DEV-11516 FIXED VWR-5308: Current wind noise generation is CPU intensive.

Minor optimizations to wind generation.  Patch by gigs and aimee, reviewed by merov.

7 files changed, 185 insertions, 120 deletions

diff --git a/indra/llaudio/llaudioengine.cpp b/indra/llaudio/llaudioengine.cpp
index b92ccd1d77..9f4c108dff 100644
--- a/indra/llaudio/llaudioengine.cpp
+++ b/indra/llaudio/llaudioengine.cpp
@@ -548,12 +548,11 @@ void LLAudioEngine::enableWind(bool enable)
 {
 	if (enable && (!mEnableWind))
 	{
-		initWind();
-		mEnableWind = enable;
+		mEnableWind = initWind();
 	}
 	else if (mEnableWind && (!enable))
 	{
-		mEnableWind = enable;
+		mEnableWind = false;
 		cleanupWind();
 	}
 }
diff --git a/indra/llaudio/llaudioengine.h b/indra/llaudio/llaudioengine.h
index d287104204..5876cef4ea 100644
--- a/indra/llaudio/llaudioengine.h
+++ b/indra/llaudio/llaudioengine.h
@@ -195,7 +195,7 @@ protected:
 	virtual LLAudioBuffer *createBuffer() = 0;
 	virtual LLAudioChannel *createChannel() = 0;
 
-	virtual void initWind() = 0;
+	virtual bool initWind() = 0;
 	virtual void cleanupWind() = 0;
 	virtual void setInternalGain(F32 gain) = 0;
 
diff --git a/indra/llaudio/llaudioengine_fmod.cpp b/indra/llaudio/llaudioengine_fmod.cpp
index d7f58defca..7a8a04afa1 100644
--- a/indra/llaudio/llaudioengine_fmod.cpp
+++ b/indra/llaudio/llaudioengine_fmod.cpp
@@ -54,13 +54,12 @@ extern "C" {
 	void * F_CALLBACKAPI windCallback(void *originalbuffer, void *newbuffer, int length, void* userdata);
 }
 
-FSOUND_DSPUNIT *gWindDSP = NULL;
-
 
 LLAudioEngine_FMOD::LLAudioEngine_FMOD()
 {
 	mInited = false;
 	mWindGen = NULL;
+	mWindDSP = NULL;
 }
 
 
@@ -258,10 +257,10 @@ void LLAudioEngine_FMOD::allocateListener(void)
 
 void LLAudioEngine_FMOD::shutdown()
 {
-	if (gWindDSP)
+	if (mWindDSP)
 	{
-		FSOUND_DSP_SetActive(gWindDSP,false);
-		FSOUND_DSP_Free(gWindDSP);
+		FSOUND_DSP_SetActive(mWindDSP,false);
+		FSOUND_DSP_Free(mWindDSP);
 	}
 
 	stopInternetStream();
@@ -289,29 +288,66 @@ LLAudioChannel * LLAudioEngine_FMOD::createChannel()
 }
 
 
-void LLAudioEngine_FMOD::initWind()
+bool LLAudioEngine_FMOD::initWind()
 {
-	mWindGen = new LLWindGen<MIXBUFFERFORMAT>;
+	if (!mWindGen)
+	{
+		bool enable;
+		
+		switch (FSOUND_GetMixer())
+		{
+			case FSOUND_MIXER_MMXP5:
+			case FSOUND_MIXER_MMXP6:
+			case FSOUND_MIXER_QUALITY_MMXP5:
+			case FSOUND_MIXER_QUALITY_MMXP6:
+				enable = (typeid(MIXBUFFERFORMAT) == typeid(S16));
+				break;
+			case FSOUND_MIXER_BLENDMODE:
+				enable = (typeid(MIXBUFFERFORMAT) == typeid(S32));
+				break;
+			case FSOUND_MIXER_QUALITY_FPU:
+				enable = (typeid(MIXBUFFERFORMAT) == typeid(F32));
+				break;
+			default:
+				// FSOUND_GetMixer() does not return a valid mixer type on Darwin
+				LL_INFOS("AppInit") << "Unknown FMOD mixer type, assuming default" << LL_ENDL;
+				enable = true;
+				break;
+		}
+		
+		if (enable)
+		{
+			mWindGen = new LLWindGen<MIXBUFFERFORMAT>(FSOUND_GetOutputRate());
+		}
+		else
+		{
+			LL_WARNS("AppInit") << "Incompatible FMOD mixer type, wind noise disabled" << LL_ENDL;
+		}
+	}
+
+	mNextWindUpdate = 0.0;
 
