AppleDBDMAClip.c   [plain text]

#include <libkern/OSTypes.h>
#include <IOKit/IOReturn.h>
//#include <libkern/libkern.h>

#include <IOKit/audio/IOAudioTypes.h>
#include <IOKit/audio/IOAudioDebug.h>

#include "AppleDBDMAClipLib.h"	

#include "fp_internal.h"	

#pragma mark ------------------------ 
#pragma mark ••• Constants and Tables
#pragma mark ------------------------ 

//	floating point types
typedef	float	Float32;
typedef double	Float64;

float gOldSample 									= 0.0f;
#ifdef USE_DYNAMIC_DOWNSAMPLING
UInt32 gDRCDownsampleFactor							= 4;
#endif

static float zeroGaindBConvTable[] = {
    0.0631f,  0.0708f,  0.0794f,  0.0891f,  0.1000f,  0.1122f,  0.1259f, 
    0.1413f,  0.1585f,  0.1778f,  0.1995f,  0.2239f,  0.2512f,  0.2818f,
    0.3162f,  0.3548f,  0.3981f,  0.4467f,  0.5012f,  0.5623f,  0.6310f, 
    0.7079f,  0.7943f,  0.8913f,  1.0000f,  1.1220f,  1.2589f,  1.4125f, 
    1.5849f,  1.7783f,  1.9953f,  2.2387f,  2.5119f,  2.8184f,  3.1623f, 
    3.5481f,  3.9811f,  4.4668f,  5.0119f,  5.6234f,  6.3096f,  7.0795f, 
    7.9433f,  8.9125f,  10.0000f, 11.2202f, 12.5893f, 14.1254f, 15.8489f
};

// -12 dB to +12 dB in 0.5 dB steps
const UInt16 kInputGaindBConvTableOffset = 24;
static float inputGaindBConvTable[] = {
	0.251189f,	0.266073f,	0.281838f,	0.298538f,
	0.316228f,	0.334965f,	0.354813f,	0.375837f,
	0.398107f,	0.421697f,	0.446684f,	0.473151f,
	0.501187f,	0.530884f,	0.562341f,	0.595662f,
	0.630957f,	0.668344f,	0.707946f,	0.749894f,
	0.794328f,	0.841395f,	0.891251f,	0.944061f,
	1.000000f,	1.059254f,	1.122018f,	1.188502f,
	1.258925f,	1.333521f,	1.412538f,	1.496236f,
	1.584893f,	1.678804f,	1.778279f,	1.883649f,
	1.995262f,	2.113489f,	2.238721f,	2.371374f,
	2.511886f,	2.660725f,	2.818383f,	2.985383f,
	3.162278f,	3.349654f,	3.548134f,	3.758374f,
	3.981072f
};

static float aoa_log2table[] = {
	-9.96578428f,	-4.95455703f,	-3.97709960f,	-3.39973025f,
	-2.98850436f,	-2.66886808f,	-2.40736357f,	-2.18606493f,
	-1.99424073f,	-1.82495451f,	-1.67346265f,	-1.53637754f,
	-1.41119543f,	-1.29601340f,	-1.18935125f,	-1.09003493f,
	-0.99711749f,	-0.90982405f,	-0.82751248f,	-0.74964473f,
	-0.67576544f,	-0.60548586f,	-0.53847144f,	-0.47443221f,
	-0.41311519f,	-0.35429834f,	-0.29778575f,	-0.24340365f,
	-0.19099723f,	-0.14042794f,	-0.09157135f,	-0.04431522f,
	0.00144197f,	0.04579242f,	0.08882003f,	0.13060145f,
	0.17120683f,	0.21070056f,	0.24914190f,	0.28658548f,
	0.32308179f,	0.35867757f,	0.39341620f,	0.42733799f,
	0.46048047f,	0.49287865f,	0.52456522f,	0.55557078f,
	0.58592398f,	0.61565170f,	0.64477922f,	0.67333028f,
	0.70132726f,	0.72879125f,	0.75574218f,	0.78219886f,
	0.80817908f,	0.83369972f,	0.85877675f,	0.88342532f,
	0.90765983f,	0.93149396f,	0.95494074f,	0.97801253f,
	1.00072117f,	1.02307789f,	1.04509344f,	1.06677807f,
	1.08814160f,	1.10919338f,	1.12994238f,	1.15039720f,
	1.17056606f,	1.19045683f,	1.21007710f,	1.22943411f,
	1.24853484f,	1.26738598f,	1.28599397f,	1.30436501f,
	1.32250506f,	1.34041984f,	1.35811490f,	1.37559554f,
	1.39286692f,	1.40993397f,	1.42680147f,	1.44347404f,
	1.45995614f,	1.47625206f,	1.49236596f,	1.50830187f,
	1.52406368f,	1.53965514f,	1.55507990f,	1.57034149f,
	1.58544332f,	1.60038870f,	1.61518085f,	1.62982287f,
	1.64431778f,	1.65866850f,	1.67287788f,	1.68694867f,
	1.70088356f,	1.71468513f,	1.72835592f,	1.74189838f,
	1.75531490f,	1.76860781f,	1.78177935f,	1.79483172f,
	1.80776706f,	1.82058746f,	1.83329492f,	1.84589144f,
	1.85837893f,	1.87075925f,	1.88303424f,	1.89520566f,
	1.90727526f,	1.91924473f,	1.93111570f,	1.94288979f,
	1.95456857f,	1.96615357f,	1.97764628f,	1.98904815f
};

static float aoa_antilog2table[] = {
	0.994599423f, 0.912051693f, 0.836355090f, 0.766940999f,
	0.703287997f, 0.644917937f, 0.591392355f, 0.542309181f,
	0.497299712f, 0.456025846f, 0.418177545f, 0.383470499f,
	0.351643998f, 0.322458968f, 0.295696178f, 0.271154591f,
	0.248649856f, 0.228012923f, 0.209088772f, 0.191735250f,
	0.175821999f, 0.161229484f, 0.147848089f, 0.135577295f,
	0.124324928f, 0.114006462f, 0.104544386f, 0.095867625f,
	0.087911000f, 0.080614742f, 0.073924044f, 0.067788648f,
	0.062162464f, 0.057003231f, 0.052272193f, 0.047933812f,
	0.043955500f, 0.040307371f, 0.036962022f, 0.033894324f,
	0.031081232f, 0.028501615f, 0.026136097f, 0.023966906f,
	0.021977750f, 0.020153686f, 0.018481011f, 0.016947162f,
	0.015540616f, 0.014250808f, 0.013068048f, 0.011983453f,
	0.010988875f, 0.010076843f, 0.009240506f, 0.008473581f,
	0.007770308f, 0.007125404f, 0.006534024f, 0.005991727f,
	0.005494437f, 0.005038421f, 0.004620253f, 0.004236790f,
	0.003885154f, 0.003562702f, 0.003267012f, 0.002995863f,
	0.002747219f, 0.002519211f, 0.002310126f, 0.002118395f,
	0.001942577f, 0.001781351f, 0.001633506f, 0.001497932f,
	0.001373609f, 0.001259605f, 0.001155063f, 0.001059198f,
	0.000971288f, 0.000890675f, 0.000816753f, 0.000748966f,
	0.000686805f, 0.000629803f, 0.000577532f, 0.000529599f,
	0.000485644f, 0.000445338f, 0.000408377f, 0.000374483f,
	0.000343402f, 0.000314901f, 0.000288766f, 0.000264799f,
	0.000242822f, 0.000222669f, 0.000204188f, 0.000187241f,
	0.000171701f, 0.000157451f, 0.000144383f, 0.000132400f,
	0.000121411f, 0.000111334f, 0.000102094f, 0.000093621f,
	0.000085851f, 0.000078725f, 0.000072191f, 0.000066200f,
	0.000060706f, 0.000055667f, 0.000051047f, 0.000046810f,
	0.000042925f, 0.000039363f, 0.000036096f, 0.000033100f,
	0.000030353f, 0.000027834f, 0.000025524f, 0.000023405f,
	0.000021463f, 0.000019681f, 0.000018048f, 0.000016550f
};

static float aoa_antilog2table_expand[] = {	// handles input signals at threshold and down to -60 below
	1.000000,	1.055645,	1.114387,	1.176397,
	1.241858,	1.310961,	1.383910,	1.460918,
	1.542211,	1.628027,	1.718619,	1.814252,
	1.915207,	2.021779,	2.134281,	2.253043,
	2.378414,	2.510762,	2.650473,	2.797959,
	2.953652,	3.118009,	3.291511,	3.474668,
	3.668016,	3.872124,	4.087589,	4.315043,
	4.555155,	4.808627,	5.076204,	5.358670,
	5.656854,	5.971631,	6.303923,	6.654706,
	7.025009,	7.415917,	7.828576,	8.264199,
	8.724062,	9.209514,	9.721979,	10.262960,
	10.834044,	11.436907,	12.073315,	12.745137,
	13.454343,	14.203012,	14.993341,	15.827648,
	16.708381,	17.638121,	18.619598,	19.655689,
	20.749433,	21.904039,	23.122893,	24.409570,
	25.767845,	27.201702,	28.715345,	30.313216,
	32.000000,	33.780646,	35.660376,	37.644704,
	39.739450,	41.950759,	44.285116,	46.749369,
	49.350746,	52.096877,	54.995818,	58.056070,
	61.286610,	64.696914,	68.296986,	72.097384,
	76.109255,	80.344368,	84.815145,	89.534699,
	94.516873,	99.776282,	105.328351,	111.189365,
	117.376518,	123.907955,	130.802835,	138.081382,
	145.764945,	153.876062,	162.438523,	171.477443,
	181.019336,	191.092189,	201.725548,	212.950602,
	224.800277,	237.309328,	250.514448,	264.454369,
	279.169980,	294.704443,	311.103324,	328.414724,
	346.689420,	365.981015,	386.346093,	407.844391,
	430.538965,	454.496382,	479.786914,	506.484743,
	534.668177,	564.419883,	595.827128,	628.982034,
	663.981852,	700.929241,	739.932573,	781.106253,
	824.571050,	870.454453,	918.891046,	970.022903
};
const static UInt16 kAOAAntiLog2TableRatioExpand		= 12.800256081f;


#pragma mark ------------------------ 
#pragma mark ••• Processing Routines
#pragma mark ------------------------ 

