDirectHrtfBase<ApplyCoeffs>(LeftOut, RightOut, data, State, NumChans, BufferSize); }
+    const ALfloat (*data)[BUFFERSIZE], float2 *RESTRICT AccumSamples, DirectHrtfState *State,
+    const ALsizei NumChans, const ALsizei BufferSize)
+{
+    MixDirectHrtfBase<ApplyCoeffs>(LeftOut, RightOut, data, AccumSamples, State, NumChans,
+        BufferSize);
+}
 
 
 template<>
diff --git a/Alc/mixer/mixer_neon.cpp b/Alc/mixer/mixer_neon.cpp
index 9e8324aa..cdd96296 100644
--- a/Alc/mixer/mixer_neon.cpp
+++ b/Alc/mixer/mixer_neon.cpp
@@ -136,11 +136,10 @@ const ALfloat *Resample_<BSincTag,NEONTag>(const InterpState *state, const ALflo
 }
 
 
-static inline void ApplyCoeffs(ALsizei Offset, HrirArray<ALfloat> &Values, const ALsizei IrSize,
+static inline void ApplyCoeffs(ALsizei /*Offset*/, float2 *RESTRICT Values, const ALsizei IrSize,
     const HrirArray<ALfloat> &Coeffs, const ALfloat left, const ALfloat right)
 {
     ASSUME(IrSize >= 2);
-    ASSUME(&Values != &Coeffs);
 
     float32x4_t leftright4;
     {
@@ -152,43 +151,43 @@ static inline void ApplyCoeffs(ALsizei Offset, HrirArray<ALfloat> &Values, const
 
     for(ALsizei c{0};c < IrSize;c += 2)
     {
-        const ALsizei o0 = (Offset+c)&HRIR_MASK;
-        const ALsizei o1 = (o0+1)&HRIR_MASK;
-        float32x4_t vals = vcombine_f32(vld1_f32((float32_t*)&Values[o0][0]),
-                                        vld1_f32((float32_t*)&Values[o1][0]));
+        float32x4_t vals = vcombine_f32(vld1_f32((float32_t*)&Values[c  ][0]),
+                                        vld1_f32((float32_t*)&Values[c+1][0]));
         float32x4_t coefs = vld1q_f32((float32_t*)&Coeffs[c][0]);
 
         vals = vmlaq_f32(vals, coefs, leftright4);
 
-        vst1_f32((float32_t*)&Values[o0][0], vget_low_f32(vals));
-        vst1_f32((float32_t*)&Values[o1][0], vget_high_f32(vals));
+        vst1_f32((float32_t*)&Values[c  ][0], vget_low_f32(vals));
+        vst1_f32((float32_t*)&Values[c+1][0], vget_high_f32(vals));
     }
 }
 
 template<>
 void MixHrtf_<NEONTag>(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut, const ALfloat *data,
-    ALsizei Offset, const ALsizei OutPos, const ALsizei IrSize, MixHrtfParams *hrtfparams,
-    HrtfState *hrtfstate, const ALsizei BufferSize)
+    float2 *RESTRICT AccumSamples, const ALsizei OutPos, const ALsizei IrSize,
+    MixHrtfParams *hrtfparams, const ALsizei BufferSize)
 {
-    MixHrtfBase<ApplyCoeffs>(LeftOut, RightOut, data, Offset, OutPos, IrSize, hrtfparams,
-        hrtfstate, BufferSize);
+    MixHrtfBase<ApplyCoeffs>(LeftOut, RightOut, data, AccumSamples, OutPos, IrSize, hrtfparams,
+        BufferSize);
 }
 
 template<>
 void MixHrtfBlend_<NEONTag>(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut,
-    const ALfloat *data, ALsizei Offset, const ALsizei OutPos, const ALsizei IrSize,
-    const HrtfParams *oldparams, MixHrtfParams *newparams, HrtfState *hrtfstate,
-    const ALsizei BufferSize)
+    const ALfloat *data, float2 *RESTRICT AccumSamples, const ALsizei OutPos, const ALsizei IrSize,
+    const HrtfParams *oldparams, MixHrtfParams *newparams, const ALsizei BufferSize)
 {
-    MixHrtfBlendBase<ApplyCoeffs>(LeftOut, RightOut, data, Offset, OutPos, IrSize, oldparams,
-        newparams, hrtfstate, BufferSize);
+    MixHrtfBlendBase<ApplyCoeffs>(LeftOut, RightOut, data, AccumSamples, OutPos, IrSize, oldparams,
+        newparams, BufferSize);
 }
 
