Further improve HRTF methods to avoid masking in the inner loops

4 files changed, 163 insertions, 100 deletions

diff --git a/Alc/mixer/hrtf_inc.cpp b/Alc/mixer/hrtf_inc.cpp
index e82bad85..caac7e54 100644
--- a/Alc/mixer/hrtf_inc.cpp
+++ b/Alc/mixer/hrtf_inc.cpp
@@ -14,114 +14,178 @@ static inline void ApplyCoeffs(ALsizei Offset, ALfloat (&Values)[HRIR_LENGTH][2]
 
 
 void MixHrtf(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut,
-             const ALfloat *data, ALsizei Offset, ALsizei OutPos,
+             const ALfloat *data, ALsizei Offset, const ALsizei OutPos,
              const ALsizei IrSize, MixHrtfParams *hrtfparams, HrtfState *hrtfstate,
-             ALsizei BufferSize)
+             const ALsizei BufferSize)
 {
-    const ALfloat (&Coeffs)[HRIR_LENGTH][2] = *hrtfparams->Coeffs;
-    const ALsizei Delay[2] = { hrtfparams->Delay[0], hrtfparams->Delay[1] };
-    const ALfloat gainstep = hrtfparams->GainStep;
-    const ALfloat gain = hrtfparams->Gain;
-    ALfloat g, stepcount = 0.0f;
-    ALfloat left, right;
-    ALsizei i;
-
+    ASSUME(OutPos >= 0);
     ASSUME(IrSize >= 4);
     ASSUME(BufferSize > 0);
 
-    LeftOut  += OutPos;
-    RightOut += OutPos;
-    for(i = 0;i < BufferSize;i++)
-    {
-        hrtfstate->History[Offset&HRTF_HISTORY_MASK] = *(data++);
-
-        g = gain + gainstep*stepcount;
-        left = hrtfstate->History[(Offset-Delay[0])&HRTF_HISTORY_MASK]*g;
-        right = hrtfstate->History[(Offset-Delay[1])&HRTF_HISTORY_MASK]*g;
+    const ALfloat (&Coeffs)[HRIR_LENGTH][2] = *hrtfparams->Coeffs;
+    const ALfloat gainstep{hrtfparams->GainStep};
+    const ALfloat gain{hrtfparams->Gain};
+    ALfloat stepcount{0.0f};
 
-        hrtfstate->Values[(Offset+IrSize-1)&HRIR_MASK][0] = 0.0f;
-        hrtfstate->Values[(Offset+IrSize-1)&HRIR_MASK][1] = 0.0f;
+    ALsizei HistOffset{Offset&HRTF_HISTORY_MASK};
+    ALsizei Delay[2]{
+        (HistOffset-hrtfparams->Delay[0])&HRTF_HISTORY_MASK,
+        (HistOffset-hrtfparams->Delay[1])&HRTF_HISTORY_MASK };
 
-        ApplyCoeffs(Offset, hrtfstate->Values, IrSize, Coeffs, left, right);
-        *(LeftOut++)  += hrtfstate->Values[Offset&HRIR_MASK][0];
-        *(RightOut++) += hrtfstate->Values[Offset&HRIR_MASK][1];
+    Offset &= HRIR_MASK;
+    ALsizei HeadOffset{(Offset+IrSize-1)&HRIR_MASK};
 