-	if (!gWindDSP)
+	if (mWindGen && !mWindDSP)
 	{
-		gWindDSP = FSOUND_DSP_Create(&windCallback, FSOUND_DSP_DEFAULTPRIORITY_CLEARUNIT + 20, mWindGen);
+		mWindDSP = FSOUND_DSP_Create(&windCallback, FSOUND_DSP_DEFAULTPRIORITY_CLEARUNIT + 20, mWindGen);
 	}
-	if (gWindDSP)
+	if (mWindDSP)
 	{
-		FSOUND_DSP_SetActive(gWindDSP, true);
+		FSOUND_DSP_SetActive(mWindDSP, true);
+		return true;
 	}
-	mNextWindUpdate = 0.0;
+	
+	return false;
 }
 
 
 void LLAudioEngine_FMOD::cleanupWind()
 {
-	if (gWindDSP)
+	if (mWindDSP)
 	{
-		FSOUND_DSP_SetActive(gWindDSP, false);
-		FSOUND_DSP_Free(gWindDSP);
-		gWindDSP = NULL;
+		FSOUND_DSP_SetActive(mWindDSP, false);
+		FSOUND_DSP_Free(mWindDSP);
+		mWindDSP = NULL;
 	}
 
 	delete mWindGen;
@@ -740,30 +776,12 @@ void * F_CALLBACKAPI windCallback(void *originalbuffer, void *newbuffer, int len
 	// originalbuffer = fmod's original mixbuffer.
 	// newbuffer = the buffer passed from the previous DSP unit.
 	// length = length in samples at this mix time.
-	// param = user parameter passed through in FSOUND_DSP_Create.
-	//
-	// modify the buffer in some fashion
+	// userdata = user parameter passed through in FSOUND_DSP_Create.
 
 	LLWindGen<LLAudioEngine_FMOD::MIXBUFFERFORMAT> *windgen =
 		(LLWindGen<LLAudioEngine_FMOD::MIXBUFFERFORMAT> *)userdata;
-	U8 stride;
-
-#if LL_DARWIN
-	stride = sizeof(LLAudioEngine_FMOD::MIXBUFFERFORMAT);
-#else
-	int mixertype = FSOUND_GetMixer();
-	if (mixertype == FSOUND_MIXER_BLENDMODE ||
-	    mixertype == FSOUND_MIXER_QUALITY_FPU)
-	{
-		stride = 4;
-	}
-	else
-	{
-		stride = 2;
-	}
-#endif
-
-	newbuffer = windgen->windGenerate((LLAudioEngine_FMOD::MIXBUFFERFORMAT *)newbuffer, length, stride);
+	
+	newbuffer = windgen->windGenerate((LLAudioEngine_FMOD::MIXBUFFERFORMAT *)newbuffer, length);
 
 	return newbuffer;
 }
diff --git a/indra/llaudio/llaudioengine_fmod.h b/indra/llaudio/llaudioengine_fmod.h
index 3968657cba..0e386a3884 100644
--- a/indra/llaudio/llaudioengine_fmod.h
+++ b/indra/llaudio/llaudioengine_fmod.h
@@ -55,15 +55,15 @@ public:
 
 	virtual void shutdown();
 
-	/*virtual*/ void initWind();
+	/*virtual*/ bool initWind();
 	/*virtual*/ void cleanupWind();
 
 	/*virtual*/void updateWind(LLVector3 direction, F32 camera_height_above_water);
 
 #if LL_DARWIN
-        typedef S32 MIXBUFFERFORMAT;
+	typedef S32 MIXBUFFERFORMAT;
 #else
-        typedef S16 MIXBUFFERFORMAT;
+	typedef S16 MIXBUFFERFORMAT;
 #endif
 
 protected:
@@ -83,6 +83,7 @@ protected:
 	void* mUserData;
 
 	LLWindGen<MIXBUFFERFORMAT> *mWindGen;
+	FSOUND_DSPUNIT *mWindDSP;
 };
 
 
diff --git a/indra/llaudio/llaudioengine_openal.cpp b/indra/llaudio/llaudioengine_openal.cpp
index a5982ccbd6..887c791790 100644
--- a/indra/llaudio/llaudioengine_openal.cpp
+++ b/indra/llaudio/llaudioengine_openal.cpp
@@ -370,7 +370,7 @@ U32 LLAudioBufferOpenAL::getLength()
 