 template<>
 void MixDirectHrtf_<NEONTag>(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut,
-    const ALfloat (*data)[BUFFERSIZE], DirectHrtfState *State, const ALsizei NumChans,
-    const ALsizei BufferSize)
-{ MixDirectHrtfBase<ApplyCoeffs>(LeftOut, RightOut, data, State, NumChans, BufferSize); }
+    const ALfloat (*data)[BUFFERSIZE], float2 *RESTRICT AccumSamples, DirectHrtfState *State,
+    const ALsizei NumChans, const ALsizei BufferSize)
+{
+    MixDirectHrtfBase<ApplyCoeffs>(LeftOut, RightOut, data, AccumSamples, State, NumChans,
+        BufferSize);
+}
 
 
 template<>
diff --git a/Alc/mixer/mixer_sse.cpp b/Alc/mixer/mixer_sse.cpp
index 69aeaa62..629fa428 100644
--- a/Alc/mixer/mixer_sse.cpp
+++ b/Alc/mixer/mixer_sse.cpp
@@ -76,97 +76,75 @@ const ALfloat *Resample_<BSincTag,SSETag>(const InterpState *state, const ALfloa
 }
 
 
-static inline void ApplyCoeffs(ALsizei Offset, HrirArray<ALfloat> &Values, const ALsizei IrSize,
+static inline void ApplyCoeffs(ALsizei Offset, float2 *RESTRICT Values, const ALsizei IrSize,
     const HrirArray<ALfloat> &Coeffs, const ALfloat left, const ALfloat right)
 {
     const __m128 lrlr{_mm_setr_ps(left, right, left, right)};
 
     ASSUME(IrSize >= 2);
-    ASSUME(&Values != &Coeffs);
 
-    ASSUME(Offset >= 0 && Offset < HRIR_LENGTH);
     if((Offset&1))
     {
-        ALsizei count{mini(IrSize-1, HRIR_LENGTH - Offset)};
-        ASSUME(count >= 1);
-
         __m128 imp0, imp1;
         __m128 coeffs{_mm_load_ps(&Coeffs[0][0])};
-        __m128 vals{_mm_loadl_pi(_mm_setzero_ps(), reinterpret_cast<__m64*>(&Values[Offset][0]))};
+        __m128 vals{_mm_loadl_pi(_mm_setzero_ps(), reinterpret_cast<__m64*>(&Values[0][0]))};
         imp0 = _mm_mul_ps(lrlr, coeffs);
         vals = _mm_add_ps(imp0, vals);
-        _mm_storel_pi(reinterpret_cast<__m64*>(&Values[Offset][0]), vals);
-        ++Offset;
-        for(ALsizei i{1};;)
+        _mm_storel_pi(reinterpret_cast<__m64*>(&Values[0][0]), vals);
+        ALsizei i{1};
+        for(;i < IrSize-1;i += 2)
         {
-            for(;i < count;i += 2)
-            {
-                coeffs = _mm_load_ps(&Coeffs[i+1][0]);
-                vals = _mm_load_ps(&Values[Offset][0]);
-                imp1 = _mm_mul_ps(lrlr, coeffs);
-                imp0 = _mm_shuffle_ps(imp0, imp1, _MM_SHUFFLE(1, 0, 3, 2));
-                vals = _mm_add_ps(imp0, vals);
-                _mm_store_ps(&Values[Offset][0], vals);
-                imp0 = imp1;
-                Offset += 2;
-            }
-            Offset &= HRIR_MASK;
-            if(i >= IrSize-1)
-                break;
-            count = IrSize-1;
+            coeffs = _mm_load_ps(&Coeffs[i+1][0]);
+            vals = _mm_load_ps(&Values[i][0]);
+            imp1 = _mm_mul_ps(lrlr, coeffs);
+            imp0 = _mm_shuffle_ps(imp0, imp1, _MM_SHUFFLE(1, 0, 3, 2));
+            vals = _mm_add_ps(imp0, vals);
+            _mm_store_ps(&Values[i][0], vals);
+            imp0 = imp1;
         }
-        vals = _mm_loadl_pi(vals, reinterpret_cast<__m64*>(&Values[Offset][0]));
+        vals = _mm_loadl_pi(vals, reinterpret_cast<__m64*>(&Values[i][0]));
         imp0 = _mm_movehl_ps(imp0, imp0);
         vals = _mm_add_ps(imp0, vals);
-        _mm_storel_pi(reinterpret_cast<__m64*>(&Values[Offset][0]), vals);
+        _mm_storel_pi(reinterpret_cast<__m64*>(&Values[i][0]), vals);
     }
     else
     {
-        ALsizei count{mini(IrSize, HRIR_LENGTH - Offset)};
-        ASSUME(count >= 2);
-
-        for(ALsizei i{0};;)
+        for(ALsizei i{0};i < IrSize;i += 2)
         {
-            for(;i < count;i += 2)
-            {
-                __m128 coeffs{_mm_load_ps(&Coeffs[i][0])};
-                __m128 vals{_mm_load_ps(&Values[Offset][0])};
-                vals = _mm_add_ps(vals, _mm_mul_ps(lrlr, coeffs));
-                _mm_store_ps(&Values[Offset][0], vals);
-                Offset += 2;
-            }
-            if(i >= IrSize)
-                break;
-            Offset = 0;
-            count = IrSize;
+            __m128 coeffs{_mm_load_ps(&Coeffs[i][0])};
+            __m128 vals{_mm_load_ps(&Values[i][0])};
+            vals = _mm_add_ps(vals, _mm_mul_ps(lrlr, coeffs));
+            _mm_store_ps(&Values[i][0], vals);
         }
     }
 }
 
