Pass the MixRow buffer size as a span

7 files changed, 57 insertions, 61 deletions

diff --git a/alc/alu.h b/alc/alu.h
index 117c932d..c63e33fd 100644
--- a/alc/alu.h
+++ b/alc/alu.h
@@ -317,8 +317,8 @@ struct ALvoice {
 using MixerFunc = void(*)(const ALfloat *data, const al::span<FloatBufferLine> OutBuffer,
     ALfloat *CurrentGains, const ALfloat *TargetGains, const ALsizei Counter, const ALsizei OutPos,
     const ALsizei BufferSize);
-using RowMixerFunc = void(*)(ALfloat *OutBuffer, const al::span<const ALfloat> Gains,
-    const ALfloat *InSamples, const ALsizei InStride, const ALsizei BufferSize);
+using RowMixerFunc = void(*)(const al::span<float> OutBuffer, const al::span<const float> Gains,
+    const float *InSamples, const size_t InStride);
 using HrtfMixerFunc = void(*)(FloatBufferLine &LeftOut, FloatBufferLine &RightOut,
     const ALfloat *InSamples, float2 *AccumSamples, const ALsizei OutPos, const ALsizei IrSize,
     MixHrtfFilter *hrtfparams, const ALsizei BufferSize);
diff --git a/alc/bformatdec.cpp b/alc/bformatdec.cpp
index ce4c4b3b..7d681bc6 100644
--- a/alc/bformatdec.cpp
+++ b/alc/bformatdec.cpp
@@ -162,10 +162,11 @@ void BFormatDec::process(const al::span<FloatBufferLine> OutBuffer,
         {
             if LIKELY(enabled&1)
             {
-                MixRowSamples(outbuf.data(), {(*mixmtx)[sHFBand], mNumChannels},
-                    mSamplesHF->data(), mSamplesHF->size(), SamplesToDo);
-                MixRowSamples(outbuf.data(), {(*mixmtx)[sLFBand], mNumChannels},
-                    mSamplesLF->data(), mSamplesLF->size(), SamplesToDo);
+                const al::span<float> outspan{outbuf.data(), outbuf.data()+SamplesToDo};
+                MixRowSamples(outspan, {(*mixmtx)[sHFBand], mNumChannels}, mSamplesHF->data(),
+                    mSamplesHF->size());
+                MixRowSamples(outspan, {(*mixmtx)[sLFBand], mNumChannels}, mSamplesLF->data(),
+                    mSamplesLF->size());
             }
             ++mixmtx;
             enabled >>= 1;
@@ -178,8 +179,8 @@ void BFormatDec::process(const al::span<FloatBufferLine> OutBuffer,
         for(FloatBufferLine &outbuf : OutBuffer)
         {
             if LIKELY(enabled&1)
-                MixRowSamples(outbuf.data(), {*mixmtx, mNumChannels}, InSamples->data(),
-                    InSamples->size(), SamplesToDo);
+                MixRowSamples({outbuf.data(), outbuf.data()+SamplesToDo}, {*mixmtx, mNumChannels},
+                    InSamples->data(), InSamples->size());
             ++mixmtx;
             enabled >>= 1;
         }
diff --git a/alc/effects/reverb.cpp b/alc/effects/reverb.cpp
index 8ff4feac..5f8daabe 100644
--- a/alc/effects/reverb.cpp
+++ b/alc/effects/reverb.cpp
@@ -421,19 +421,20 @@ struct ReverbState final : public EffectState {
         ASSUME(todo > 0);
 
         /* Convert back to B-Format, and mix the results to output. */
+        const al::span<float> tmpspan{mTempLine.begin(), mTempLine.begin()+todo};
         for(size_t c{0u};c < NUM_LINES;c++)
         {
-            std::fill_n(mTempLine.begin(), todo, 0.0f);
-            MixRowSamples(mTempLine.data(), {A2B[c], NUM_LINES}, mEarlySamples[0].data(),
-                mEarlySamples[0].size(), todo);
+            std::fill(tmpspan.begin(), tmpspan.end(), 0.0f);
+            MixRowSamples(tmpspan, {A2B[c], NUM_LINES}, mEarlySamples[0].data(),
+                mEarlySamples[0].size());
             MixSamples(mTempLine.data(), samplesOut, mEarly.CurrentGain[c],
                 mEarly.PanGain[c], counter, offset, todo);
         }
         for(size_t c{0u};c < NUM_LINES;c++)
         {
-            std::fill_n(mTempLine.begin(), todo, 0.0f);
-            MixRowSamples(mTempLine.data(), {A2B[c], NUM_LINES}, mLateSamples[0].data(),
-                mLateSamples[0].size(), todo);
+            std::fill(tmpspan.begin(), tmpspan.end(), 0.0f);
+            MixRowSamples(tmpspan, {A2B[c], NUM_LINES}, mLateSamples[0].data(),
+                mLateSamples[0].size());
             MixSamples(mTempLine.data(), samplesOut, mLate.CurrentGain[c], mLate.PanGain[c],
                 counter, offset, todo);
         }
@@ -444,11 +445,12 @@ struct ReverbState final : public EffectState {
     {
         ASSUME(todo > 0);
 