-        stepcount += 1.0f;
-        Offset++;
+    LeftOut  += OutPos;
+    RightOut += OutPos;
+    for(ALsizei i{0};i < BufferSize;)
+    {
+        /* 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_hist{HRTF_HISTORY_LENGTH - maxi(maxi(HistOffset, Delay[0]), Delay[1])};
+        const ALsizei todo_hrir{HRIR_LENGTH - maxi(HeadOffset, Offset)};
+        const ALsizei todo{mini(BufferSize-i, mini(todo_hist, todo_hrir)) + i};
+        ASSUME(todo > i);
+
+        for(;i < todo;++i)
+        {
+            hrtfstate->Values[HeadOffset][0] = 0.0f;
+            hrtfstate->Values[HeadOffset][1] = 0.0f;
+            ++HeadOffset;
+
+            hrtfstate->History[HistOffset++] = *(data++);
+
+            const ALfloat g{gain + gainstep*stepcount};
+            const ALfloat left{hrtfstate->History[Delay[0]++] * g};
+            const ALfloat right{hrtfstate->History[Delay[1]++] * g};
+            ApplyCoeffs(Offset, hrtfstate->Values, IrSize, Coeffs, left, right);
+
+            *(LeftOut++)  += hrtfstate->Values[Offset][0];
+            *(RightOut++) += hrtfstate->Values[Offset][1];
+            ++Offset;
+
+            stepcount += 1.0f;
+        }
+
+        HeadOffset &= HRIR_MASK;
+        HistOffset &= HRTF_HISTORY_MASK;
+        Delay[0] &= HRTF_HISTORY_MASK;
+        Delay[1] &= HRTF_HISTORY_MASK;
+        Offset &= HRIR_MASK;
     }
     hrtfparams->Gain = gain + gainstep*stepcount;
 }
 
 void MixHrtfBlend(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut,
-                  const ALfloat *data, ALsizei Offset, ALsizei OutPos,
+                  const ALfloat *data, ALsizei Offset, const ALsizei OutPos,
                   const ALsizei IrSize, const HrtfParams *oldparams,
                   MixHrtfParams *newparams, HrtfState *hrtfstate,
-                  ALsizei BufferSize)
+                  const ALsizei BufferSize)
 {
     const ALfloat (&OldCoeffs)[HRIR_LENGTH][2] = oldparams->Coeffs;
-    const ALsizei OldDelay[2] = { oldparams->Delay[0], oldparams->Delay[1] };
-    const ALfloat oldGain = oldparams->Gain;
-    const ALfloat oldGainStep = -oldGain / (ALfloat)BufferSize;
+    const ALfloat oldGain{oldparams->Gain};
+    const ALfloat oldGainStep{-oldGain / (ALfloat)BufferSize};
     const ALfloat (&NewCoeffs)[HRIR_LENGTH][2] = *newparams->Coeffs;
-    const ALsizei NewDelay[2] = { newparams->Delay[0], newparams->Delay[1] };
-    const ALfloat newGain = newparams->Gain;
-    const ALfloat newGainStep = newparams->GainStep;
-    ALfloat g, stepcount = 0.0f;
-    ALfloat left, right;
-    ALsizei i;
+    const ALfloat newGainStep{newparams->GainStep};
+    ALfloat stepcount{0.0f};
 
+    ASSUME(OutPos >= 0);
     ASSUME(IrSize >= 4);
     ASSUME(BufferSize > 0);
 