// ------------------------------------------------------------------------
// Delay right or left channel
// ------------------------------------------------------------------------
void delayLR(float* inFloatBufferPtr, float* outFloatBufferPtr, UInt32 numSamples, DelayStructPtr inDelay) 
{
	register UInt32 i, writeIndex, bufferMaxIndex, delayTimeInt, numFrames;
    register float* inPtr;
    register float* outPtr;
	register float* delayBuffer;
	register float delayFrac, temp, inL, inR;
	register SInt32 delayIndexInt;
	
	inPtr = inFloatBufferPtr;
	outPtr = outFloatBufferPtr;
	
	delayBuffer = inDelay->buffer;
	delayTimeInt = inDelay->delayTimeInt;
	delayFrac = inDelay->delayFrac;
	writeIndex = inDelay->writeIndex;
	bufferMaxIndex = kNumDelayBufferSamples - 1;
	
	numFrames = numSamples >> 1;

	delayIndexInt = (SInt32)writeIndex - (SInt32)delayTimeInt;			
	if (delayIndexInt < 0) {
		delayIndexInt += kNumDelayBufferSamples;
	}

	if (0 == inDelay->channel) {
		for (i = 0; i < numFrames; i++ ) {
			delayBuffer[writeIndex] = *(inPtr++);
			inR = *(inPtr++);

			inL = delayBuffer[delayIndexInt];
			if (delayIndexInt == 0) {
				temp = delayBuffer[bufferMaxIndex];
			} else {
				temp = delayBuffer[delayIndexInt - 1];
			}
			temp -= inL;
			temp *= delayFrac;
			inL += temp;

			*(outPtr++) = inL;
			*(outPtr++) = inR;		

			writeIndex++;
			if (writeIndex > bufferMaxIndex) {
				writeIndex = 0;
			}
			delayIndexInt++;
			if (delayIndexInt > bufferMaxIndex) {
				delayIndexInt = 0;
			}
		}
	} else if (1 == inDelay->channel) {
		for (i = 0; i < numFrames; i++ ) {
			inL = *(inPtr++);
			delayBuffer[writeIndex] = *(inPtr++);

			inR = delayBuffer[delayIndexInt];
			if (delayIndexInt == 0) {
				temp = delayBuffer[bufferMaxIndex];
			} else {
				temp = delayBuffer[delayIndexInt - 1];
			}
			temp -= inR;
			temp *= delayFrac;
			inR += temp;

			*(outPtr++) = inL;
			*(outPtr++) = inR;		

			writeIndex++;
			if (writeIndex > bufferMaxIndex) {
				writeIndex = 0;
			}
			delayIndexInt++;
			if (delayIndexInt > bufferMaxIndex) {
				delayIndexInt = 0;
			}
		}
	}
	inDelay->writeIndex = writeIndex;
	inDelay->delayIndexInt = delayIndexInt;
}		

void volume (float* inFloatBufferPtr, UInt32 numSamples, float* inLeftVolume, float* inRightVolume, float* inPreviousLeftVolume, float* inPreviousRightVolume)
{
	register UInt32 i;
	register UInt32 numFrames;
	register UInt32 leftOver;
    register float* inPtr;
    register float* outPtr;
	register float leftGain;
	register float rightGain;
	register float inL0;
	register float inR0;
	register float inL1;
	register float inR1;
	register float inL2;
	register float inR2;
	register float inL3;
	register float inR3;
	register float timeConstant, oneMinusTimeConstant, oldLeftGain, oldRightGain, leftGain_s, rightGain_s;
	
	inPtr = inFloatBufferPtr;
	outPtr = inFloatBufferPtr;  
	
	leftGain = *inLeftVolume;
	rightGain = *inRightVolume;
	
	numFrames = numSamples >> 1;
	leftOver = numFrames % 4;
	numSamples = numFrames >> 2;

	timeConstant = 0.000453411f;
	oneMinusTimeConstant = 1.0f - timeConstant;

	oldLeftGain = *inPreviousLeftVolume;
	oldRightGain = *inPreviousRightVolume;

    for (i = 0; i < numSamples; i++ ) 
    {
		leftGain_s = leftGain*timeConstant;
		rightGain_s = rightGain*timeConstant;

		leftGain_s += oneMinusTimeConstant*oldLeftGain;
		rightGain_s += oneMinusTimeConstant*oldRightGain;

		oldLeftGain = leftGain_s;
		oldRightGain = rightGain_s;

		inL0 = *(inPtr++);					

		inR0 = *(inPtr++);			

		inL1 = *(inPtr++);			
		inL0 *= leftGain_s;

		inR1 = *(inPtr++);		
		inR0 *= rightGain_s;

		leftGain_s = leftGain*timeConstant;
		rightGain_s = rightGain*timeConstant;

		leftGain_s += oneMinusTimeConstant*oldLeftGain;
		rightGain_s += oneMinusTimeConstant*oldRightGain;

		oldLeftGain = leftGain_s;
		oldRightGain = rightGain_s;

		inL2 = *(inPtr++);					
		inL1 *= leftGain_s;
		*(outPtr++) = inL0;			

		inR2 = *(inPtr++);			
		inR1 *= rightGain_s;
		*(outPtr++) = inR0;		

		leftGain_s = leftGain*timeConstant;
		rightGain_s = rightGain*timeConstant;

		leftGain_s += oneMinusTimeConstant*oldLeftGain;
		rightGain_s += oneMinusTimeConstant*oldRightGain;

		oldLeftGain = leftGain_s;
		oldRightGain = rightGain_s;

		inL3 = *(inPtr++);			
		inL2 *= leftGain_s;
		*(outPtr++) = inL1;		

		inR3 = *(inPtr++);		
		inR2 *= rightGain_s;
		*(outPtr++) = inR1;		

		leftGain_s = leftGain*timeConstant;
		rightGain_s = rightGain*timeConstant;

		leftGain_s += oneMinusTimeConstant*oldLeftGain;
		rightGain_s += oneMinusTimeConstant*oldRightGain;

		oldLeftGain = leftGain_s;
		oldRightGain = rightGain_s;

		inL3 *= leftGain_s;
		*(outPtr++) = inL2;			

		inR3 *= rightGain_s;
		*(outPtr++) = inR2;				

		*(outPtr++) = inL3;

		*(outPtr++) = inR3;		
	}

    for (i = 0; i < leftOver; i ++ ) 
    {
		leftGain_s = leftGain*timeConstant;
		rightGain_s = rightGain*timeConstant;
		leftGain_s += oneMinusTimeConstant*oldLeftGain;
		rightGain_s += oneMinusTimeConstant*oldRightGain;
		inL0 = *(inPtr++);
		inR0 = *(inPtr++);
		inL0 *= leftGain_s;
		inR0 *= rightGain_s;
		oldLeftGain = leftGain_s;
		oldRightGain = rightGain_s;
		*(outPtr++) = inL0;			
		*(outPtr++) = inR0;		
	}

	*inPreviousLeftVolume = oldLeftGain;
	*inPreviousRightVolume = oldRightGain;
}

// ------------------------------------------------------------------------
// Apply stereo gain, needed on machines like Q27 to compensate for
// different speaker output due to enclosure differences.
// ------------------------------------------------------------------------
void balanceAdjust(float* inFloatBufferPtr, float* outFloatBufferPtr, UInt32 numSamples, GainStructPtr inGain) 
{
	register UInt32 i;
	register UInt32 numFrames;
	register UInt32 leftOver;
    register float* inPtr;
    register float* outPtr;
	register float leftGain;
	register float rightGain;
	register float inL0;
	register float inR0;
	register float inL1;
	register float inR1;
	register float inL2;
	register float inR2;
	register float inL3;
	register float inR3;
	
	inPtr = inFloatBufferPtr;
	outPtr = outFloatBufferPtr;  
	
	leftGain = inGain->leftSoftVolume;
	rightGain = inGain->rightSoftVolume;
	
	numFrames = numSamples >> 1;
	leftOver = numFrames % 4;
	numSamples = numFrames >> 2;
	
    for (i = 0; i < numSamples; i++ ) 
    {
		inL0 = *(inPtr++);					

		inR0 = *(inPtr++);			

		inL1 = *(inPtr++);			
		inL0 *= leftGain;

		inR1 = *(inPtr++);		
		inR0 *= rightGain;

		inL2 = *(inPtr++);					
		inL1 *= leftGain;
		*(outPtr++) = inL0;			

		inR2 = *(inPtr++);			
		inR1 *= rightGain;
		*(outPtr++) = inR0;		

		inL3 = *(inPtr++);			
		inL2 *= leftGain;
		*(outPtr++) = inL1;		

		inR3 = *(inPtr++);		
		inR2 *= rightGain;
		*(outPtr++) = inR1;		

		inL3 *= leftGain;
		*(outPtr++) = inL2;			

		inR3 *= rightGain;
		*(outPtr++) = inR2;				

		*(outPtr++) = inL3;

		*(outPtr++) = inR3;		
	}

    for (i = 0; i < leftOver; i ++ ) 
    {
		inL0 = *(inPtr++);
		inR0 = *(inPtr++);
		inL0 *= leftGain;
		inR0 *= rightGain;
		*(outPtr++) = inL0;			
		*(outPtr++) = inR0;		
	}
}

// ------------------------------------------------------------------------
// Mix left and right channels together, and mute the right channel
// ------------------------------------------------------------------------
void mixAndMuteRightChannel(float* inFloatBufferPtr, float* outFloatBufferPtr, UInt32 numSamples) 
{
	UInt32 i, leftOver;
    register float* inPtr;
    register float* outPtr;
	register float inL0;
	register float inL1;
	register float inL2;
	register float inL3;
	register float inR0;
	register float inR1;
	register float inR2;
	register float inR3;
	
	inPtr = inFloatBufferPtr;  
	outPtr = outFloatBufferPtr;  
	
	leftOver = numSamples % 8;
	numSamples = numSamples >> 3;
	
    for (i = 0; i < numSamples; i++ ) 
    {
		inL0 = *(inPtr++);
		inR0 = *(inPtr++);
		
		inL1 = *(inPtr++);
		inR1 = *(inPtr++);
		inL0 += inR0;
		
		inL2 = *(inPtr++);
		inR2 = *(inPtr++);
		inL1 += inR1;
		inL0 *= kMixingToMonoScale;
		
		inL3 = *(inPtr++);
		inR3 = *(inPtr++);
		*(outPtr++) = inL0;
		*(outPtr++) = 0.0f;
		inL1 *= kMixingToMonoScale;
		inL2 += inR2;
		
		inL3 += inR3;
		*(outPtr++) = inL1;
		*(outPtr++) = 0.0f;
		inL2 *= kMixingToMonoScale;

		*(outPtr++) = inL2;
		*(outPtr++) = 0.0f;
		inL3 *= kMixingToMonoScale;

		*(outPtr++) = inL3;
		*(outPtr++) = 0.0f;
	}

    for (i = 0; i < leftOver; i += 2 ) 
    {
		inL0 = *(inPtr++);
		inR0 =  *(inPtr++);
		inL0 += inR0;
		inL0 *= kMixingToMonoScale;
		*(outPtr++) = inL0;
		*(outPtr++) = 0.0f;
	}
}