+        const al::span<float> tmpspan{mTempLine.begin(), mTempLine.begin()+todo};
         for(size_t c{0u};c < NUM_LINES;c++)
         {
-            std::fill_n(mTempLine.begin(), todo, 0.0f);
-            MixRowSamples(mTempLine.data(), {A2B[c], NUM_LINES}, mEarlySamples[0].data(),
-                mEarlySamples[0].size(), todo);
+            std::fill(tmpspan.begin(), tmpspan.end(), 0.0f);
+            MixRowSamples(tmpspan, {A2B[c], NUM_LINES}, mEarlySamples[0].data(),
+                mEarlySamples[0].size());
 
             /* Apply scaling to the B-Format's HF response to "upsample" it to
              * higher-order output.
@@ -461,9 +463,9 @@ struct ReverbState final : public EffectState {
         }
         for(size_t c{0u};c < NUM_LINES;c++)
         {
-            std::fill_n(mTempLine.begin(), todo, 0.0f);
-            MixRowSamples(mTempLine.data(), {A2B[c], NUM_LINES}, mLateSamples[0].data(),
-                mLateSamples[0].size(), todo);
+            std::fill(tmpspan.begin(), tmpspan.end(), 0.0f);
+            MixRowSamples(tmpspan, {A2B[c], NUM_LINES}, mLateSamples[0].data(),
+                mLateSamples[0].size());
 
             const ALfloat hfscale{(c==0) ? mOrderScales[0] : mOrderScales[1]};
             mAmbiSplitter[1][c].applyHfScale(mTempLine.data(), hfscale, todo);
@@ -1457,11 +1459,11 @@ void ReverbState::process(const ALsizei samplesToDo, const FloatBufferLine *REST
     ASSUME(offset >= 0);
 
     /* Convert B-Format to A-Format for processing. */
+    const al::span<float> tmpspan{mTempLine.begin(), mTempLine.begin()+samplesToDo};
     for(size_t c{0u};c < NUM_LINES;c++)
     {
-        std::fill_n(mTempLine.begin(), samplesToDo, 0.0f);
-        MixRowSamples(mTempLine.data(), {B2A[c], B2A[c]+numInput}, samplesIn->data(),
-            samplesIn->size(), samplesToDo);
+        std::fill(tmpspan.begin(), tmpspan.end(), 0.0f);
+        MixRowSamples(tmpspan, {B2A[c], B2A[c]+numInput}, samplesIn->data(), samplesIn->size());
 
         /* Band-pass the incoming samples and feed the initial delay line. */
         mFilter[c].Lp.process(mTempLine.data(), mTempLine.data(), samplesToDo);
diff --git a/alc/mixer/defs.h b/alc/mixer/defs.h
index 19fb42d0..adaed8f5 100644
--- a/alc/mixer/defs.h
+++ b/alc/mixer/defs.h
@@ -32,7 +32,8 @@ const ALfloat *Resample_(const InterpState *state, const ALfloat *RESTRICT src,
 template<InstSetType InstTag>
 void Mix_(const ALfloat *data, const al::span<FloatBufferLine> OutBuffer, ALfloat *CurrentGains, const ALfloat *TargetGains, const ALsizei Counter, const ALsizei OutPos, const ALsizei BufferSize);
 template<InstSetType InstTag>
-void MixRow_(ALfloat *OutBuffer, const al::span<const ALfloat> Gains, const ALfloat *InSamples, const ALsizei InStride, const ALsizei BufferSize);
+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 ALfloat *InSamples, float2 *AccumSamples, const ALsizei OutPos, const ALsizei IrSize, MixHrtfFilter *hrtfparams, const ALsizei BufferSize);
diff --git a/alc/mixer/mixer_c.cpp b/alc/mixer/mixer_c.cpp
index 4985933c..dd094040 100644
--- a/alc/mixer/mixer_c.cpp
+++ b/alc/mixer/mixer_c.cpp
@@ -189,20 +189,18 @@ void Mix_<CTag>(const ALfloat *data, const al::span<FloatBufferLine> OutBuffer,
  * stepping is necessary.
  */
 template<>
-void MixRow_<CTag>(ALfloat *OutBuffer, const al::span<const ALfloat> Gains,
-    const ALfloat *InSamples, const ALsizei InStride, const ALsizei BufferSize)
+void MixRow_<CTag>(const al::span<float> OutBuffer, const al::span<const float> Gains,
+    const float *InSamples, const size_t InStride)
 {
-    ASSUME(BufferSize > 0);
-
-    for(const ALfloat gain : Gains)
+    for(const float gain : Gains)
     {
-        const ALfloat *RESTRICT src{InSamples};
+        const float *RESTRICT src{InSamples};
         InSamples += InStride;
 
         if(!(std::fabs(gain) > GAIN_SILENCE_THRESHOLD))
             continue;
 