+    ALsizei HistOffset{Offset&HRTF_HISTORY_MASK};
+    ALsizei OldDelay[2]{
+        (HistOffset-oldparams->Delay[0])&HRTF_HISTORY_MASK,
+        (HistOffset-oldparams->Delay[1])&HRTF_HISTORY_MASK };
+    ALsizei NewDelay[2]{
+        (HistOffset-newparams->Delay[0])&HRTF_HISTORY_MASK,
+        (HistOffset-newparams->Delay[1])&HRTF_HISTORY_MASK };
+
+    Offset &= HRIR_MASK;
+    ALsizei HeadOffset{(Offset+IrSize-1)&HRIR_MASK};
+
     LeftOut  += OutPos;
     RightOut += OutPos;
-    for(i = 0;i < BufferSize;i++)
+    for(ALsizei i{0};i < BufferSize;)
     {
-        hrtfstate->Values[(Offset+IrSize-1)&HRIR_MASK][0] = 0.0f;
-        hrtfstate->Values[(Offset+IrSize-1)&HRIR_MASK][1] = 0.0f;
-
-        hrtfstate->History[Offset&HRTF_HISTORY_MASK] = *(data++);
-
-        g = oldGain + oldGainStep*stepcount;
-        left = hrtfstate->History[(Offset-OldDelay[0])&HRTF_HISTORY_MASK]*g;
-        right = hrtfstate->History[(Offset-OldDelay[1])&HRTF_HISTORY_MASK]*g;
-        ApplyCoeffs(Offset, hrtfstate->Values, IrSize, OldCoeffs, left, right);
-
-        g = newGain + newGainStep*stepcount;
-        left = hrtfstate->History[(Offset-NewDelay[0])&HRTF_HISTORY_MASK]*g;
-        right = hrtfstate->History[(Offset-NewDelay[1])&HRTF_HISTORY_MASK]*g;
-        ApplyCoeffs(Offset, hrtfstate->Values, IrSize, NewCoeffs, left, right);
-
-        *(LeftOut++)  += hrtfstate->Values[Offset&HRIR_MASK][0];
-        *(RightOut++) += hrtfstate->Values[Offset&HRIR_MASK][1];
-
-        stepcount += 1.0f;
-        Offset++;
+        const ALsizei todo_hist{HRTF_HISTORY_LENGTH -
+            maxi(maxi(maxi(maxi(HistOffset, OldDelay[0]), OldDelay[1]), NewDelay[0]), NewDelay[1])
+        };
+        const ALsizei todo_hrir{HRIR_LENGTH - maxi(HeadOffset, Offset)};
+        const ALsizei todo{mini(BufferSize-i, mini(todo_hist, todo_hrir)) + i};
+        ASSUME(todo > i);
+
+        for(;i < todo;++i)
+        {
+            hrtfstate->Values[HeadOffset][0] = 0.0f;
+            hrtfstate->Values[HeadOffset][1] = 0.0f;
+            ++HeadOffset;
+
+            hrtfstate->History[HistOffset++] = *(data++);
+
+            ALfloat g{oldGain + oldGainStep*stepcount};
+            ALfloat left{hrtfstate->History[OldDelay[0]++] * g};
+            ALfloat right{hrtfstate->History[OldDelay[1]++] * g};
+            ApplyCoeffs(Offset, hrtfstate->Values, IrSize, OldCoeffs, left, right);
+
+            g = newGainStep*stepcount;
+            left = hrtfstate->History[NewDelay[0]++] * g;
+            right = hrtfstate->History[NewDelay[1]++] * g;
+            ApplyCoeffs(Offset, hrtfstate->Values, IrSize, NewCoeffs, left, right);
+
+            *(LeftOut++)  += hrtfstate->Values[Offset][0];
+            *(RightOut++) += hrtfstate->Values[Offset][1];
+            ++Offset;
+
+            stepcount += 1.0f;
+        }
+
+        HeadOffset &= HRIR_MASK;
+        HistOffset &= HRTF_HISTORY_MASK;
+        OldDelay[0] &= HRTF_HISTORY_MASK;
+        OldDelay[1] &= HRTF_HISTORY_MASK;
+        NewDelay[0] &= HRTF_HISTORY_MASK;
+        NewDelay[1] &= HRTF_HISTORY_MASK;
+        Offset &= HRIR_MASK;
     }
-    newparams->Gain = newGain + newGainStep*stepcount;
+    newparams->Gain = newGainStep*stepcount;
 }
 
 void MixDirectHrtf(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut,
                    const ALfloat *data, DirectHrtfState *State, const ALsizei Chan,
                    const ALsizei BufferSize)
 {
-    const ALsizei IrSize{State->IrSize};
-    ALsizei Offset{State->Offset};
-    ALfloat (&Values)[HRIR_LENGTH][2] = State->Chan[Chan].Values;
+    ASSUME(Chan >= 0);
+    ASSUME(BufferSize > 0);
+
     const ALfloat (&Coeffs)[HRIR_LENGTH][2] = State->Chan[Chan].Coeffs;
+    ALfloat (&Values)[HRIR_LENGTH][2] = State->Chan[Chan].Values;
+    ALsizei Offset{State->Offset&HRIR_MASK};
+    const ALsizei IrSize{State->IrSize};
 