 // ------------
 
-void LLAudioEngine_OpenAL::initWind()
+bool LLAudioEngine_OpenAL::initWind()
 {
 	ALenum error;
 	llinfos << "LLAudioEngine_OpenAL::initWind() start" << llendl;
@@ -397,10 +397,12 @@ void LLAudioEngine_OpenAL::initWind()
 	if(mWindBuf==NULL)
 	{
 		llerrs << "LLAudioEngine_OpenAL::initWind() Error creating wind memory buffer" << llendl;
-		mEnableWind=false;
+		return false;
 	}
 
 	llinfos << "LLAudioEngine_OpenAL::initWind() done" << llendl;
+
+	return true;
 }
 
 void LLAudioEngine_OpenAL::cleanupWind()
@@ -508,14 +510,14 @@ void LLAudioEngine_OpenAL::updateWind(LLVector3 wind_vec, F32 camera_altitude)
 		alGenBuffers(1,&buffer);
 		if((error=alGetError()) != AL_NO_ERROR)
 		{
-			llwarns << "LLAudioEngine_OpenAL::initWind() Error creating wind buffer: " << error << llendl;
+			llwarns << "LLAudioEngine_OpenAL::updateWind() Error creating wind buffer: " << error << llendl;
 			break;
 		}
 
 		alBufferData(buffer,
 			     AL_FORMAT_STEREO16,
 			     mWindGen->windGenerate(mWindBuf,
-						    mWindBufSamples, 2),
+						    mWindBufSamples),
 			     mWindBufBytes,
 			     mWindBufFreq);
 		error = alGetError();
diff --git a/indra/llaudio/llaudioengine_openal.h b/indra/llaudio/llaudioengine_openal.h
index 5aca03e195..16125b2476 100644
--- a/indra/llaudio/llaudioengine_openal.h
+++ b/indra/llaudio/llaudioengine_openal.h
@@ -57,23 +57,23 @@ class LLAudioEngine_OpenAL : public LLAudioEngine
 		LLAudioBuffer* createBuffer();
 		LLAudioChannel* createChannel();
 
-		/*virtual*/ void initWind();
+		/*virtual*/ bool initWind();
 		/*virtual*/ void cleanupWind();
 		/*virtual*/ void updateWind(LLVector3 direction, F32 camera_altitude);
 
 	private:
 		void * windDSP(void *newbuffer, int length);
-	        typedef S16 WIND_SAMPLE_T;
-        	LLWindGen<WIND_SAMPLE_T> *mWindGen;
-        	S16 *mWindBuf;
-        	U32 mWindBufFreq;
-        	U32 mWindBufSamples;
-        	U32 mWindBufBytes;
-        	ALuint mWindSource;
-	        int mNumEmptyWindALBuffers;
-
-        	static const int MAX_NUM_WIND_BUFFERS = 80;
-        	static const float WIND_BUFFER_SIZE_SEC = 0.05f; // 1/20th sec
+        typedef S16 WIND_SAMPLE_T;
+    	LLWindGen<WIND_SAMPLE_T> *mWindGen;
+    	S16 *mWindBuf;
+    	U32 mWindBufFreq;
+    	U32 mWindBufSamples;
+    	U32 mWindBufBytes;
+    	ALuint mWindSource;
+        int mNumEmptyWindALBuffers;
+
+    	static const int MAX_NUM_WIND_BUFFERS = 80;
+    	static const float WIND_BUFFER_SIZE_SEC = 0.05f; // 1/20th sec
 };
 
 class LLAudioChannelOpenAL : public LLAudioChannel
diff --git a/indra/llaudio/llwindgen.h b/indra/llaudio/llwindgen.h
index 847bfa6e9d..1908b2545f 100644
--- a/indra/llaudio/llwindgen.h
+++ b/indra/llaudio/llwindgen.h
@@ -33,104 +33,149 @@
 #define WINDGEN_H
 
 #include "llcommon.h"
-#include "llrand.h"
 
 template <class MIXBUFFERFORMAT_T>
 class LLWindGen
 {
 public:
-	LLWindGen() :
+	LLWindGen(const U32 sample_rate = 44100) :
 		mTargetGain(0.f),
 		mTargetFreq(100.f),
 		mTargetPanGainR(0.5f),
-		mbuf0(0.0),
-		mbuf1(0.0),
-		mbuf2(0.0),
-		mbuf3(0.0),
-		mbuf4(0.0),
-		mbuf5(0.0),
-		mY0(0.0),
-		mY1(0.0),
+		mInputSamplingRate(sample_rate),
+		mSubSamples(2),
+		mFilterBandWidth(50.f),
+		mBuf0(0.0f),
+		mBuf1(0.0f),
+		mBuf2(0.0f),
+		mY0(0.0f),
+		mY1(0.0f),
 		mCurrentGain(0.f),
 		mCurrentFreq(100.f),
-		mCurrentPanGainR(0.5f) {};
-
-	static const U32 getInputSamplingRate() {return mInputSamplingRate;}
+		mCurrentPanGainR(0.5f)
+	{
+		mSamplePeriod = (F32)mSubSamples / (F32)mInputSamplingRate;
+		mB2 = expf(-F_TWO_PI * mFilterBandWidth * mSamplePeriod);
+	}
 
