Merge pull request #952 from soyersoyer/reverbspeed

Mixing/reverb speed improvements

src/Makefile

@@ -10,7 +10,7 @@ OBJS = main.o kernel.o minidexed.o config.o userinterface.o uimenu.o \
        mididevice.o midikeyboard.o serialmididevice.o pckeyboard.o \
        sysexfileloader.o performanceconfig.o perftimer.o \
        effect_platervbstereo.o uibuttons.o midipin.o \
-       arm_float_to_q23.o \
+       arm_float_to_q23.o arm_scale_zip_f32.o \
        net/ftpdaemon.o net/ftpworker.o net/applemidi.o net/udpmidi.o net/mdnspublisher.o udpmididevice.o
 
 OPTIMIZE = -O3

src/arm_scale_zip_f32.c

@@ -0,0 +1,85 @@
+#include "arm_scale_zip_f32.h"
+
+/**
+  Scale two vectors and zip after.  For floating-point data, the algorithm used is:
+
+  <pre>
+      pDst[n] = pSrc1[n] * scale, pDst[n+1] = pSrc2[n] * scale   0 <= n < blockSize.
+  </pre>
+
+ */
+
+/**
+* @brief Scale two floating-point vector with a scalar and zip after.
+* @param[in]  pSrc1      points to the input vector 1
+* @param[in]  pSrc2      points to the input vector 2
+* @param[in]  scale      scale scalar
+* @param[out] pDst       points to the output vector
+* @param[in]  blockSize  number of samples in the vector
+*/
+
+#if defined(ARM_MATH_NEON_EXPERIMENTAL)
+void arm_scale_zip_f32(
+  const float32_t * pSrc1,
+  const float32_t * pSrc2,
+        float32_t scale,
+        float32_t * pDst,
+        uint32_t blockSize)
+{
+    uint32_t blkCnt;                               /* Loop counter */
+
+    f32x2x2_t res;
+
+    /* Compute 2 outputs at a time */
+    blkCnt = blockSize >> 1U;
+
+    while (blkCnt > 0U)
+    {
+        res.val[0] = vmul_n_f32(vld1_f32(pSrc1), scale);
+        res.val[1] = vmul_n_f32(vld1_f32(pSrc2), scale);
+        vst2_f32(pDst, res);
+
+        /* Increment pointers */
+        pSrc1 += 2;
+        pSrc2 += 2;
+        pDst += 4;
+        
+        /* Decrement the loop counter */
+        blkCnt--;
+    }
+
+    /* If the blockSize is not a multiple of 4, compute any remaining output samples here.
+    ** No loop unrolling is used. */
+    blkCnt = blockSize & 1;
+
+    while (blkCnt > 0U)
+    {
+        *pDst++ = *pSrc1++ * scale;
+        *pDst++ = *pSrc2++ * scale;
+
+        /* Decrement the loop counter */
+        blkCnt--;
+    }
+}
+#else
+void arm_scale_zip_f32(
+  const float32_t * pSrc1,
+  const float32_t * pSrc2,
+        float32_t scale,
+        float32_t * pDst,
+        uint32_t blockSize)
+{
+  uint32_t blkCnt;                               /* Loop counter */
+
+  blkCnt = blockSize;
+
+  while (blkCnt > 0U)
+  {
+      *pDst++ = *pSrc1++ * scale;
+      *pDst++ = *pSrc2++ * scale;
+      
+      /* Decrement the loop counter */
+      blkCnt--;
+  }
+}
+#endif

src/arm_scale_zip_f32.h

@@ -0,0 +1,22 @@
+#pragma once
+
+#include "arm_math_types.h"
+
+#ifdef __cplusplus
+extern "C"
+{
+#endif
+
+/**
+* @brief Scale two floating-point vector with a scalar and zip after.
+* @param[in]  pSrc1      points to the input vector 1
+* @param[in]  pSrc2      points to the input vector 2
+* @param[in]  scale      scale scalar
+* @param[out] pDst       points to the output vector
+* @param[in]  blockSize  number of samples in the vector
+*/
+void arm_scale_zip_f32(const float32_t * pSrc1, const float32_t * pSrc2, float32_t scale, float32_t * pDst, uint32_t blockSize);
+
+#ifdef __cplusplus
+}
+#endif

src/effect_mixer.hpp

@@ -124,31 +124,10 @@ public:
 		assert(in);
 