     ASSUME(IrSize >= 4);
-    ASSUME(BufferSize > 0);
 
-    for(ALsizei i{0};i < BufferSize;i++)
+    ALsizei HeadOffset{(Offset+IrSize-1)&HRIR_MASK};
+    for(ALsizei i{0};i < BufferSize;)
     {
-        Values[(Offset+IrSize)&HRIR_MASK][0] = 0.0f;
-        Values[(Offset+IrSize)&HRIR_MASK][1] = 0.0f;
-        Offset++;
-
-        const ALfloat insample{*(data++)};
-        ApplyCoeffs(Offset, Values, IrSize, Coeffs, insample, insample);
-        *(LeftOut++)  += Values[Offset&HRIR_MASK][0];
-        *(RightOut++) += Values[Offset&HRIR_MASK][1];
+        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{*(data++)};
+            ApplyCoeffs(Offset, Values, IrSize, Coeffs, insample, insample);
+
+            *(LeftOut++)  += Values[Offset][0];
+            *(RightOut++) += Values[Offset][1];
+            ++Offset;
+        }
+        HeadOffset &= HRIR_MASK;
+        Offset &= HRIR_MASK;
     }
 }
diff --git a/Alc/mixer/mixer_c.cpp b/Alc/mixer/mixer_c.cpp
index d98b8e2e..22d3642e 100644
--- a/Alc/mixer/mixer_c.cpp
+++ b/Alc/mixer/mixer_c.cpp
@@ -107,24 +107,23 @@ static inline void ApplyCoeffs(ALsizei Offset, ALfloat (&Values)[HRIR_LENGTH][2]
                                const ALsizei IrSize, const ALfloat (&Coeffs)[HRIR_LENGTH][2],
                                const ALfloat left, const ALfloat right)
 {
-    ALsizei off{Offset&HRIR_MASK};
-    ALsizei count{mini(IrSize, HRIR_LENGTH - off)};
-
+    ASSUME(Offset >= 0 && Offset < HRIR_LENGTH);
     ASSUME(IrSize >= 2);
     ASSUME(&Values != &Coeffs);
-    ASSUME(count > 0);
 
+    ALsizei count{mini(IrSize, HRIR_LENGTH - Offset)};
+    ASSUME(count > 0);
     for(ALsizei c{0};;)
     {
         for(;c < count;++c)
         {
-            Values[off][0] += Coeffs[c][0] * left;
-            Values[off][1] += Coeffs[c][1] * right;
-            ++off;
+            Values[Offset][0] += Coeffs[c][0] * left;
+            Values[Offset][1] += Coeffs[c][1] * right;
+            ++Offset;
         }
         if(c >= IrSize)
             break;
-        off = 0;
+        Offset = 0;
         count = IrSize;
     }
 }
diff --git a/Alc/mixer/mixer_sse.cpp b/Alc/mixer/mixer_sse.cpp
index 5d82e5ae..2637883b 100644
--- a/Alc/mixer/mixer_sse.cpp
+++ b/Alc/mixer/mixer_sse.cpp
@@ -90,45 +90,45 @@ static inline void ApplyCoeffs(ALsizei Offset, ALfloat (&Values)[HRIR_LENGTH][2]
     ASSUME(IrSize >= 2);
     ASSUME(&Values != &Coeffs);
 
-    ALsizei off{Offset&HRIR_MASK};
+    ASSUME(Offset >= 0 && Offset < HRIR_LENGTH);
     if((Offset&1))
     {
-        ALsizei count{mini(IrSize-1, HRIR_LENGTH - off)};
+        ALsizei count{mini(IrSize-1, HRIR_LENGTH - Offset)};
         ASSUME(count >= 1);
 