#pragma mark ------------------------ 
#pragma mark ••• iSub Processing Routines
#pragma mark ------------------------ 

void iSubDownSampleLinearAndConvert( float* inData, float* srcPhase, float* srcState, UInt32 adaptiveSampleRate, UInt32 outputSampleRate, UInt32 sampleIndex, UInt32 maxSampleIndex, SInt16 *iSubBufferMemory, SInt32 *iSubBufferOffset, UInt32 iSubBufferLen, UInt32 *loopCount )
{
    Float32		iSubSampleFloat;
    SInt16		iSubSampleInt;
	UInt32		baseIndex;
	float*		inDataPtr;
	
    float 		x0, x1, temp;
    float 		phaseInc;		// phase increment = Fs_in/Fs_out
    float 		phase;			// current phase location

	inDataPtr = inData;
	baseIndex = sampleIndex;
	phase = *srcPhase;
	phaseInc = ((float)adaptiveSampleRate)/((float)outputSampleRate);
	
	// linear interpolation src (good enough since we have a 4th order lp in front of us, 
	// down -90 dB at Nyquist for 6kHz sample rate)
	while (sampleIndex < maxSampleIndex) {
				
		if (phase >= 1.0f) {	
			phase -= 1.0f;
			sampleIndex+=2;
		} else {   
			// check for beginning of frame case, use saved last sample if needed
			if (sampleIndex == baseIndex) { 
				x0 = *srcState;
			} else {
				// mix x[n-1] to mono
				x0 = inDataPtr[sampleIndex-2];
				temp = inDataPtr[sampleIndex-1];
				x0 = 0.5f*(x0 + temp);
			}
			
			// mix x[n] to mono
			x1 = inDataPtr[sampleIndex];
			temp = inDataPtr[sampleIndex+1];
			x1 = 0.5f*(x1 + temp);
							
			// linearly interpolate between x0 and x1
			iSubSampleFloat = x0 + phase*(x1 - x0);
			
//#if 0
			// clip
			if (iSubSampleFloat > 1.0f) {
				iSubSampleFloat = 1.0f;
			} else if (iSubSampleFloat < -1.0f) {
				iSubSampleFloat = -1.0f;
			}
			
			// convert to fixed
			iSubSampleInt = (SInt16) (iSubSampleFloat * 32767.0f);
//#else			
//			scaled = iSubSampleFloat * scale + round;
//			converted = __fctiw( scaled );
//#endif
			
			// check for end of buffer condition
			if (*iSubBufferOffset >= (SInt32)iSubBufferLen) {
				*iSubBufferOffset = 0;
				(*loopCount)++;
			}
			
			// byteswap to USB format and copy to iSub buffer
			iSubBufferMemory[(*iSubBufferOffset)++] = ((((UInt16)iSubSampleInt) << 8) & 0xFF00) | ((((UInt16)iSubSampleInt) >> 8) & 0x00FF);

			// increment phase and update input buffer pointer
			phase += phaseInc;		
		}
	}
	if (phase < 1) {
		// mix and save last sample in buffer to mono if it will be needed for the next loop
		x1 = inDataPtr[maxSampleIndex-2];
		temp = inDataPtr[maxSampleIndex-1];
		*srcState = 0.5f*(x1 + temp);
	} else {
		*srcState = 0.0f;
	}
	// cache current phase for use next time we enter the clip loop
	*srcPhase = phase;
	
	return;
}

// fourth order coefficient setting functions
Boolean Set4thOrderCoefficients (Float32 *b0, Float32 *b1, Float32 *b2, Float32 *a1, Float32 *a2, UInt32 samplingRate)
{
    Boolean 	success = TRUE;

    switch ( samplingRate )
    {
        case 8000:  *b0 =  0.00782020803350;
                    *b1 =  0.01564041606699;
                    *b2 =  0.00782020803350;
                    *a1 = -1.73472576880928;
                    *a2 =  0.76600660094326;
                    break;
       case 11025:  *b0 =  0.00425905333005;
                    *b1 =  0.00851810666010;
                    *b2 =  0.00425905333005;
                    *a1 = -1.80709136077571;
                    *a2 =  0.82412757409590;
                    break;
       case 22050:  *b0 =  0.00111491512001;
                    *b1 =  0.00222983024003;
                    *b2 =  0.00111491512001;
                    *a1 = -1.90335434048751;
                    *a2 =  0.90781400096756;
                    break;
       case 32000:  *b0 =  0.00053716977481;
                    *b1 =  0.00107433954962;
                    *b2 =  0.00053716977481;
                    *a1 = -1.93338022587993;
                    *a2 =  0.93552890497918;
                    break;
	   case 44100: 	*b0 =  0.00028538351548666;
                    *b1 =  0.00057076703097332;
                    *b2 =  0.00028538351548666;
                    *a1 = -1.95165117996464;
                    *a2 =  0.95279271402659;
                    break;
       case 48000:  *b0 =  0.00024135904904198;
                    *b1 =  0.00048271809808396;
                    *b2 =  0.00024135904904198;
                    *a1 = -1.95557824031504;
                    *a2 =  0.95654367651120;
                    break;
        case 96000: *b0 =  0.00006100617875806425;
                    *b1 =  0.0001220123575161285;
                    *b2 =  0.00006100617875806425;
                    *a1 = -1.977786483776763;
                    *a2 =  0.9780305084917958;
                    break;
        default:    // debugIOLog (3, "\nNot a registered frequency...");
                    success = FALSE;
                    break;
    }

    return(success);
}

// this function sets the parameters of a second order all-pass filter that is used to compensate for the phase
// shift of the 4th order lowpass IIR filter used in the iSub crossover.  Note that a0 and b2 are both 1.0.
Boolean Set4thOrderPhaseCompCoefficients (Float32 *b0, Float32 *b1, Float32 *a1, Float32 *a2, UInt32 samplingRate)
{
    Boolean 	success = TRUE;

    switch ( samplingRate )
    {
        case 8000:  *a1 = -1.734725768809275;
                    *a2 =  0.7660066009432638;
                    *b0 =  *a2;
                    *b1 =  *a1;
                    break;
        case 11025: *a1 = -1.807091360775707;
                    *a2 =  0.8241275740958973;
                    *b0 =  *a2;
                    *b1 =  *a1;
                    break;
        case 22050: *a1 = -1.903354340487510;
                    *a2 =  0.9078140009675627;
                    *b0 =  *a2;
                    *b1 =  *a1;
                    break;
        case 32000: *a1 = -1.93338022587993;
                    *a2 =  0.93552890497918;
                    *b0 =  *a2;
                    *b1 =  *a1;
                    break;
        case 44100: *a1 = -1.951651179964643;
                    *a2 =  0.9527927140265903;
                    *b0 =  *a2;
                    *b1 =  *a1;
                    break;
        case 48000: *a1 = -1.955578240315035;
                    *a2 =  0.9565436765112033;
                    *b0 =  *a2;
                    *b1 =  *a1;
                    break;
        case 96000: *a1 = -1.977786483776763;
                    *a2 =  0.9780305084917958;
                    *b0 =  *a2;
                    *b1 =  *a1;
                    break;
        default:    
                    success = FALSE;
                    break;
    }

    return(success);
}

// stereo 4th order LR crossover
// this needs lots of optimization!
void StereoCrossover4thOrderPhaseComp (Float32 *in, Float32 *low, Float32 *high, UInt32 frames, UInt32 SamplingRate, PreviousValues *section1State, PreviousValues *section2State, PreviousValues *phaseCompState)
{
    UInt32	i;
    Float32	inL, inR, outL1, outR1, outL, outR, inPhaseCompL, inPhaseCompR;

    // shared coefficients for second order sections
    Float32	b0, b1, b2, a1, a2;
    // coefficients for phase compensator
    Float32	bp0, bp1, ap1, ap2;
    // taps for second order section 1
    Float32	inLTap1, inLTap2, inRTap1, inRTap2;
    Float32	outLTap1, outLTap2, outRTap1, outRTap2;
    // taps for second order section 2
    Float32	inLTap1_2, inLTap2_2, inRTap1_2, inRTap2_2;
    Float32	outLTap1_2, outLTap2_2, outRTap1_2, outRTap2_2;
    // taps for phase compensator
    Float32	inLTap1_p, inLTap2_p, inRTap1_p, inRTap2_p;
    Float32	outLTap1_p, outLTap2_p, outRTap1_p, outRTap2_p;

    // copy to state local variables to avoid structure referencing during inner loop
    // section 1
    inLTap1 = section1State->xl_1;
    inLTap2 = section1State->xl_2;
    inRTap1 = section1State->xr_1;
    inRTap2 = section1State->xr_2;

    outLTap1 = section1State->yl_1;
    outLTap2 = section1State->yl_2;
    outRTap1 = section1State->yr_1;
    outRTap2 = section1State->yr_2;
    
    // section 2
    inLTap1_2 = section2State->xl_1;
    inLTap2_2 = section2State->xl_2;
    inRTap1_2 = section2State->xr_1;
    inRTap2_2 = section2State->xr_2;

    outLTap1_2 = section2State->yl_1;
    outLTap2_2 = section2State->yl_2;
    outRTap1_2 = section2State->yr_1;
    outRTap2_2 = section2State->yr_2;
    
    // phase compensator
    inLTap1_p = phaseCompState->xl_1;
    inLTap2_p = phaseCompState->xl_2;
    inRTap1_p = phaseCompState->xr_1;
    inRTap2_p = phaseCompState->xr_2;

    outLTap1_p = phaseCompState->yl_1;
    outLTap2_p = phaseCompState->yl_2;
    outRTap1_p = phaseCompState->yr_1;
    outRTap2_p = phaseCompState->yr_2;
 
    // set all coefficients
    if (Set4thOrderCoefficients (&b0, &b1, &b2, &a1, &a2, SamplingRate) == FALSE)
        return;
    if (Set4thOrderPhaseCompCoefficients (&bp0, &bp1, &ap1, &ap2, SamplingRate) == FALSE)
        return;

    // need to unroll this loop to get rid of stalls!
    for ( i = 0 ; i < frames ; i ++ )
    {
        inL = in[2*i];
        inR = in[2*i+1];
        
        // Low-pass filter first pass
        outL1 = b0*inL + b1*inLTap1 + b2*inLTap2 - a1*outLTap1 - a2*outLTap2;
        outR1 = b0*inR + b1*inRTap1 + b2*inRTap2 - a1*outRTap1 - a2*outRTap2;
 
        // update section 1 filter taps
        inLTap2 = inLTap1;
        inRTap2 = inRTap1;
        inLTap1 = inL;
        inRTap1 = inR;
        outLTap2 = outLTap1;
        outRTap2 = outRTap1;
        outLTap1 = outL1;
        outRTap1 = outR1;
        