 		// left
-		arm_scale_f32(in, panorama[channel][0], tmp, buffer_length);
-		if(multiplier[channel]!=UNITY_GAIN)
-			arm_scale_f32(tmp,multiplier[channel],tmp,buffer_length);
-		arm_add_f32(sumbufL, tmp, sumbufL, buffer_length);
-		// right
-		arm_scale_f32(in, panorama[channel][1], tmp, buffer_length);
-		if(multiplier[channel]!=UNITY_GAIN)
-			arm_scale_f32(tmp,multiplier[channel],tmp,buffer_length);
-		arm_add_f32(sumbufR, tmp, sumbufR, buffer_length);
-	}
-
-	void doAddMix(uint8_t channel, float32_t* inL, float32_t* inR)
-	{
-		float32_t tmp[buffer_length];
-
-		assert(inL);
-		assert(inR);
-
-		// left
-		if(multiplier[channel]!=UNITY_GAIN)
-			arm_scale_f32(inL,multiplier[channel],tmp,buffer_length);
+		arm_scale_f32(in, panorama[channel][0] * multiplier[channel], tmp, buffer_length);
 		arm_add_f32(sumbufL, tmp, sumbufL, buffer_length);
 		// right
-		if(multiplier[channel]!=UNITY_GAIN)
-			arm_scale_f32(inR,multiplier[channel],tmp,buffer_length);
+		arm_scale_f32(in, panorama[channel][1] * multiplier[channel], tmp, buffer_length);
 		arm_add_f32(sumbufR, tmp, sumbufR, buffer_length);
 	}
 
@@ -168,6 +147,20 @@ public:
 			arm_fill_f32(0.0f, sumbufR, buffer_length);
 	}
 
+	void getBuffers(float32_t (*buffers[2]))
+	{
+		buffers[0] = sumbufL;
+		buffers[1] = sumbufR;
+	}
+
+	void zeroFill()
+	{
+		if(sumbufL)
+			arm_fill_f32(0.0f, sumbufL, buffer_length);
+		if(sumbufR)
+			arm_fill_f32(0.0f, sumbufR, buffer_length);
+	}
+
 protected:
 	using AudioMixer<NN>::sumbufL;
 	using AudioMixer<NN>::multiplier;

src/minidexed.cpp

@@ -30,6 +30,7 @@
 #include <stdio.h>
 #include <assert.h>
 #include "arm_float_to_q23.h"
+#include "arm_scale_zip_f32.h"
 
 const char WLANFirmwarePath[] = "SD:firmware/";
 const char WLANConfigFile[]   = "SD:wpa_supplicant.conf";
@@ -1354,8 +1355,6 @@ void CMiniDexed::ProcessSound (void)
 			float32_t tmp_float[nFrames*Channels];
 			int32_t tmp_int[nFrames*Channels];
 
-			if(nMasterVolume > 0.0)
-			{
 			// Convert dual float array (8 chan) to single int16 array (8 chan)
 			for(uint16_t i=0; i<nFrames;i++)
 			{
@@ -1363,23 +1362,12 @@ void CMiniDexed::ProcessSound (void)
 				// reading directly from the TG OutputLevel buffer with
 				// no additional processing.
 				for (uint8_t tg = 0; tg < Channels; tg++)
-					{
-						if(nMasterVolume >0.0 && nMasterVolume <1.0)
 				{
 					tmp_float[(i*Channels)+tg]=m_OutputLevel[tg][i] * nMasterVolume;
 				}
-						else if(nMasterVolume == 1.0)
-						{
-							tmp_float[(i*Channels)+tg]=m_OutputLevel[tg][i];
-						}
-					}
 			}
+
 			arm_float_to_q23(tmp_float,tmp_int,nFrames*Channels);
-			}
-			else
-			{
-				arm_fill_q31(0, tmp_int, nFrames*Channels);
-			}
 