         __m128 imp0, imp1;
         coeffs = _mm_load_ps(&Coeffs[0][0]);
-        vals = _mm_loadl_pi(vals, (__m64*)&Values[off][0]);
+        vals = _mm_loadl_pi(vals, (__m64*)&Values[Offset][0]);
         imp0 = _mm_mul_ps(lrlr, coeffs);
         vals = _mm_add_ps(imp0, vals);
-        _mm_storel_pi((__m64*)&Values[off][0], vals);
-        ++off;
+        _mm_storel_pi((__m64*)&Values[Offset][0], vals);
+        ++Offset;
         for(ALsizei i{1};;)
         {
             for(;i < count;i += 2)
             {
                 coeffs = _mm_load_ps(&Coeffs[i+1][0]);
-                vals = _mm_load_ps(&Values[off][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[off][0], vals);
+                _mm_store_ps(&Values[Offset][0], vals);
                 imp0 = imp1;
-                off += 2;
+                Offset += 2;
             }
-            off &= HRIR_MASK;
+            Offset &= HRIR_MASK;
             if(i >= IrSize-1)
                 break;
             count = IrSize-1;
         }
-        vals = _mm_loadl_pi(vals, (__m64*)&Values[off][0]);
+        vals = _mm_loadl_pi(vals, (__m64*)&Values[Offset][0]);
         imp0 = _mm_movehl_ps(imp0, imp0);
         vals = _mm_add_ps(imp0, vals);
-        _mm_storel_pi((__m64*)&Values[off][0], vals);
+        _mm_storel_pi((__m64*)&Values[Offset][0], vals);
     }
     else
     {
-        ALsizei count{mini(IrSize, HRIR_LENGTH - off)};
+        ALsizei count{mini(IrSize, HRIR_LENGTH - Offset)};
         ASSUME(count >= 2);
 
         for(ALsizei i{0};;)
@@ -136,14 +136,14 @@ static inline void ApplyCoeffs(ALsizei Offset, ALfloat (&Values)[HRIR_LENGTH][2]
             for(;i < count;i += 2)
             {
                 coeffs = _mm_load_ps(&Coeffs[i][0]);
-                vals = _mm_load_ps(&Values[off][0]);
+                vals = _mm_load_ps(&Values[Offset][0]);
                 vals = _mm_add_ps(vals, _mm_mul_ps(lrlr, coeffs));
-                _mm_store_ps(&Values[off][0], vals);
-                off += 2;
+                _mm_store_ps(&Values[Offset][0], vals);
+                Offset += 2;
             }
             if(i >= IrSize)
                 break;
-            off = 0;
+            Offset = 0;
             count = IrSize;
         }
     }
diff --git a/OpenAL32/Include/alu.h b/OpenAL32/Include/alu.h
index 4fcc4c9c..08e6319e 100644
--- a/OpenAL32/Include/alu.h
+++ b/OpenAL32/Include/alu.h
@@ -280,12 +280,12 @@ using RowMixerFunc = void(*)(ALfloat *OutBuffer, const ALfloat *gains,
     const ALfloat (*RESTRICT data)[BUFFERSIZE], ALsizei InChans, ALsizei InPos,
     ALsizei BufferSize);
 using HrtfMixerFunc = void(*)(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut,
-    const ALfloat *data, ALsizei Offset, ALsizei OutPos, const ALsizei IrSize,
-    MixHrtfParams *hrtfparams, HrtfState *hrtfstate, ALsizei BufferSize);
+    const ALfloat *data, ALsizei Offset, const ALsizei OutPos, const ALsizei IrSize,
+    MixHrtfParams *hrtfparams, HrtfState *hrtfstate, const ALsizei BufferSize);
 using HrtfMixerBlendFunc = void(*)(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut,
-    const ALfloat *data, ALsizei Offset, ALsizei OutPos, const ALsizei IrSize,
+    const ALfloat *data, ALsizei Offset, const ALsizei OutPos, const ALsizei IrSize,
     const HrtfParams *oldparams, MixHrtfParams *newparams, HrtfState *hrtfstate,
-    ALsizei BufferSize);
+    const ALsizei BufferSize);
 using HrtfDirectMixerFunc = void(*)(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut,
     const ALfloat *data, DirectHrtfState *State, const ALsizei Chan, const ALsizei BufferSize);