        // Low-pass filter second pass
        outL = b0*outL1 + b1*inLTap1_2 + b2*inLTap2_2 - a1*outLTap1_2 - a2*outLTap2_2;
        outR = b0*outR1 + b1*inRTap1_2 + b2*inRTap2_2 - a1*outRTap1_2 - a2*outRTap2_2;
        
        // update section 2 filter taps
        inLTap2_2 = inLTap1_2;
        inRTap2_2 = inRTap1_2;
        inLTap1_2 = outL1;
        inRTap1_2 = outR1;
        outLTap2_2 = outLTap1_2;
        outRTap2_2 = outRTap1_2;
        outLTap1_2 = outL;
        outRTap1_2 = outR;

        // phase compensate the input, note that b2 is 1.0
        inPhaseCompL = bp0*inL + bp1*inLTap1_p + inLTap2_p - ap1*outLTap1_p - ap2*outLTap2_p;
        inPhaseCompR = bp0*inR + bp1*inRTap1_p + inRTap2_p - ap1*outRTap1_p - ap2*outRTap2_p;
        
        // update phase compensate filter taps
        inLTap2_p = inLTap1_p;
        inRTap2_p = inRTap1_p;
        inLTap1_p = inL;
        inRTap1_p = inR;
        outLTap2_p = outLTap1_p;
        outRTap2_p = outRTap1_p;
        outLTap1_p = inPhaseCompL;
        outRTap1_p = inPhaseCompR;

        // Storage
        low[2*i] = outL;
        low[2*i+1] = outR;
        high[2*i] = inPhaseCompL-outL;
        high[2*i+1] = inPhaseCompR-outR;
    }

    // update state structures
    
    // section 1 state
    section1State->xl_1 = inLTap1;
    section1State->xl_2 = inLTap2;
    section1State->xr_1 = inRTap1;
    section1State->xr_2 = inRTap2;

    section1State->yl_1 = outLTap1;
    section1State->yl_2 = outLTap2;
    section1State->yr_1 = outRTap1;
    section1State->yr_2 = outRTap2;
    
    // section 2 state
    section2State->xl_1 = inLTap1_2;
    section2State->xl_2 = inLTap2_2;
    section2State->xr_1 = inRTap1_2;
    section2State->xr_2 = inRTap2_2;

    section2State->yl_1 = outLTap1_2;
    section2State->yl_2 = outLTap2_2;
    section2State->yr_1 = outRTap1_2;
    section2State->yr_2 = outRTap2_2;
    
    // phase compensator state
    phaseCompState->xl_1 = inLTap1_p;
    phaseCompState->xl_2 = inLTap2_p;
    phaseCompState->xr_1 = inRTap1_p;
    phaseCompState->xr_2 = inRTap2_p;

    phaseCompState->yl_1 = outLTap1_p;
    phaseCompState->yl_2 = outLTap2_p;
    phaseCompState->yr_1 = outRTap1_p;
    phaseCompState->yr_2 = outRTap2_p;

    return;
}

// stereo 4th order LR crossover
// this needs lots of optimization!
void StereoLowPass4thOrder (Float32 *in, Float32 *low, UInt32 frames, UInt32 SamplingRate, PreviousValues *section1State, PreviousValues *section2State)
{
    UInt32	i;
    Float32	inL, inR, outL1, outR1, outL, outR;

    // shared coefficients for second order sections
    Float32	b0, b1, b2, a1, a2;
    // taps for second order section 1
    Float32	inLTap1, inLTap2, inRTap1, inRTap2;
    Float32	outLTap1, outLTap2, outRTap1, outRTap2;
    // taps for second order section 2
    Float32	inLTap1_2, inLTap2_2, inRTap1_2, inRTap2_2;
    Float32	outLTap1_2, outLTap2_2, outRTap1_2, outRTap2_2;

    // copy to state local variables to avoid structure referencing during inner loop
    // section 1
    inLTap1 = section1State->xl_1;
    inLTap2 = section1State->xl_2;
    inRTap1 = section1State->xr_1;
    inRTap2 = section1State->xr_2;

    outLTap1 = section1State->yl_1;
    outLTap2 = section1State->yl_2;
    outRTap1 = section1State->yr_1;
    outRTap2 = section1State->yr_2;
    
    // section 2
    inLTap1_2 = section2State->xl_1;
    inLTap2_2 = section2State->xl_2;
    inRTap1_2 = section2State->xr_1;
    inRTap2_2 = section2State->xr_2;

    outLTap1_2 = section2State->yl_1;
    outLTap2_2 = section2State->yl_2;
    outRTap1_2 = section2State->yr_1;
    outRTap2_2 = section2State->yr_2;
 
    // set all coefficients
    if (Set4thOrderCoefficients (&b0, &b1, &b2, &a1, &a2, SamplingRate) == FALSE)
        return;

    // need to unroll this loop to get rid of stalls!
    for ( i = 0 ; i < frames ; i ++ )
    {
        inL = in[2*i];
        inR = in[2*i+1];
        
        // Low-pass filter first pass
        outL1 = b0*inL + b1*inLTap1 + b2*inLTap2 - a1*outLTap1 - a2*outLTap2;
        outR1 = b0*inR + b1*inRTap1 + b2*inRTap2 - a1*outRTap1 - a2*outRTap2;
 
        // update section 1 filter taps
        inLTap2 = inLTap1;
        inRTap2 = inRTap1;
        inLTap1 = inL;
        inRTap1 = inR;
        outLTap2 = outLTap1;
        outRTap2 = outRTap1;
        outLTap1 = outL1;
        outRTap1 = outR1;
        
        // Low-pass filter second pass
        outL = b0*outL1 + b1*inLTap1_2 + b2*inLTap2_2 - a1*outLTap1_2 - a2*outLTap2_2;
        outR = b0*outR1 + b1*inRTap1_2 + b2*inRTap2_2 - a1*outRTap1_2 - a2*outRTap2_2;
        
        // update section 2 filter taps
        inLTap2_2 = inLTap1_2;
        inRTap2_2 = inRTap1_2;
        inLTap1_2 = outL1;
        inRTap1_2 = outR1;
        outLTap2_2 = outLTap1_2;
        outRTap2_2 = outRTap1_2;
        outLTap1_2 = outL;
        outRTap1_2 = outR;

        // Storage
        low[2*i] = outL;
        low[2*i+1] = outR;
    }

    // update state structures
    
    // section 1 state
    section1State->xl_1 = inLTap1;
    section1State->xl_2 = inLTap2;
    section1State->xr_1 = inRTap1;
    section1State->xr_2 = inRTap2;

    section1State->yl_1 = outLTap1;
    section1State->yl_2 = outLTap2;
    section1State->yr_1 = outRTap1;
    section1State->yr_2 = outRTap2;
    
    // section 2 state
    section2State->xl_1 = inLTap1_2;
    section2State->xl_2 = inLTap2_2;
    section2State->xr_1 = inRTap1_2;
    section2State->xr_2 = inRTap2_2;

    section2State->yl_1 = outLTap1_2;
    section2State->yl_2 = outLTap2_2;
    section2State->yr_1 = outRTap1_2;
    section2State->yr_2 = outRTap2_2;

    return;
}

// new routines [2964790]
#pragma mark ------------------------ 
#pragma mark ••• Conversion Routines
#pragma mark ------------------------ 

#if	defined(__ppc__)

// this behaves incorrectly in Float32ToSwapInt24 if not declared volatile
#define __lwbrx( index, base )	({ register long result; __asm__ __volatile__("lwbrx %0, %1, %2" : "=r" (result) : "b%" (index), "r" (base) : "memory" ); result; } )

#define __lhbrx(index, base)	\
  ({ register signed short lhbrxResult; \
	 __asm__ ("lhbrx %0, %1, %2" : "=r" (lhbrxResult) : "b%" (index), "r" (base) : "memory"); \
	 /*return*/ lhbrxResult; } )
	// dsw: make signed to get sign-extension

#define __rlwimi( rA, rS, cnt, mb, me ) \
	({ __asm__ __volatile__( "rlwimi %0, %2, %3, %4, %5" : "=r" (rA) : "0" (rA), "r" (rS), "n" (cnt), "n" (mb), "n" (me) ); /*return*/ rA; })

#define __stwbrx( value, index, base ) \
	__asm__( "stwbrx %0, %1, %2" : : "r" (value), "b%" (index), "r" (base) : "memory" )

#define __rlwimi_volatile( rA, rS, cnt, mb, me ) \
	({ __asm__ __volatile__( "rlwimi %0, %2, %3, %4, %5" : "=r" (rA) : "0" (rA), "r" (rS), "n" (cnt), "n" (mb), "n" (me) ); /*return*/ rA; })

#define __stfiwx( value, offset, addr )			\
	asm( "stfiwx %0, %1, %2" : /*no result*/ : "f" (value), "b%" (offset), "r" (addr) : "memory" )

static inline double __fctiw( register double B )
{
	register double result;
	asm( "fctiw %0, %1" : "=f" (result) : "f" (B)  );
	return result;
}

void Int8ToFloat32( SInt8 *src, float *dest, unsigned int count )
{
	register float bias;
	register long exponentMask = ((0x97UL - 8) << 23) | 0x8000;	//FP exponent + bias for sign
	register long int0, int1, int2, int3;
	register float float0, float1, float2, float3;
	register unsigned long loopCount;
	union
	{
		float	f;
		long	i;
	}exponent;

	exponent.i = exponentMask;
	bias = exponent.f;

	src--;
	if( count >= 8 )
	{
		//Software Cycle 1
		int0 = (++src)[0];

		//Software Cycle 2
		int1 = (++src)[0];
		int0 += exponentMask;

		//Software Cycle 3
		int2 = (++src)[0];
		int1 += exponentMask;
		((long*) dest)[0] = int0;

		//Software Cycle 4
		int3 = (++src)[0];
		int2 += exponentMask;
		((long*) dest)[1] = int1;
		//delay one loop for the store to complete

		//Software Cycle 5
		int0 = (++src)[0];
		int3 += exponentMask;
		((long*) dest)[2] = int2;
		float0 = dest[0];

		//Software Cycle 6
		int1 = (++src)[0];
		int0 += exponentMask;
		((long*) dest)[3] = int3;
		float1 = dest[1];
		float0 -= bias;

		//Software Cycle 7
		int2 = (++src)[0];
		int1 += exponentMask;
		((long*) dest)[4] = int0;
		float2 = dest[2];
		float1 -= bias;

		dest--;
		//Software Cycle 8
		int3 = (++src)[0];
		int2 += exponentMask;
		((long*) dest)[6] = int1;
		float3 = dest[4];
		float2 -= bias;
		(++dest)[0] = float0;

		count -= 8;
		loopCount = count / 4;
		count &= 3;
		while( loopCount-- )
		{

			//Software Cycle A
			int0 = (++src)[0];
			int3 += exponentMask;
			((long*) dest)[6] = int2;
			float0 = dest[4];
			float3 -= bias;
			(++dest)[0] = float1;

