From a4bc0a46e9e0fe001d4c14909b47ed4ac8d81960 Mon Sep 17 00:00:00 2001
From: Chris Robinson <chris.kcat@gmail.com>
Date: Sun, 26 Jan 2014 01:34:39 -0800
Subject: [PATCH] Implement dry and wet mixers for Neon

Code provided by Philippe Simons <simons.philippe@gmail.com>.
---
 Alc/ALc.c        |  6 +++--
 Alc/ALu.c        |  8 ++++++
 Alc/mixer_defs.h |  2 ++
 Alc/mixer_neon.c | 80 +++++++++++++++++++++++++++++++++++++++++++++++++++++---
 4 files changed, 91 insertions(+), 5 deletions(-)

diff --git a/Alc/ALc.c b/Alc/ALc.c
index fea0320d..7d8f14cd 100644
--- a/Alc/ALc.c
+++ b/Alc/ALc.c
@@ -1745,8 +1745,8 @@ static ALCenum UpdateDeviceParams(ALCdevice *device, const ALCint *attrList)
 
         device->UpdateSize = (ALuint64)device->UpdateSize * freq /
                              device->Frequency;
-        /* SSE does best with the update size being a multiple of 4 */
-        if((CPUCapFlags&CPU_CAP_SSE))
+        /* SSE and Neon do best with the update size being a multiple of 4 */
+        if((CPUCapFlags&(CPU_CAP_SSE|CPU_CAP_NEON)) != 0)
             device->UpdateSize = (device->UpdateSize+3)&~3;
 
         device->Frequency = freq;
@@ -1861,6 +1861,8 @@ static ALCenum UpdateDeviceParams(ALCdevice *device, const ALCint *attrList)
     {
         if((CPUCapFlags&CPU_CAP_SSE))
             WARN("SSE performs best with multiple of 4 update sizes (%u)\n", device->UpdateSize);
+        if((CPUCapFlags&CPU_CAP_NEON))
+            WARN("NEON performs best with multiple of 4 update sizes (%u)\n", device->UpdateSize);
     }
 
     SetMixerFPUMode(&oldMode);
diff --git a/Alc/ALu.c b/Alc/ALu.c
index 34ac6687..bf09450f 100644
--- a/Alc/ALu.c
+++ b/Alc/ALu.c
@@ -118,6 +118,10 @@ static DryMixerFunc SelectDirectMixer(void)
     if((CPUCapFlags&CPU_CAP_SSE))
         return MixDirect_SSE;
 #endif
+#ifdef HAVE_NEON
+    if((CPUCapFlags&CPU_CAP_NEON))
+        return MixDirect_Neon;
+#endif
 
     return MixDirect_C;
 }
@@ -128,6 +132,10 @@ static WetMixerFunc SelectSendMixer(void)
     if((CPUCapFlags&CPU_CAP_SSE))
         return MixSend_SSE;
 #endif
+#ifdef HAVE_NEON
+    if((CPUCapFlags&CPU_CAP_NEON))
+        return MixSend_Neon;
+#endif
 
     return MixSend_C;
 }
diff --git a/Alc/mixer_defs.h b/Alc/mixer_defs.h
index 5e43af15..f8968a0a 100644
--- a/Alc/mixer_defs.h
+++ b/Alc/mixer_defs.h
@@ -27,5 +27,7 @@ void MixSend_SSE(const struct SendParams*,const ALfloat*restrict,ALuint,ALuint,A
 
 /* Neon mixers */
 void MixDirect_Hrtf_Neon(const struct DirectParams*,const ALfloat*restrict,ALuint,ALuint,ALuint,ALuint);
+void MixDirect_Neon(const struct DirectParams*,const ALfloat*restrict,ALuint,ALuint,ALuint,ALuint);
+void MixSend_Neon(const struct SendParams*,const ALfloat*restrict,ALuint,ALuint,ALuint);
 
 #endif /* MIXER_DEFS_H */
diff --git a/Alc/mixer_neon.c b/Alc/mixer_neon.c
index 571221be..0aa450ad 100644
--- a/Alc/mixer_neon.c
+++ b/Alc/mixer_neon.c
@@ -14,11 +14,15 @@ static inline void ApplyCoeffsStep(const ALuint IrSize,
                                    ALfloat (*restrict Coeffs)[2],
                                    const ALfloat (*restrict CoeffStep)[2])
 {
+    float32x4_t coeffs, deltas;
     ALuint c;
-    for(c = 0;c < IrSize;c++)
+
+    for(c = 0;c < IrSize;c += 2)
     {
-        Coeffs[c][0] += CoeffStep[c][0];
-        Coeffs[c][1] += CoeffStep[c][1];
+        coeffs = vld1q_f32(&Coeffs[c][0]);
+        deltas = vld1q_f32(&CoeffStep[c][0]);
+        coeffs = vaddq_f32(coeffs, deltas);
+        vst1q_f32(&Coeffs[c][0], coeffs);
     }
 }
 
@@ -54,3 +58,73 @@ static inline void ApplyCoeffs(ALuint Offset, ALfloat (*restrict Values)[2],
 #define SUFFIX Neon
 #include "mixer_inc.c"
 #undef SUFFIX
+
+
+void MixDirect_Neon(const DirectParams *params, const ALfloat *restrict data, ALuint srcchan,
+  ALuint OutPos, ALuint SamplesToDo, ALuint BufferSize)
+{
+    ALfloat (*restrict OutBuffer)[BUFFERSIZE] = params->OutBuffer;
+    ALfloat *restrict ClickRemoval = params->ClickRemoval;
+    ALfloat *restrict PendingClicks = params->PendingClicks;
+    ALfloat DrySend;
+    float32x4_t gain;
+    ALuint pos;
+    ALuint c;
+
+    for(c = 0;c < MaxChannels;c++)
+    {
+        DrySend = params->Gains[srcchan][c];
+        if(!(DrySend > GAIN_SILENCE_THRESHOLD))
+            continue;
+
+        if(OutPos == 0)
+            ClickRemoval[c] -= data[0]*DrySend;
+
+        gain = vdupq_n_f32(DrySend);
+        for(pos = 0;BufferSize-pos > 3;pos += 4)
+        {
+            const float32x4_t val4 = vld1q_f32(&data[pos]);
+            float32x4_t dry4 = vld1q_f32(&OutBuffer[c][OutPos+pos]);
+            dry4 = vaddq_f32(dry4, vmulq_f32(val4, gain));
+            vst1q_f32(&OutBuffer[c][OutPos+pos], dry4);
+        }
+        for(;pos < BufferSize;pos++)
+            OutBuffer[c][OutPos+pos] += data[pos]*DrySend;
+
+        if(OutPos+pos == SamplesToDo)
+            PendingClicks[c] += data[pos]*DrySend;
+    }
+}
+
+
+void MixSend_Neon(const SendParams *params, const ALfloat *restrict data,
+  ALuint OutPos, ALuint SamplesToDo, ALuint BufferSize)
+{
+    ALfloat (*restrict OutBuffer)[BUFFERSIZE] = params->OutBuffer;
+    ALfloat *restrict ClickRemoval = params->ClickRemoval;
+    ALfloat *restrict PendingClicks = params->PendingClicks;
+    ALfloat WetGain;
+    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;
+
+    if(OutPos+pos == SamplesToDo)
+        PendingClicks[0] += data[pos] * WetGain;
+}
-- 
2.11.4.GIT