Add gain stepping to the send mixers

5 files changed, 168 insertions, 64 deletions

diff --git a/Alc/ALc.c b/Alc/ALc.c
index bf296974..7399a0c7 100644
--- a/Alc/ALc.c
+++ b/Alc/ALc.c
@@ -1931,6 +1931,13 @@ static ALCenum UpdateDeviceParams(ALCdevice *device, const ALCint *attrList)
         {
             ALactivesource *src = context->ActiveSources[pos];
             ALsource *source = src->Source;
+            ALuint s = device->NumAuxSends;
+            while(s < MAX_SENDS)
+            {
+                src->Send[s].Moving = AL_FALSE;
+                src->Send[s].Counter = 0;
+                s++;
+            }
 
             src->Update(src, context);
             source->NeedsUpdate = AL_FALSE;
diff --git a/Alc/ALu.c b/Alc/ALu.c
index 10e2a45d..d00620eb 100644
--- a/Alc/ALu.c
+++ b/Alc/ALu.c
@@ -565,7 +565,27 @@ ALvoid CalcNonAttnSourceParams(ALactivesource *src, const ALCcontext *ALContext)
         src->DryMix = SelectDirectMixer();
     }
     for(i = 0;i < NumSends;i++)
-        src->Send[i].Gain = WetGain[i];
+    {
+        if(src->Send[i].Moving)
+        {
+            ALfloat cur = maxf(src->Send[i].Gain.Current, GAIN_SILENCE_THRESHOLD);
+            ALfloat trg = maxf(src->Send[i].Gain.Target, GAIN_SILENCE_THRESHOLD);
+            if(fabs(trg - cur) >= GAIN_SILENCE_THRESHOLD)
+                src->Send[i].Gain.Step = powf(trg/cur, 1.0f/64.0f);
+            else
+                src->Send[i].Gain.Step = 1.0f;
+            src->Send[i].Gain.Current = cur;
+            src->Send[i].Counter = 64;
+        }
+        else
+        {
+            src->Send[i].Gain.Current = WetGain[i];
+            src->Send[i].Gain.Target = WetGain[i];
+            src->Send[i].Gain.Step = 1.0f;
+            src->Send[i].Counter = 0;
+            src->Send[i].Moving  = AL_TRUE;
+        }
+    }
     src->WetMix = SelectSendMixer();
 
     {
@@ -1050,7 +1070,27 @@ ALvoid CalcSourceParams(ALactivesource *src, const ALCcontext *ALContext)
         src->DryMix = SelectDirectMixer();
     }
     for(i = 0;i < NumSends;i++)
-        src->Send[i].Gain = WetGain[i];
+    {
+        if(src->Send[i].Moving)
+        {
+            ALfloat cur = maxf(src->Send[i].Gain.Current, GAIN_SILENCE_THRESHOLD);
+            ALfloat trg = maxf(src->Send[i].Gain.Target, GAIN_SILENCE_THRESHOLD);
+            if(fabs(trg - cur) >= GAIN_SILENCE_THRESHOLD)
+                src->Send[i].Gain.Step = powf(trg/cur, 1.0f/64.0f);
+            else
+                src->Send[i].Gain.Step = 1.0f;
+            src->Send[i].Gain.Current = cur;
+            src->Send[i].Counter = 64;
+        }
+        else
+        {
+            src->Send[i].Gain.Current = WetGain[i];
+            src->Send[i].Gain.Target = WetGain[i];
+            src->Send[i].Gain.Step = 1.0f;
+            src->Send[i].Counter = 0;
+            src->Send[i].Moving  = AL_TRUE;
+        }
+    }
     src->WetMix = SelectSendMixer();
 
     {
diff --git a/Alc/mixer_c.c b/Alc/mixer_c.c
index a4028beb..c5091322 100644
--- a/Alc/mixer_c.c
+++ b/Alc/mixer_c.c
@@ -115,22 +115,30 @@ void MixDirect_C(DirectParams *params, const ALfloat *restrict data, ALuint srcc
 