			//Software Cycle B
			int1 = (++src)[0];
			int0 += exponentMask;
			((long*) dest)[6] = int3;
			float1 = dest[4];
			float0 -= bias;
			(++dest)[0] = float2;

			//Software Cycle C
			int2 = (++src)[0];
			int1 += exponentMask;
			((long*) dest)[6] = int0;
			float2 = dest[4];
			float1 -= bias;
			(++dest)[0] = float3;

			//Software Cycle D
			int3 = (++src)[0];
			int2 += exponentMask;
			((long*) dest)[6] = int1;
			float3 = dest[4];
			float2 -= bias;
			(++dest)[0] = float0;
		}

		//Software Cycle 7
		int3 += exponentMask;
		((long*) dest)[6] = int2;
		float0 = dest[4];
		float3 -= bias;
		(++dest)[0] = float1;

		//Software Cycle 6
		((long*) dest)[6] = int3;
		float1 = dest[4];
		float0 -= bias;
		(++dest)[0] = float2;

		//Software Cycle 5
		float2 = dest[4];
		float1 -= bias;
		(++dest)[0] = float3;

		//Software Cycle 4
		float3 = dest[4];
		float2 -= bias;
		(++dest)[0] = float0;

		//Software Cycle 3
		float3 -= bias;
		(++dest)[0] = float1;

		//Software Cycle 2
		(++dest)[0] = float2;

		//Software Cycle 1
		(++dest)[0] = float3;

		dest++;
	}


	while( count-- )
	{
		register long value = (++src)[0];
		value += exponentMask;
		((long*) dest)[0] = value;
		dest[0] -= bias;
		dest++;
	}
}

// bitDepth may be less than 16, e.g. for low-aligned 12 bit samples
void NativeInt16ToFloat32( signed short *src, float *dest, unsigned int count, int bitDepth )
{
	register float bias;
	register long exponentMask = ((0x97UL - bitDepth) << 23) | 0x8000;	//FP exponent + bias for sign
	register long int0, int1, int2, int3;
	register float float0, float1, float2, float3;
	register unsigned long loopCount;
	union
	{
		float	f;
		long	i;
	} exponent;

	exponent.i = exponentMask;
	bias = exponent.f;

	src--;
	if( count >= 8 )
	{
		//Software Cycle 1
		int0 = (++src)[0];

		//Software Cycle 2
		int1 = (++src)[0];
		int0 += exponentMask;

		//Software Cycle 3
		int2 = (++src)[0];
		int1 += exponentMask;
		((long*) dest)[0] = int0;

		//Software Cycle 4
		int3 = (++src)[0];
		int2 += exponentMask;
		((long*) dest)[1] = int1;
		//delay one loop for the store to complete

		//Software Cycle 5
		int0 = (++src)[0];
		int3 += exponentMask;
		((long*) dest)[2] = int2;
		float0 = dest[0];

		//Software Cycle 6
		int1 = (++src)[0];
		int0 += exponentMask;
		((long*) dest)[3] = int3;
		float1 = dest[1];
		float0 -= bias;

		//Software Cycle 7
		int2 = (++src)[0];
		int1 += exponentMask;
		((long*) dest)[4] = int0;
		float2 = dest[2];
		float1 -= bias;

		dest--;
		//Software Cycle 8
		int3 = (++src)[0];
		int2 += exponentMask;
		((long*) dest)[6] = int1;
		float3 = dest[4];
		float2 -= bias;
		(++dest)[0] = float0;

		count -= 8;
		loopCount = count / 4;
		count &= 3;
		while( loopCount-- )
		{

			//Software Cycle A
			int0 = (++src)[0];
			int3 += exponentMask;
			((long*) dest)[6] = int2;
			float0 = dest[4];
			float3 -= bias;
			(++dest)[0] = float1;

			//Software Cycle B
			int1 = (++src)[0];
			int0 += exponentMask;
			((long*) dest)[6] = int3;
			float1 = dest[4];
			float0 -= bias;
			(++dest)[0] = float2;

			//Software Cycle C
			int2 = (++src)[0];
			int1 += exponentMask;
			((long*) dest)[6] = int0;
			float2 = dest[4];
			float1 -= bias;
			(++dest)[0] = float3;

			//Software Cycle D
			int3 = (++src)[0];
			int2 += exponentMask;
			((long*) dest)[6] = int1;
			float3 = dest[4];
			float2 -= bias;
			(++dest)[0] = float0;
		}

		//Software Cycle 7
		int3 += exponentMask;
		((long*) dest)[6] = int2;
		float0 = dest[4];
		float3 -= bias;
		(++dest)[0] = float1;

		//Software Cycle 6
		((long*) dest)[6] = int3;
		float1 = dest[4];
		float0 -= bias;
		(++dest)[0] = float2;

		//Software Cycle 5
		float2 = dest[4];
		float1 -= bias;
		(++dest)[0] = float3;

		//Software Cycle 4
		float3 = dest[4];
		float2 -= bias;
		(++dest)[0] = float0;

		//Software Cycle 3
		float3 -= bias;
		(++dest)[0] = float1;

		//Software Cycle 2
		(++dest)[0] = float2;

		//Software Cycle 1
		(++dest)[0] = float3;

		dest++;
	}


	while( count-- )
	{
		register long value = (++src)[0];
		value += exponentMask;
		((long*) dest)[0] = value;
		dest[0] -= bias;
		dest++;
	}
}

void NativeInt24ToFloat32( long *src, float *dest, unsigned int count, int bitDepth )
{
	union
	{
		double			d[4];
		unsigned int	i[8];
	} transfer;
	register double			dBias;
	register unsigned int	loopCount, load0SignMask;
	register unsigned long	load0, load1, load2;
	register unsigned long	int0, int1, int2, int3;
	register double			d0, d1, d2, d3;
	register float		f0, f1, f2, f3;

	transfer.i[0] = transfer.i[2] = transfer.i[4] = transfer.i[6] = (0x434UL - bitDepth) << 20; //0x41C00000UL;
	transfer.i[1] = 0x00800000;
	int0 = int1 = int2 = int3 = 0;
	load0SignMask = 0x80000080UL;

	dBias = transfer.d[0];

	src--;
	dest--;

	if( count >= 8 )
	{
		count -= 8;
		loopCount = count / 4;
		count &= 3;

		//Virtual cycle 1
		load0 = (++src)[0];

		//Virtual cycle 2
		load1 = (++src)[0];
		load0 ^= load0SignMask;

		//Virtual cycle 3
		load2 = (++src)[0];
		load1 ^= 0x00008000UL;
		int0 = load0 >> 8;
		int1 = __rlwimi( int1, load0, 16, 8, 15);

		//Virtual cycle 4
		//No load3 -- already loaded last cycle
		load2 ^= 0x00800000UL;
		int1 = __rlwimi( int1, load1, 16, 16, 31);
		int2 = __rlwimi( int2, load1, 8, 8, 23 );
		transfer.i[1] = int0;

		//Virtual cycle 5
		load0 = (++src)[0];
		int2 = __rlwimi( int2, load2, 8, 24, 31 );
		int3 = load2 & 0x00FFFFFF;
		transfer.i[3] = int1;

		//Virtual cycle 6
		load1 = (++src)[0];
		load0 ^= load0SignMask;
		transfer.i[5] = int2;
		d0 = transfer.d[0];

		//Virtual cycle 7
		load2 = (++src)[0];
		load1 ^= 0x00008000UL;
		int0 = load0 >> 8;
		int1 = __rlwimi( int1, load0, 16, 8, 15 );
		transfer.i[7] = int3;
		d1 = transfer.d[1];
		d0 -= dBias;

		//Virtual cycle 8
		//No load3 -- already loaded last cycle
		load2 ^= 0x00800000UL;
		int1 = __rlwimi( int1, load1, 16, 16, 31);
		int2 = __rlwimi( int2, load1, 8, 8, 23 );
		transfer.i[1] = int0;
		d2 = transfer.d[2];
		d1 -= dBias;
		f0 = d0;

		while( loopCount-- )
		{
			//Virtual cycle A
			load0 = (++src)[0];
			int2 = __rlwimi( int2, load2, 8, 24, 31 );
			int3 = load2 & 0x00FFFFFF;
			transfer.i[3] = int1;
			d3 = transfer.d[3];
			d2 -= dBias;
			f1 = d1;
			(++dest)[0] = f0;

			//Virtual cycle B
			load1 = (++src)[0];
			load0 ^= load0SignMask;
			transfer.i[5] = int2;
			d0 = transfer.d[0];
			d3 -= dBias;
			f2 = d2;
			(++dest)[0] = f1;

			//Virtual cycle C
			load2 = (++src)[0];
			load1 ^= 0x00008000UL;
			int0 = load0 >> 8;
			int1 = __rlwimi( int1, load0, 16, 8, 15 );
			transfer.i[7] = int3;
			d1 = transfer.d[1];
			d0 -= dBias;
			f3 = d3;
			(++dest)[0] = f2;

			//Virtual cycle D
			load2 ^= 0x00800000UL;
			int1 = __rlwimi( int1, load1, 16, 16, 31);
			int2 = __rlwimi( int2, load1, 8, 8, 23 );
			transfer.i[1] = int0;
			d2 = transfer.d[2];
			d1 -= dBias;
			f0 = d0;
			(++dest)[0] = f3;
		}

		//Virtual cycle 8
		int2 = __rlwimi( int2, load2, 8, 24, 31 );
		int3 = load2 & 0x00FFFFFF;
		transfer.i[3] = int1;
		d3 = transfer.d[3];
		d2 -= dBias;
		f1 = d1;
		(++dest)[0] = f0;

		//Virtual cycle 7
		transfer.i[5] = int2;
		d0 = transfer.d[0];
		d3 -= dBias;
		f2 = d2;
		(++dest)[0] = f1;

		//Virtual cycle 6
		transfer.i[7] = int3;
		d1 = transfer.d[1];
		d0 -= dBias;
		f3 = d3;
		(++dest)[0] = f2;

		//Virtual cycle 5
		d2 = transfer.d[2];
		d1 -= dBias;
		f0 = d0;
		(++dest)[0] = f3;

		//Virtual cycle 4
		d3 = transfer.d[3];
		d2 -= dBias;
		f1 = d1;
		(++dest)[0] = f0;

		//Virtual cycle 3
		d3 -= dBias;
		f2 = d2;
		(++dest)[0] = f1;

		//Virtual cycle 2
		f3 = d3;
		(++dest)[0] = f2;