 template<>
 void MixHrtf_<SSETag>(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut, const ALfloat *data,
-    ALsizei Offset, const ALsizei OutPos, const ALsizei IrSize, MixHrtfParams *hrtfparams,
-    HrtfState *hrtfstate, const ALsizei BufferSize)
+    float2 *RESTRICT AccumSamples, const ALsizei OutPos, const ALsizei IrSize,
+    MixHrtfParams *hrtfparams, const ALsizei BufferSize)
 {
-    MixHrtfBase<ApplyCoeffs>(LeftOut, RightOut, data, Offset, OutPos, IrSize, hrtfparams,
-        hrtfstate, BufferSize);
+    MixHrtfBase<ApplyCoeffs>(LeftOut, RightOut, data, AccumSamples, OutPos, IrSize, hrtfparams,
+        BufferSize);
 }
 
 template<>
 void MixHrtfBlend_<SSETag>(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut,
-    const ALfloat *data, ALsizei Offset, const ALsizei OutPos, const ALsizei IrSize,
-    const HrtfParams *oldparams, MixHrtfParams *newparams, HrtfState *hrtfstate,
-    const ALsizei BufferSize)
+    const ALfloat *data, float2 *RESTRICT AccumSamples, const ALsizei OutPos, const ALsizei IrSize,
+    const HrtfParams *oldparams, MixHrtfParams *newparams, const ALsizei BufferSize)
 {
-    MixHrtfBlendBase<ApplyCoeffs>(LeftOut, RightOut, data, Offset, OutPos, IrSize, oldparams,
-        newparams, hrtfstate, BufferSize);
+    MixHrtfBlendBase<ApplyCoeffs>(LeftOut, RightOut, data, AccumSamples, OutPos, IrSize, oldparams,
+        newparams, BufferSize);
 }
 
 template<>
 void MixDirectHrtf_<SSETag>(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut,
-    const ALfloat (*data)[BUFFERSIZE], DirectHrtfState *State, const ALsizei NumChans,
-    const ALsizei BufferSize)
-{ MixDirectHrtfBase<ApplyCoeffs>(LeftOut, RightOut, data, State, NumChans, BufferSize); }
+    const ALfloat (*data)[BUFFERSIZE], float2 *RESTRICT AccumSamples, DirectHrtfState *State,
+    const ALsizei NumChans, const ALsizei BufferSize)
+{
+    MixDirectHrtfBase<ApplyCoeffs>(LeftOut, RightOut, data, AccumSamples, State, NumChans,
+        BufferSize);
+}
 
 
 template<>