+	const U32 getInputSamplingRate() { return mInputSamplingRate; }
+	
 	// newbuffer = the buffer passed from the previous DSP unit.
 	// numsamples = length in samples-per-channel at this mix time.
-	// stride = number of bytes between start of each sample.
 	// NOTE: generates L/R interleaved stereo
-	MIXBUFFERFORMAT_T* windGenerate(MIXBUFFERFORMAT_T *newbuffer, int numsamples, int stride)
+	MIXBUFFERFORMAT_T* windGenerate(MIXBUFFERFORMAT_T *newbuffer, int numsamples)
 	{
-		U8 *cursamplep = (U8*)newbuffer;
+		MIXBUFFERFORMAT_T *cursamplep = newbuffer;
+		
+		// Filter coefficients
+		F32 a0 = 0.0f, b1 = 0.0f;
 		
-		double bandwidth = 50.0F;
-		double a0,b1,b2;
+		// No need to clip at normal volumes
+		bool clip = mCurrentGain > 2.0f;
 		
-		// calculate resonant filter coeffs
-		b2 = exp(-(F_TWO_PI) * (bandwidth / mInputSamplingRate));
+		bool interp_freq = false; 
 		
-		while (numsamples--)
+		//if the frequency isn't changing much, we don't need to interpolate in the inner loop
+		if (llabs(mTargetFreq - mCurrentFreq) < (mCurrentFreq * 0.112))
 		{
-			mCurrentFreq = (float)((0.999 * mCurrentFreq) + (0.001 * mTargetFreq));
-			mCurrentGain = (float)((0.999 * mCurrentGain) + (0.001 * mTargetGain));
-			mCurrentPanGainR = (float)((0.999 * mCurrentPanGainR) + (0.001 * mTargetPanGainR));
-			b1 = (-4.0 * b2) / (1.0 + b2) * cos(F_TWO_PI * (mCurrentFreq / mInputSamplingRate));
-			a0 = (1.0 - b2) * sqrt(1.0 - (b1 * b1) / (4.0 * b2));
-			double nextSample;
+			// calculate resonant filter coefficients
+			mCurrentFreq = mTargetFreq;
+			b1 = (-4.0f * mB2) / (1.0f + mB2) * cosf(F_TWO_PI * (mCurrentFreq * mSamplePeriod));
+			a0 = (1.0f - mB2) * sqrtf(1.0f - (b1 * b1) / (4.0f * mB2));
+		}
+		else
+		{
+			interp_freq = true;
+		}
+		
+		while (numsamples)
+		{
+			F32 next_sample;
+			
+			// Start with white noise
+			// This expression is fragile, rearrange it and it will break!
+			next_sample = (F32)rand() * (1.0f / (F32)(RAND_MAX / (U16_MAX / 8))) + (F32)(S16_MIN / 8);
 			
-			// start with white noise
-			nextSample = ll_frand(2.0f) - 1.0f;
+			// Apply a pinking filter
+			// Magic numbers taken from PKE method at http://www.firstpr.com.au/dsp/pink-noise/
+			mBuf0 = mBuf0 * 0.99765f + next_sample * 0.0990460f;
+			mBuf1 = mBuf1 * 0.96300f + next_sample * 0.2965164f;
+			mBuf2 = mBuf2 * 0.57000f + next_sample * 1.0526913f;
 			
-			// apply pinking filter
-			mbuf0 = 0.997f * mbuf0 + 0.0126502f * nextSample; 
-			mbuf1 = 0.985f * mbuf1 + 0.0139083f * nextSample;
-			mbuf2 = 0.950f * mbuf2 + 0.0205439f * nextSample;
-			mbuf3 = 0.850f * mbuf3 + 0.0387225f * nextSample;
-			mbuf4 = 0.620f * mbuf4 + 0.0465932f * nextSample;
-			mbuf5 = 0.250f * mbuf5 + 0.1093477f * nextSample;
+			next_sample = mBuf0 + mBuf1 + mBuf2 + next_sample * 0.1848f;
 			