		//Virtual cycle 1
		(++dest)[0] = f3;
	}

	src = (long*) ((char*) src + 1 );
	while( count-- )
	{
		int0 = ((unsigned char*)(src = (long*)( (char*) src + 3 )))[0];
		int1 = ((unsigned short*)( (char*) src + 1 ))[0];
		int0 ^= 0x00000080UL;
		int1 = __rlwimi( int1, int0, 16, 8, 15 );
		transfer.i[1] = int1;

		d0 = transfer.d[0];
		d0 -= dBias;
		f0 = d0;
		(++dest)[0] = f0;
   }
}

// bitDepth may be less than 32, e.g. for 24 bits low-aligned in 32-bit words
void NativeInt32ToFloat32( long *src, float *dest, unsigned int count, int bitDepth )
{
	union
	{
		double			d[4];
		unsigned int	i[8];
	}transfer;
	register double 		dBias;
	register unsigned int 	loopCount;
	register long			int0, int1, int2, int3;
	register double			d0, d1, d2, d3;
	register float			f0, f1, f2, f3;

	transfer.i[0] = transfer.i[2] = transfer.i[4] = transfer.i[6] = (0x434UL - bitDepth) << 20;
		//0x41400000UL;
	transfer.i[1] = 0x80000000;

	dBias = transfer.d[0];

	src--;
	dest--;

	if( count >= 8 )
	{
		count -= 8;
		loopCount = count / 4;
		count &= 3;

		//Virtual cycle 1
		int0 = (++src)[0];

		//Virtual cycle 2
		int1 = (++src)[0];
		int0 ^= 0x80000000UL;

		//Virtual cycle 3
		int2 = (++src)[0];
		int1 ^= 0x80000000UL;
		transfer.i[1] = int0;

		//Virtual cycle 4
		int3 = (++src)[0];
		int2 ^= 0x80000000UL;
		transfer.i[3] = int1;

		//Virtual cycle 5
		int0 = (++src)[0];
		int3 ^= 0x80000000UL;
		transfer.i[5] = int2;

		//Virtual cycle 6
		int1 = (++src)[0];
		int0 ^= 0x80000000UL;
		transfer.i[7] = int3;
		d0 = transfer.d[0];

		//Virtual cycle 7
		int2 = (++src)[0];
		int1 ^= 0x80000000UL;
		transfer.i[1] = int0;
		d1 = transfer.d[1];
		d0 -= dBias;

		//Virtual cycle 8
		int3 = (++src)[0];
		int2 ^= 0x80000000UL;
		transfer.i[3] = int1;
		d2 = transfer.d[2];
		d1 -= dBias;
		f0 = d0;

		while( loopCount-- )
		{
			//Virtual cycle A
			int0 = (++src)[0];
			int3 ^= 0x80000000UL;
			transfer.i[5] = int2;
			d3 = transfer.d[3];
			d2 -= dBias;
			f1 = d1;
			(++dest)[0] = f0;

			//Virtual cycle B
			int1 = (++src)[0];
			int0 ^= 0x80000000UL;
			transfer.i[7] = int3;
			d0 = transfer.d[0];
			d3 -= dBias;
			f2 = d2;
			(++dest)[0] = f1;

			//Virtual cycle C
			int2 = (++src)[0];
			int1 ^= 0x80000000UL;
			transfer.i[1] = int0;
			d1 = transfer.d[1];
			d0 -= dBias;
			f3 = d3;
			(++dest)[0] = f2;

			//Virtual cycle D
			int3 = (++src)[0];
			int2 ^= 0x80000000UL;
			transfer.i[3] = int1;
			d2 = transfer.d[2];
			d1 -= dBias;
			f0 = d0;
			(++dest)[0] = f3;
		}

		//Virtual cycle 8
		int3 ^= 0x80000000UL;
		transfer.i[5] = int2;
		d3 = transfer.d[3];
		d2 -= dBias;
		f1 = d1;
		(++dest)[0] = f0;

		//Virtual cycle 7
		transfer.i[7] = int3;
		d0 = transfer.d[0];
		d3 -= dBias;
		f2 = d2;
		(++dest)[0] = f1;

		//Virtual cycle 6
		d1 = transfer.d[1];
		d0 -= dBias;
		f3 = d3;
		(++dest)[0] = f2;

		//Virtual cycle 5
		d2 = transfer.d[2];
		d1 -= dBias;
		f0 = d0;
		(++dest)[0] = f3;

		//Virtual cycle 4
		d3 = transfer.d[3];
		d2 -= dBias;
		f1 = d1;
		(++dest)[0] = f0;

		//Virtual cycle 3
		d3 -= dBias;
		f2 = d2;
		(++dest)[0] = f1;

		//Virtual cycle 2
		f3 = d3;
		(++dest)[0] = f2;

		//Virtual cycle 1
		(++dest)[0] = f3;
	}


	while( count-- )
	{
		int0 = (++src)[0];
		int0 ^= 0x80000000UL;
		transfer.i[1] = int0;

		d0 = transfer.d[0];
		d0 -= dBias;
		f0 = d0;
		(++dest)[0] = f0;
	}
}

// bitDepth may be less than 16, e.g. for low-aligned 12 bit samples
void NativeInt16ToFloat32Gain( signed short *src, float *dest, unsigned int count, int bitDepth, float* inGainLPtr, float* inGainRPtr )
{
	register float bias, gainL, gainR;
	register long exponentMask = ((0x97UL - bitDepth) << 23) | 0x8000;	//FP exponent + bias for sign
	register long int0, int1, int2, int3;
	register float float0, float1, float2, float3;
	register unsigned long loopCount;
	union
	{
		float	f;
		long	i;
	} exponent;

	exponent.i = exponentMask;
	bias = exponent.f;

	if (inGainLPtr) {
		gainL = *inGainLPtr;
	} else {
		gainL = 1.0f;
	}
	if (inGainRPtr) {
		gainR = *inGainRPtr;	
	} else {
		gainR = 1.0f;
	}
	
	src--;
	if( count >= 8 )
	{
		//Software Cycle 1
		int0 = (++src)[0];

		//Software Cycle 2
		int1 = (++src)[0];
		int0 += exponentMask;

		//Software Cycle 3
		int2 = (++src)[0];
		int1 += exponentMask;
		((long*) dest)[0] = int0;

		//Software Cycle 4
		int3 = (++src)[0];
		int2 += exponentMask;
		((long*) dest)[1] = int1;
		//delay one loop for the store to complete

		//Software Cycle 5
		int0 = (++src)[0];
		int3 += exponentMask;
		((long*) dest)[2] = int2;
		float0 = dest[0];

		//Software Cycle 6
		int1 = (++src)[0];
		int0 += exponentMask;
		((long*) dest)[3] = int3;
		float1 = dest[1];
		float0 -= bias;

		//Software Cycle 7
		int2 = (++src)[0];
		int1 += exponentMask;
		((long*) dest)[4] = int0;
		float0 *= gainL;
		float2 = dest[2];
		float1 -= bias;

		dest--;
		//Software Cycle 8
		int3 = (++src)[0];
		int2 += exponentMask;
		((long*) dest)[6] = int1;
		float1 *= gainR;
		float3 = dest[4];
		float2 -= bias;
		(++dest)[0] = float0;

		count -= 8;
		loopCount = count / 4;
		count &= 3;
		while( loopCount-- )
		{

			//Software Cycle A
			int0 = (++src)[0];
			int3 += exponentMask;
			((long*) dest)[6] = int2;
			float2 *= gainL;
			float0 = dest[4];
			float3 -= bias;
			(++dest)[0] = float1;

			//Software Cycle B
			int1 = (++src)[0];
			int0 += exponentMask;
			((long*) dest)[6] = int3;
			float3 *= gainR;
			float1 = dest[4];
			float0 -= bias;
			(++dest)[0] = float2;

			//Software Cycle C
			int2 = (++src)[0];
			int1 += exponentMask;
			((long*) dest)[6] = int0;
			float0 *= gainL;
			float2 = dest[4];
			float1 -= bias;
			(++dest)[0] = float3;

			//Software Cycle D
			int3 = (++src)[0];
			int2 += exponentMask;
			((long*) dest)[6] = int1;
			float1 *= gainR;
			float3 = dest[4];
			float2 -= bias;
			(++dest)[0] = float0;
		}

		//Software Cycle 7
		int3 += exponentMask;
		((long*) dest)[6] = int2;
		float2 *= gainL;
		float0 = dest[4];
		float3 -= bias;
		(++dest)[0] = float1;

		//Software Cycle 6
		((long*) dest)[6] = int3;
		float3 *= gainR;
		float1 = dest[4];
		float0 -= bias;
		(++dest)[0] = float2;

		//Software Cycle 5
		float0 *= gainL;
		float2 = dest[4];
		float1 -= bias;
		(++dest)[0] = float3;

		//Software Cycle 4
		float1 *= gainR;
		float3 = dest[4];
		float2 -= bias;
		(++dest)[0] = float0;

		//Software Cycle 3
		float2 *= gainL;
		float3 -= bias;
		(++dest)[0] = float1;

		//Software Cycle 2
		float3 *= gainR;
		(++dest)[0] = float2;

		//Software Cycle 1
		(++dest)[0] = float3;

		dest++;
	}

	loopCount = count/2;
	while( loopCount-- )
	{
		register long value = (++src)[0];
		value += exponentMask;
		((long*) dest)[0] = value;
		value = (++src)[0];
		dest[0] -= bias;
		dest[0] *= gainL;
		value += exponentMask;
		dest++;
		((long*) dest)[0] = value;
		dest[0] -= bias;
		dest[0] *= gainR;
		dest++;
	}
	if (count % 2) {
		register long value = (++src)[0];
		value += exponentMask;
		((long*) dest)[0] = value;
		dest[0] -= bias;
		dest[0] *= gainL;
	}
}

// bitDepth may be less than 16, e.g. for low-aligned 12 bit samples
void NativeInt16ToFloat32CopyRightToLeft( signed short *src, float *dest, unsigned int count, int bitDepth )
{
	register float bias;
	register long exponentMask = ((0x97UL - bitDepth) << 23) | 0x8000;	//FP exponent + bias for sign
	register long int0, int1, int2, int3;
	register float float0, float1, float2, float3;
	register unsigned long loopCount;
	union
	{
		float	f;
		long	i;
	} exponent;

	exponent.i = exponentMask;
	bias = exponent.f;

	if( count >= 8 )
	{
		//Software Cycle 1
		int0 = (++src)[0];		// right 1

		//Software Cycle 2
		int1 = (src++)[0];		// reuse right 1, skip left 1
		int0 += exponentMask;

		//Software Cycle 3
		int2 = (++src)[0];		// right 2
		int1 += exponentMask;
		((long*) dest)[0] = int0;

		//Software Cycle 4
		int3 = (src++)[0];		// reuse right 2, skip left 2
		int2 += exponentMask;
		((long*) dest)[1] = int1;
		//delay one loop for the store to complete