 			// Prevent PCM510x analog mute from kicking in
 			for (uint8_t tg = 0; tg < Channels; tg++) 
@@ -1404,36 +1392,35 @@ void CMiniDexed::ProcessSound (void)
 			float32_t tmp_float[nFrames*2];
 			int32_t tmp_int[nFrames*2];
 
-			if(nMasterVolume > 0.0)
-			{
+			// get the mix buffer of all TGs
+			float32_t *SampleBuffer[2];
+			tg_mixer->getBuffers(SampleBuffer);
+
+			tg_mixer->zeroFill();
+
 			for (uint8_t i = 0; i < m_nToneGenerators; i++)
 			{
 				tg_mixer->doAddMix(i,m_OutputLevel[i]);
-					reverb_send_mixer->doAddMix(i,m_OutputLevel[i]);
 			}
 			// END TG mixing
 
-				// BEGIN create SampleBuffer for holding audio data
-				float32_t SampleBuffer[2][nFrames];
-				// END create SampleBuffer for holding audio data
-
-				// get the mix of all TGs
-				tg_mixer->getMix(SampleBuffer[indexL], SampleBuffer[indexR]);
-
 			// BEGIN adding reverb
 			if (m_nParameter[ParameterReverbEnable])
 			{
 				float32_t ReverbBuffer[2][nFrames];
-					float32_t ReverbSendBuffer[2][nFrames];
 
-					arm_fill_f32(0.0f, ReverbBuffer[indexL], nFrames);
-					arm_fill_f32(0.0f, ReverbBuffer[indexR], nFrames);
-					arm_fill_f32(0.0f, ReverbSendBuffer[indexR], nFrames);
-					arm_fill_f32(0.0f, ReverbSendBuffer[indexL], nFrames);
+				float32_t *ReverbSendBuffer[2];
+				reverb_send_mixer->getBuffers(ReverbSendBuffer);
+
+				reverb_send_mixer->zeroFill();
+
+				for (uint8_t i = 0; i < m_nToneGenerators; i++)
+				{
+					reverb_send_mixer->doAddMix(i,m_OutputLevel[i]);
+				}
 
 				m_ReverbSpinLock.Acquire ();
 
-					reverb_send_mixer->getMix(ReverbSendBuffer[indexL], ReverbSendBuffer[indexR]);
 				reverb->doReverb(ReverbSendBuffer[indexL],ReverbSendBuffer[indexR],ReverbBuffer[indexL], ReverbBuffer[indexR],nFrames);
 
 				// scale down and add left reverb buffer by reverb level 
@@ -1455,25 +1442,9 @@ void CMiniDexed::ProcessSound (void)
 			}
 
 			// Convert dual float array (left, right) to single int16 array (left/right)
-				for(uint16_t i=0; i<nFrames;i++)
-				{
-					if(nMasterVolume >0.0 && nMasterVolume <1.0)
-					{
-						tmp_float[i*2]=SampleBuffer[indexL][i] * nMasterVolume;
-						tmp_float[(i*2)+1]=SampleBuffer[indexR][i] * nMasterVolume;
-					}
-					else if(nMasterVolume == 1.0)
-					{
-						tmp_float[i*2]=SampleBuffer[indexL][i];
-						tmp_float[(i*2)+1]=SampleBuffer[indexR][i];
-					}
-				}
+			arm_scale_zip_f32(SampleBuffer[indexL], SampleBuffer[indexR], nMasterVolume, tmp_float, nFrames);
+
 			arm_float_to_q23(tmp_float,tmp_int,nFrames*2);
-			}
-			else
-			{
-				arm_fill_q31(0, tmp_int, nFrames * 2);
-			}
 
 			// Prevent PCM510x analog mute from kicking in
 			if (tmp_int[nFrames * 2 - 1] == 0)