-			nextSample = mbuf0 + mbuf1 + mbuf2 + mbuf3 + mbuf4 + mbuf5;
+			if (interp_freq)
+			{
+				// calculate and interpolate resonant filter coefficients
+				mCurrentFreq = (0.999f * mCurrentFreq) + (0.001f * mTargetFreq);
+				b1 = (-4.0f * mB2) / (1.0f + mB2) * cosf(F_TWO_PI * (mCurrentFreq * mSamplePeriod));
+				a0 = (1.0f - mB2) * sqrtf(1.0f - (b1 * b1) / (4.0f * mB2));
+			}
 			
-			// do a resonant filter on the noise
-			nextSample = (double)( a0 * nextSample - b1 * mY0 - b2 * mY1 );
+			// Apply a resonant low-pass filter on the pink noise
+	        next_sample = a0 * next_sample - b1 * mY0 - mB2 * mY1;
 			mY1 = mY0;
-			mY0 = nextSample;
+			mY0 = next_sample;
 			
-			nextSample *= mCurrentGain;
+			mCurrentGain = (0.999f * mCurrentGain) + (0.001f * mTargetGain);
+			mCurrentPanGainR = (0.999f * mCurrentPanGainR) + (0.001f * mTargetPanGainR);
 			
-			MIXBUFFERFORMAT_T	sample;
+			// For a 3dB pan law use:
+			// next_sample *= mCurrentGain * ((mCurrentPanGainR*(mCurrentPanGainR-1)*1.652+1.413);
+		    next_sample *= mCurrentGain;
 			
-			sample = llfloor(((F32)nextSample*32768.f*(1.0f - mCurrentPanGainR))+0.5f);
-			*(MIXBUFFERFORMAT_T*)cursamplep = llclamp(sample, (MIXBUFFERFORMAT_T)-32768, (MIXBUFFERFORMAT_T)32767);
-			cursamplep += stride;
-
-			sample = llfloor(((F32)nextSample*32768.f*mCurrentPanGainR)+0.5f);
-			*(MIXBUFFERFORMAT_T*)cursamplep = llclamp(sample, (MIXBUFFERFORMAT_T)-32768, (MIXBUFFERFORMAT_T)32767);
-			cursamplep += stride;
+			// delta is used to interpolate between synthesized samples
+			F32 delta = (next_sample - mLastSample) / (F32)mSubSamples;
+			
+			// Fill the audio buffer, clipping if necessary
+			for (U8 i=mSubSamples; i && numsamples; --i, --numsamples) 
+			{
+				mLastSample = mLastSample + delta;
+				S32	sample_right = (S32)(mLastSample * mCurrentPanGainR);
+				S32	sample_left = (S32)mLastSample - sample_right;
+				
+				if (!clip)
+				{
+					*cursamplep = (MIXBUFFERFORMAT_T)sample_left;
+					++cursamplep;
+					*cursamplep = (MIXBUFFERFORMAT_T)sample_right;
+					++cursamplep;
+				}
+				else
+				{
+					*cursamplep = (MIXBUFFERFORMAT_T)llclamp(sample_left, (S32)S16_MIN, (S32)S16_MAX);
+					++cursamplep;
+					*cursamplep = (MIXBUFFERFORMAT_T)llclamp(sample_right, (S32)S16_MIN, (S32)S16_MAX);
+					++cursamplep;
+				}
+			}
 		}
 		
 		return newbuffer;
 	}
-
+	
+public:
 	F32 mTargetGain;
 	F32 mTargetFreq;
 	F32 mTargetPanGainR;
-
+	
 private:
-	static const U32 mInputSamplingRate = 44100;
-	F64 mbuf0;
-	F64 mbuf1;
-	F64 mbuf2;
-	F64 mbuf3;
-	F64 mbuf4;
-	F64 mbuf5;
-	F64 mY0;
-	F64 mY1;
+	U32 mInputSamplingRate;
+	U8  mSubSamples;
+	F32 mSamplePeriod;
+	F32 mFilterBandWidth;
+	F32 mB2;
+	
+	F32 mBuf0;
+	F32 mBuf1;
+	F32 mBuf2;
+	F32 mY0;
+	F32 mY1;
+	
 	F32 mCurrentGain;
 	F32 mCurrentFreq;
 	F32 mCurrentPanGainR;
+	F32 mLastSample;
 };
 
 #endif