		//Software Cycle 5
		int0 = (++src)[0];		// right 3
		int3 += exponentMask;
		((long*) dest)[2] = int2;
		float0 = dest[0];

		//Software Cycle 6
		int1 = (src++)[0];		// reuse right 3, skip left 3
		int0 += exponentMask;
		((long*) dest)[3] = int3;
		float1 = dest[1];
		float0 -= bias;

		//Software Cycle 7
		int2 = (++src)[0];		// right 4
		int1 += exponentMask;
		((long*) dest)[4] = int0;
		float2 = dest[2];
		float1 -= bias;

		dest--;
		//Software Cycle 8
		int3 = (src++)[0];		// reuse left 4, skip right 4
		int2 += exponentMask;
		((long*) dest)[6] = int1;
		float3 = dest[4];
		float2 -= bias;
		(++dest)[0] = float0;

		count -= 8;
		loopCount = count / 4;
		count &= 3;
		while( loopCount-- )
		{

			//Software Cycle A
			int0 = (++src)[0];
			int3 += exponentMask;
			((long*) dest)[6] = int2;
			float0 = dest[4];
			float3 -= bias;
			(++dest)[0] = float1;

			//Software Cycle B
			int1 = (src++)[0];
			int0 += exponentMask;
			((long*) dest)[6] = int3;
			float1 = dest[4];
			float0 -= bias;
			(++dest)[0] = float2;

			//Software Cycle C
			int2 = (++src)[0];
			int1 += exponentMask;
			((long*) dest)[6] = int0;
			float2 = dest[4];
			float1 -= bias;
			(++dest)[0] = float3;

			//Software Cycle D
			int3 = (src++)[0];
			int2 += exponentMask;
			((long*) dest)[6] = int1;
			float3 = dest[4];
			float2 -= bias;
			(++dest)[0] = float0;
		}

		//Software Cycle 7
		int3 += exponentMask;
		((long*) dest)[6] = int2;
		float0 = dest[4];
		float3 -= bias;
		(++dest)[0] = float1;

		//Software Cycle 6
		((long*) dest)[6] = int3;
		float1 = dest[4];
		float0 -= bias;
		(++dest)[0] = float2;

		//Software Cycle 5
		float2 = dest[4];
		float1 -= bias;
		(++dest)[0] = float3;

		//Software Cycle 4
		float3 = dest[4];
		float2 -= bias;
		(++dest)[0] = float0;

		//Software Cycle 3
		float3 -= bias;
		(++dest)[0] = float1;

		//Software Cycle 2
		(++dest)[0] = float2;

		//Software Cycle 1
		(++dest)[0] = float3;

		dest++;
	}

	loopCount = count/2;
	while( loopCount-- )
	{
		register long value = (++src)[0];
		value += exponentMask;
		((long*) dest)[0] = value;
		dest[0] -= bias;
		dest++;
		++src;
		((long*) dest)[0] = value;
		dest[0] -= bias;
		dest++;
	}
	if (count % 2) {
		register long value = (++src)[0];
		value += exponentMask;
		((long*) dest)[0] = value;
		dest[0] -= bias;
	}
}

// bitDepth may be less than 32, e.g. for 24 bits low-aligned in 32-bit words
void NativeInt32ToFloat32Gain( signed long *src, float *dest, unsigned int count, int bitDepth, float* inGainLPtr, float* inGainRPtr )
{
	union
	{
		double			d[4];
		unsigned int	i[8];
	} transfer;
	register double 		dBias, gainL, gainR;
	register unsigned int 	loopCount;
	register long			int0, int1, int2, int3;
	register double			d0, d1, d2, d3;
	register float			f0, f1, f2, f3;

	transfer.i[0] = transfer.i[2] = transfer.i[4] = transfer.i[6] = (0x434UL - bitDepth) << 20;
		//0x41400000UL;
	transfer.i[1] = 0x80000000;

	dBias = transfer.d[0];

	if (inGainLPtr) {
		gainL = *inGainLPtr;
	} else {
		gainL = 1.0f;
	}
	if (inGainRPtr) {
		gainR = *inGainRPtr;	
	} else {
		gainR = 1.0f;
	}

	src--;
	dest--;

	if( count >= 8 )
	{
		count -= 8;
		loopCount = count / 4;
		count &= 3;

		//Virtual cycle 1
		int0 = (++src)[0];

		//Virtual cycle 2
		int1 = (++src)[0];
		int0 ^= 0x80000000UL;

		//Virtual cycle 3
		int2 = (++src)[0];
		int1 ^= 0x80000000UL;
		transfer.i[1] = int0;

		//Virtual cycle 4
		int3 = (++src)[0];
		int2 ^= 0x80000000UL;
		transfer.i[3] = int1;

		//Virtual cycle 5
		int0 = (++src)[0];
		int3 ^= 0x80000000UL;
		transfer.i[5] = int2;

		//Virtual cycle 6
		int1 = (++src)[0];
		int0 ^= 0x80000000UL;
		transfer.i[7] = int3;
		d0 = transfer.d[0];

		//Virtual cycle 7
		int2 = (++src)[0];
		int1 ^= 0x80000000UL;
		transfer.i[1] = int0;
		d1 = transfer.d[1];
		d0 -= dBias;

		//Virtual cycle 8
		int3 = (++src)[0];
		int2 ^= 0x80000000UL;
		transfer.i[3] = int1;
		d2 = transfer.d[2];
		d1 -= dBias;
		f0 = d0*gainL;

		while( loopCount-- )
		{
			//Virtual cycle A
			int0 = (++src)[0];
			int3 ^= 0x80000000UL;
			transfer.i[5] = int2;
			d3 = transfer.d[3];
			d2 -= dBias;
			f1 = d1*gainR;
			(++dest)[0] = f0;

			//Virtual cycle B
			int1 = (++src)[0];
			int0 ^= 0x80000000UL;
			transfer.i[7] = int3;
			d0 = transfer.d[0];
			d3 -= dBias;
			f2 = d2*gainL;
			(++dest)[0] = f1;

			//Virtual cycle C
			int2 = (++src)[0];
			int1 ^= 0x80000000UL;
			transfer.i[1] = int0;
			d1 = transfer.d[1];
			d0 -= dBias;
			f3 = d3*gainR;
			(++dest)[0] = f2;

			//Virtual cycle D
			int3 = (++src)[0];
			int2 ^= 0x80000000UL;
			transfer.i[3] = int1;
			d2 = transfer.d[2];
			d1 -= dBias;
			f0 = d0*gainL;
			(++dest)[0] = f3;
		}

		//Virtual cycle 8
		int3 ^= 0x80000000UL;
		transfer.i[5] = int2;
		d3 = transfer.d[3];
		d2 -= dBias;
		f1 = d1*gainR;
		(++dest)[0] = f0;

		//Virtual cycle 7
		transfer.i[7] = int3;
		d0 = transfer.d[0];
		d3 -= dBias;
		f2 = d2*gainL;
		(++dest)[0] = f1;

		//Virtual cycle 6
		d1 = transfer.d[1];
		d0 -= dBias;
		f3 = d3*gainR;
		(++dest)[0] = f2;

		//Virtual cycle 5
		d2 = transfer.d[2];
		d1 -= dBias;
		f0 = d0*gainL;
		(++dest)[0] = f3;

		//Virtual cycle 4
		d3 = transfer.d[3];
		d2 -= dBias;
		f1 = d1*gainR;
		(++dest)[0] = f0;

		//Virtual cycle 3
		d3 -= dBias;
		f2 = d2*gainL;
		(++dest)[0] = f1;

		//Virtual cycle 2
		f3 = d3*gainR;
		(++dest)[0] = f2;

		//Virtual cycle 1
		(++dest)[0] = f3;
	}


	loopCount = count/2;
	while( loopCount-- )
	{
		int0 = (++src)[0];
		int0 ^= 0x80000000UL;
		transfer.i[1] = int0;

		d0 = transfer.d[0];
		d0 -= dBias;
		f0 = d0*gainL;
		(++dest)[0] = f0;

		int0 = (++src)[0];
		int0 ^= 0x80000000UL;
		transfer.i[1] = int0;

		d0 = transfer.d[0];
		d0 -= dBias;
		f0 = d0*gainR;
		(++dest)[0] = f0;
	}
	if (count % 2) {
		int0 = (++src)[0];
		int0 ^= 0x80000000UL;
		transfer.i[1] = int0;

		d0 = transfer.d[0];
		d0 -= dBias;
		f0 = d0*gainL;
		(++dest)[0] = f0;
	}
}


// bitDepth may be less than 16, e.g. for low-aligned 12 bit samples
void SwapInt16ToFloat32( signed short *src, float *dest, unsigned int count, int bitDepth )
{
	register float bias;
	register long exponentMask = ((0x97UL - bitDepth) << 23) | 0x8000;	//FP exponent + bias for sign
	register long int0, int1, int2, int3;
	register float float0, float1, float2, float3;
	register unsigned long loopCount;
	union
	{
		float	f;
		long	i;
	}exponent;

	exponent.i = exponentMask;
	bias = exponent.f;

	src--;
	if( count >= 8 )
	{
		//Software Cycle 1
		int0 = __lhbrx(0, ++src);

		//Software Cycle 2
		int1 = __lhbrx(0, ++src);
		int0 += exponentMask;

		//Software Cycle 3
		int2 = __lhbrx(0, ++src);
		int1 += exponentMask;
		((long*) dest)[0] = int0;

		//Software Cycle 4
		int3 = __lhbrx(0, ++src);
		int2 += exponentMask;
		((long*) dest)[1] = int1;
		//delay one loop for the store to complete

		//Software Cycle 5
		int0 = __lhbrx(0, ++src);
		int3 += exponentMask;
		((long*) dest)[2] = int2;
		float0 = dest[0];

		//Software Cycle 6
		int1 = __lhbrx(0, ++src);
		int0 += exponentMask;
		((long*) dest)[3] = int3;
		float1 = dest[1];
		float0 -= bias;

		//Software Cycle 7
		int2 = __lhbrx(0, ++src);
		int1 += exponentMask;
		((long*) dest)[4] = int0;
		float2 = dest[2];
		float1 -= bias;

		dest--;
		//Software Cycle 8
		int3 = __lhbrx(0, ++src);
		int2 += exponentMask;
		((long*) dest)[6] = int1;
		float3 = dest[4];
		float2 -= bias;
		(++dest)[0] = float0;

		count -= 8;
		loopCount = count / 4;
		count &= 3;
		while( loopCount-- )
		{

			//Software Cycle A
			int0 = __lhbrx(0, ++src);
			int3 += exponentMask;
			((long*) dest)[6] = int2;
			float0 = dest[4];
			float3 -= bias;
			(++dest)[0] = float1;