-        for(ALsizei i{0};i < BufferSize;i++)
-            OutBuffer[i] += src[i] * gain;
+        std::transform(OutBuffer.begin(), OutBuffer.end(), src, OutBuffer.begin(),
+            [gain](const ALfloat cur, const ALfloat src) -> ALfloat { return cur + src*gain; });
     }
 }
diff --git a/alc/mixer/mixer_neon.cpp b/alc/mixer/mixer_neon.cpp
index c6ea2c60..829d4ea1 100644
--- a/alc/mixer/mixer_neon.cpp
+++ b/alc/mixer/mixer_neon.cpp
@@ -276,11 +276,9 @@ void Mix_<NEONTag>(const ALfloat *data, const al::span<FloatBufferLine> OutBuffe
 }
 
 template<>
-void MixRow_<NEONTag>(ALfloat *OutBuffer, const al::span<const ALfloat> Gains,
-    const ALfloat *InSamples, const ALsizei InStride, const ALsizei BufferSize)
+void MixRow_<NEONTag>(const al::span<ALfloat> OutBuffer, const al::span<const ALfloat> Gains,
+    const ALfloat *InSamples, const ALsizei InStride)
 {
-    ASSUME(BufferSize > 0);
-
     for(const ALfloat gain : Gains)
     {
         const ALfloat *RESTRICT src{InSamples};
@@ -289,20 +287,19 @@ void MixRow_<NEONTag>(ALfloat *OutBuffer, const al::span<const ALfloat> Gains,
         if(!(std::fabs(gain) > GAIN_SILENCE_THRESHOLD))
             continue;
 
-        ALsizei pos{0};
-        if LIKELY(BufferSize > 3)
+        auto out_iter = OutBuffer.begin();
+        if(size_t todo{OutBuffer.size() >> 2})
         {
-            ALsizei todo{BufferSize >> 2};
-            float32x4_t gain4{vdupq_n_f32(gain)};
+            const float32x4_t gain4{vdupq_n_f32(gain)};
             do {
-                const float32x4_t val4 = vld1q_f32(&src[pos]);
-                float32x4_t dry4 = vld1q_f32(&OutBuffer[pos]);
+                const float32x4_t val4 = vld1q_f32(src);
+                float32x4_t dry4 = vld1q_f32(out_iter);
                 dry4 = vmlaq_f32(dry4, val4, gain4);
-                vst1q_f32(&OutBuffer[pos], dry4);
-                pos += 4;
+                vst1q_f32(out_iter, dry4);
+                out_iter += 4; src += 4;
             } while(--todo);
         }
-        for(;pos < BufferSize;pos++)
-            OutBuffer[pos] += src[pos]*gain;
+        std::transform(out_iter, OutBuffer.end(), src, out_iter,
+            [gain](const ALfloat cur, const ALfloat src) -> ALfloat { return cur + src*gain; });
     }
 }
diff --git a/alc/mixer/mixer_sse.cpp b/alc/mixer/mixer_sse.cpp
index b5958c8e..b3454898 100644
--- a/alc/mixer/mixer_sse.cpp
+++ b/alc/mixer/mixer_sse.cpp
@@ -229,33 +229,30 @@ void Mix_<SSETag>(const ALfloat *data, const al::span<FloatBufferLine> OutBuffer
 }
 
 template<>
-void MixRow_<SSETag>(ALfloat *OutBuffer, const al::span<const ALfloat> Gains,
-    const ALfloat *InSamples, const ALsizei InStride, const ALsizei BufferSize)
+void MixRow_<SSETag>(const al::span<float> OutBuffer, const al::span<const float> Gains,
+    const float *InSamples, const size_t InStride)
 {
-    ASSUME(BufferSize > 0);
-
-    for(const ALfloat gain : Gains)
+    for(const float gain : Gains)
     {
-        const ALfloat *RESTRICT src{InSamples};
+        const float *RESTRICT src{InSamples};
         InSamples += InStride;
 
         if(!(std::fabs(gain) > GAIN_SILENCE_THRESHOLD))
             continue;
 
-        ALsizei pos{0};
-        if LIKELY(BufferSize > 3)
+        auto out_iter = OutBuffer.begin();
+        if(size_t todo{OutBuffer.size() >> 2})
         {
-            ALsizei todo{BufferSize >> 2};
             const __m128 gain4 = _mm_set1_ps(gain);
             do {
-                const __m128 val4{_mm_load_ps(&src[pos])};
-                __m128 dry4{_mm_load_ps(&OutBuffer[pos])};
+                const __m128 val4{_mm_load_ps(src)};
+                __m128 dry4{_mm_load_ps(out_iter)};
                 dry4 = _mm_add_ps(dry4, _mm_mul_ps(val4, gain4));
-                _mm_store_ps(&OutBuffer[pos], dry4);
-                pos += 4;
+                _mm_store_ps(out_iter, dry4);
+                out_iter += 4; src += 4;
             } while(--todo);
         }
-        for(;pos < BufferSize;pos++)
-            OutBuffer[pos] += src[pos]*gain;
+        std::transform(out_iter, OutBuffer.end(), src, out_iter,
+            [gain](const ALfloat cur, const ALfloat src) -> ALfloat { return cur + src*gain; });
     }
 }