 
 void MixSend_C(SendParams *params, const ALfloat *restrict data,
-  ALuint OutPos, ALuint SamplesToDo, ALuint BufferSize)
+  ALuint OutPos, ALuint UNUSED(SamplesToDo), ALuint BufferSize)
 {
     ALfloat (*restrict OutBuffer)[BUFFERSIZE] = params->OutBuffer;
-    ALfloat *restrict ClickRemoval = params->ClickRemoval;
-    ALfloat *restrict PendingClicks = params->PendingClicks;
-    ALfloat WetSend;
-    ALuint pos;
-
-    WetSend = params->Gain;
-    if(!(WetSend > GAIN_SILENCE_THRESHOLD))
-        return;
-
-    if(OutPos == 0)
-        ClickRemoval[0] -= data[0] * WetSend;
-    for(pos = 0;pos < BufferSize;pos++)
-        OutBuffer[0][OutPos+pos] += data[pos] * WetSend;
-    if(OutPos+pos == SamplesToDo)
-        PendingClicks[0] += data[pos] * WetSend;
+    ALuint Counter = maxu(params->Counter, OutPos) - OutPos;
+    ALfloat WetSend, Step;
+
+    {
+        ALuint pos = 0;
+        Step = params->Gain.Step;
+        if(Step != 1.0f && Counter > 0)
+        {
+            WetSend = params->Gain.Current;
+            for(;pos < BufferSize && pos < Counter;pos++)
+            {
+                OutBuffer[0][OutPos+pos] += data[pos]*WetSend;
+                WetSend *= Step;
+            }
+            params->Gain.Current = WetSend;
+        }
+
+        WetSend = params->Gain.Target;
+        if(!(WetSend > GAIN_SILENCE_THRESHOLD))
+            return;
+        for(;pos < BufferSize;pos++)
+            OutBuffer[0][OutPos+pos] += data[pos] * WetSend;
+    }
 }
diff --git a/Alc/mixer_neon.c b/Alc/mixer_neon.c
index 35dfd1f1..cdc889c5 100644
--- a/Alc/mixer_neon.c
+++ b/Alc/mixer_neon.c
@@ -133,33 +133,53 @@ void MixDirect_Neon(DirectParams *params, const ALfloat *restrict data, ALuint s
 
 
 void MixSend_Neon(SendParams *params, const ALfloat *restrict data,
-  ALuint OutPos, ALuint SamplesToDo, ALuint BufferSize)
+  ALuint OutPos, ALuint UNUSED(SamplesToDo), ALuint BufferSize)
 {
     ALfloat (*restrict OutBuffer)[BUFFERSIZE] = params->OutBuffer;
-    ALfloat *restrict ClickRemoval = params->ClickRemoval;
-    ALfloat *restrict PendingClicks = params->PendingClicks;
-    ALfloat WetGain;
+    ALuint Counter = maxu(params->Counter, OutPos) - OutPos;
+    ALfloat WetGain, Step;
     float32x4_t gain;
-    ALuint pos;
-
-    WetGain = params->Gain;
-    if(!(WetGain > GAIN_SILENCE_THRESHOLD))
-        return;
-
-    if(OutPos == 0)
-        ClickRemoval[0] -= data[0] * WetGain;
 
-    gain = vdupq_n_f32(WetGain);
-    for(pos = 0;BufferSize-pos > 3;pos += 4)
     {
-        const float32x4_t val4 = vld1q_f32(&data[pos]);
-        float32x4_t wet4 = vld1q_f32(&OutBuffer[0][OutPos+pos]);
-        wet4 = vaddq_f32(wet4, vmulq_f32(val4, gain));
-        vst1q_f32(&OutBuffer[0][OutPos+pos], wet4);
-    }
-    for(;pos < BufferSize;pos++)
-        OutBuffer[0][OutPos+pos] += data[pos] * WetGain;
+        ALuint pos = 0;
+        Step = params->Gain.Step;
+        if(Step != 1.0f && Counter > 0)
+        {
+            WetGain = params->Gain.Current;
+            if(BufferSize-pos > 3 && Counter-pos > 3)
+            {
+                OutBuffer[0][OutPos+pos  ] += data[pos  ]*WetGain;
+                WetGain *= Step;
+                OutBuffer[0][OutPos+pos+1] += data[pos+1]*WetGain;
+                WetGain *= Step;
+                OutBuffer[0][OutPos+pos+2] += data[pos+2]*WetGain;
+                WetGain *= Step;
+                OutBuffer[0][OutPos+pos+4] += data[pos+3]*WetGain;
+                WetGain *= Step;
+            }
+            if(!(BufferSize-pos > 3))
+            {
+                for(;pos < BufferSize && pos < Counter;pos++)
+                {
+                    OutBuffer[0][OutPos+pos] += data[pos]*WetGain;
+                    WetGain *= Step;
+                }
+            }
+            params->Gain.Current = WetGain;
+        }
 
-    if(OutPos+pos == SamplesToDo)
-        PendingClicks[0] += data[pos] * WetGain;
+        WetGain = params->Gain.Target;
+        if(!(WetGain > GAIN_SILENCE_THRESHOLD))
+            return;
+        gain = vdupq_n_f32(WetGain);
+        for(;BufferSize-pos > 3;pos += 4)
+        {
+            const float32x4_t val4 = vld1q_f32(&data[pos]);
+            float32x4_t wet4 = vld1q_f32(&OutBuffer[0][OutPos+pos]);
+            wet4 = vaddq_f32(wet4, vmulq_f32(val4, gain));
+            vst1q_f32(&OutBuffer[0][OutPos+pos], wet4);
+        }
+        for(;pos < BufferSize;pos++)
+            OutBuffer[0][OutPos+pos] += data[pos] * WetGain;
+    }
 }
diff --git a/Alc/mixer_sse.c b/Alc/mixer_sse.c
index 38435dac..4ae15201 100644
--- a/Alc/mixer_sse.c
+++ b/Alc/mixer_sse.c
@@ -204,33 +204,62 @@ void MixDirect_SSE(DirectParams *params, const ALfloat *restrict data, ALuint sr
 
 
 void MixSend_SSE(SendParams *params, const ALfloat *restrict data,
-  ALuint OutPos, ALuint SamplesToDo, ALuint BufferSize)
+  ALuint OutPos, ALuint UNUSED(SamplesToDo), ALuint BufferSize)
 {
     ALfloat (*restrict OutBuffer)[BUFFERSIZE] = params->OutBuffer;
-    ALfloat *restrict ClickRemoval = params->ClickRemoval;
-    ALfloat *restrict PendingClicks = params->PendingClicks;
-    ALfloat WetGain;
-    __m128 gain;
-    ALuint pos;
+    ALuint Counter = maxu(params->Counter, OutPos) - OutPos;
+    ALfloat WetGain, Step;
+    __m128 gain, step;
 
-    WetGain = params->Gain;
-    if(!(WetGain > GAIN_SILENCE_THRESHOLD))
-        return;
+    {
+        ALuint pos = 0;
 
-    if(OutPos == 0)
-        ClickRemoval[0] -= data[0] * WetGain;
+        Step = params->Gain.Step;
+        if(Step != 1.0f && Counter > 0)
+        {
+            WetGain = params->Gain.Current;
+            if(BufferSize-pos > 3 && Counter-pos > 3)
+            {
+                gain = _mm_set_ps(
+                    WetGain,
+                    WetGain * Step,
+                    WetGain * Step * Step,
+                    WetGain * Step * Step * Step
+                );
+                step = _mm_set1_ps(Step * Step * Step * Step);
+                do {
+                    const __m128 val4 = _mm_load_ps(&data[pos]);
+                    __m128 dry4 = _mm_load_ps(&OutBuffer[0][OutPos+pos]);
+                    dry4 = _mm_add_ps(dry4, _mm_mul_ps(val4, gain));
+                    gain = _mm_mul_ps(gain, step);
+                    _mm_store_ps(&OutBuffer[0][OutPos+pos], dry4);
+                    pos += 4;
+                } while(BufferSize-pos > 3 && Counter-pos > 3);
+                WetGain = _mm_cvtss_f32(_mm_shuffle_ps(gain, gain, _MM_SHUFFLE(3, 3, 3, 3)));
+            }
+            if(!(BufferSize-pos > 3))
+            {
+                for(;pos < BufferSize && pos < Counter;pos++)
+                {
+                    OutBuffer[0][OutPos+pos] += data[pos]*WetGain;
+                    WetGain *= Step;
+                }
+            }
+            params->Gain.Current = WetGain;
+        }
 
-    gain = _mm_set1_ps(WetGain);
-    for(pos = 0;BufferSize-pos > 3;pos += 4)
-    {
-        const __m128 val4 = _mm_load_ps(&data[pos]);
-        __m128 wet4 = _mm_load_ps(&OutBuffer[0][OutPos+pos]);
-        wet4 = _mm_add_ps(wet4, _mm_mul_ps(val4, gain));
-        _mm_store_ps(&OutBuffer[0][OutPos+pos], wet4);
+        WetGain = params->Gain.Target;
+        if(!(WetGain > GAIN_SILENCE_THRESHOLD))
+            return;
+        gain = _mm_set1_ps(WetGain);
+        for(;BufferSize-pos > 3;pos += 4)
+        {
+            const __m128 val4 = _mm_load_ps(&data[pos]);
+            __m128 wet4 = _mm_load_ps(&OutBuffer[0][OutPos+pos]);
+            wet4 = _mm_add_ps(wet4, _mm_mul_ps(val4, gain));
+            _mm_store_ps(&OutBuffer[0][OutPos+pos], wet4);
+        }
+        for(;pos < BufferSize;pos++)
+            OutBuffer[0][OutPos+pos] += data[pos] * WetGain;
     }
-    for(;pos < BufferSize;pos++)
-        OutBuffer[0][OutPos+pos] += data[pos] * WetGain;
-
-    if(OutPos+pos == SamplesToDo)
-        PendingClicks[0] += data[pos] * WetGain;
 }