			//Software Cycle B
			int1 = __lhbrx(0, ++src);
			int0 += exponentMask;
			((long*) dest)[6] = int3;
			float1 = dest[4];
			float0 -= bias;
			(++dest)[0] = float2;

			//Software Cycle C
			int2 = __lhbrx(0, ++src);
			int1 += exponentMask;
			((long*) dest)[6] = int0;
			float2 = dest[4];
			float1 -= bias;
			(++dest)[0] = float3;

			//Software Cycle D
			int3 = __lhbrx(0, ++src);
			int2 += exponentMask;
			((long*) dest)[6] = int1;
			float3 = dest[4];
			float2 -= bias;
			(++dest)[0] = float0;
		}

		//Software Cycle 7
		int3 += exponentMask;
		((long*) dest)[6] = int2;
		float0 = dest[4];
		float3 -= bias;
		(++dest)[0] = float1;

		//Software Cycle 6
		((long*) dest)[6] = int3;
		float1 = dest[4];
		float0 -= bias;
		(++dest)[0] = float2;

		//Software Cycle 5
		float2 = dest[4];
		float1 -= bias;
		(++dest)[0] = float3;

		//Software Cycle 4
		float3 = dest[4];
		float2 -= bias;
		(++dest)[0] = float0;

		//Software Cycle 3
		float3 -= bias;
		(++dest)[0] = float1;

		//Software Cycle 2
		(++dest)[0] = float2;

		//Software Cycle 1
		(++dest)[0] = float3;

		dest++;
	}


	while( count-- )
	{
		register long value = __lhbrx(0, ++src);
		value += exponentMask;
		((long*) dest)[0] = value;
		dest[0] -= bias;
		dest++;
	}
}

// CAUTION: bitDepth is ignored
void SwapInt24ToFloat32( long *src, float *dest, unsigned int count, int bitDepth )
{
	union
	{
		double		d[4];
		unsigned int	i[8];
	}transfer;
	register double			dBias;
	register unsigned int	loopCount, load2SignMask;
	register unsigned long	load0, load1, load2;
	register unsigned long	int0, int1, int2, int3;
	register double			d0, d1, d2, d3;
	register float		f0, f1, f2, f3;

	transfer.i[0] = transfer.i[2] = transfer.i[4] = transfer.i[6] = 0x41400000UL;
	transfer.i[1] = 0x80000000;
	int0 = int1 = int2 = int3 = 0;
	load2SignMask = 0x80000080UL;

	dBias = transfer.d[0];

	src--;
	dest--;

	if( count >= 8 )
	{
		count -= 8;
		loopCount = count / 4;
		count &= 3;

		//Virtual cycle 1
		load0 = (++src)[0];

		//Virtual cycle 2
		load1 = (++src)[0];
		load0 ^= 0x00008000;

		//Virtual cycle 3
		load2 = (++src)[0];
		load1 ^= 0x00800000UL;
		int0 = load0 >> 8;
		int1 = __rlwimi( int1, load0, 16, 8, 15);

		//Virtual cycle 4
		//No load3 -- already loaded last cycle
		load2 ^= load2SignMask;
		int1 = __rlwimi( int1, load1, 16, 16, 31);
		int2 = __rlwimi( int2, load1, 8, 8, 23 );
		__stwbrx( int0, 0, &transfer.i[1]);

		//Virtual cycle 5
		load0 = (++src)[0];
		int2 = __rlwimi( int2, load2, 8, 24, 31 );
		int3 = load2 & 0x00FFFFFF;
		__stwbrx( int1, 0, &transfer.i[3]);

		//Virtual cycle 6
		load1 = (++src)[0];
		load0 ^= 0x00008000;
		__stwbrx( int2, 0, &transfer.i[5]);
		d0 = transfer.d[0];

		//Virtual cycle 7
		load2 = (++src)[0];
		load1 ^= 0x00800000UL;
		int0 = load0 >> 8;
		int1 = __rlwimi( int1, load0, 16, 8, 15 );
		__stwbrx( int3, 0, &transfer.i[7]);
		d1 = transfer.d[1];
		d0 -= dBias;

		//Virtual cycle 8
		//No load3 -- already loaded last cycle
		load2 ^= load2SignMask;
		int1 = __rlwimi( int1, load1, 16, 16, 31);
		int2 = __rlwimi( int2, load1, 8, 8, 23 );
		__stwbrx( int0, 0, &transfer.i[1]);
		d2 = transfer.d[2];
		d1 -= dBias;
		f0 = d0;

		while( loopCount-- )
		{
			//Virtual cycle A
			load0 = (++src)[0];
			int2 = __rlwimi( int2, load2, 8, 24, 31 );
			int3 = load2 & 0x00FFFFFF;
			__stwbrx( int1, 0, &transfer.i[3]);
			d3 = transfer.d[3];
			d2 -= dBias;
			f1 = d1;
			(++dest)[0] = f0;

			//Virtual cycle B
			load1 = (++src)[0];
			load0 ^= 0x00008000;
			__stwbrx( int2, 0, &transfer.i[5]);
			d0 = transfer.d[0];
			d3 -= dBias;
			f2 = d2;
			(++dest)[0] = f1;

			//Virtual cycle C
			load2 = (++src)[0];
			load1 ^= 0x00800000UL;
			int0 = load0 >> 8;
			int1 = __rlwimi( int1, load0, 16, 8, 15 );
			__stwbrx( int3, 0, &transfer.i[7]);
			d1 = transfer.d[1];
			d0 -= dBias;
			f3 = d3;
			(++dest)[0] = f2;

			//Virtual cycle D
			load2 ^= load2SignMask;
			int1 = __rlwimi( int1, load1, 16, 16, 31);
			int2 = __rlwimi( int2, load1, 8, 8, 23 );
			__stwbrx( int0, 0, &transfer.i[1]);
			d2 = transfer.d[2];
			d1 -= dBias;
			f0 = d0;
			(++dest)[0] = f3;
		}

		//Virtual cycle 8
		int2 = __rlwimi( int2, load2, 8, 24, 31 );
		int3 = load2 & 0x00FFFFFF;
		__stwbrx( int1, 0, &transfer.i[3]);
		d3 = transfer.d[3];
		d2 -= dBias;
		f1 = d1;
		(++dest)[0] = f0;

		//Virtual cycle 7
		__stwbrx( int2, 0, &transfer.i[5]);
		d0 = transfer.d[0];
		d3 -= dBias;
		f2 = d2;
		(++dest)[0] = f1;

		//Virtual cycle 6
		__stwbrx( int3, 0, &transfer.i[7]);
		d1 = transfer.d[1];
		d0 -= dBias;
		f3 = d3;
		(++dest)[0] = f2;

		//Virtual cycle 5
		d2 = transfer.d[2];
		d1 -= dBias;
		f0 = d0;
		(++dest)[0] = f3;

		//Virtual cycle 4
		d3 = transfer.d[3];
		d2 -= dBias;
		f1 = d1;
		(++dest)[0] = f0;

		//Virtual cycle 3
		d3 -= dBias;
		f2 = d2;
		(++dest)[0] = f1;

		//Virtual cycle 2
		f3 = d3;
		(++dest)[0] = f2;

		//Virtual cycle 1
		(++dest)[0] = f3;
	}

	if( count > 0 )
	{

		int1 = ((unsigned char*) src)[6];
		int0 = ((unsigned short*)(++src))[0];
		int1 ^= 0x80;
		int1 = __rlwimi( int1, int0, 8, 8, 23 );
		__stwbrx( int1, 0, &transfer.i[1]);
		d0 = transfer.d[0];
		d0 -= dBias;
		f0 = d0;
		(++dest)[0] = f0;

		src = (long*) ((char*)src - 1 );
		while( --count )
		{
			int0 = (src = (long*)( (char*) src + 3 ))[0];
			int0 ^= 0x80UL;
			int0 &= 0x00FFFFFFUL;
			__stwbrx( int0, 0, &transfer.i[1]);

			d0 = transfer.d[0];
			d0 -= dBias;
			f0 = d0;
			(++dest)[0] = f0;
		}
	}
}

// bitDepth may be less than 32, e.g. for 24 bits low-aligned in 32-bit words
void SwapInt32ToFloat32( long *src, float *dest, unsigned int count, int bitDepth )
{
	union
	{
		double		d[4];
		unsigned int	i[8];
	}transfer;
	register double dBias;
	register unsigned int loopCount;
	register long	int0, int1, int2, int3;
	register double		d0, d1, d2, d3;
	register float	f0, f1, f2, f3;

	transfer.i[0] = transfer.i[2] = transfer.i[4] = transfer.i[6] = (0x434UL - bitDepth) << 20;
		//0x41400000UL;
	transfer.i[1] = 0x80000000;

	dBias = transfer.d[0];

	src--;
	dest--;

	if( count >= 8 )
	{
		count -= 8;
		loopCount = count / 4;
		count &= 3;

		//Virtual cycle 1
		int0 = __lwbrx( 0, ++src);

		//Virtual cycle 2
		int1 = __lwbrx( 0, ++src);
		int0 ^= 0x80000000UL;

		//Virtual cycle 3
		int2 = __lwbrx( 0, ++src);
		int1 ^= 0x80000000UL;
		transfer.i[1] = int0;

		//Virtual cycle 4
		int3 = __lwbrx( 0, ++src);
		int2 ^= 0x80000000UL;
		transfer.i[3] = int1;

		//Virtual cycle 5
		int0 = __lwbrx( 0, ++src);
		int3 ^= 0x80000000UL;
		transfer.i[5] = int2;

		//Virtual cycle 6
		int1 = __lwbrx( 0, ++src);
		int0 ^= 0x80000000UL;
		transfer.i[7] = int3;
		d0 = transfer.d[0];

		//Virtual cycle 7
		int2 = __lwbrx( 0, ++src);
		int1 ^= 0x80000000UL;
		transfer.i[1] = int0;
		d1 = transfer.d[1];
		d0 -= dBias;

		//Virtual cycle 8
		int3 = __lwbrx( 0, ++src);
		int2 ^= 0x80000000UL;
		transfer.i[3] = int1;
		d2 = transfer.d[2];
		d1 -= dBias;
		f0 = d0;

		while( loopCount-- )
		{
			//Virtual cycle A
			int0 = __lwbrx( 0, ++src);
			int3 ^= 0x80000000UL;
			transfer.i[5] = int2;
			d3 = transfer.d[3];
			d2 -= dBias;
			f1 = d1;
			(++dest)[0] = f0;

			//Virtual cycle B
			int1 = __lwbrx( 0, ++src);
			int0 ^= 0x80000000UL;
			transfer.i[7] = int3;
			d0 = transfer.d[0];
			d3 -= dBias;
			f2 = d2;
			(++dest)[0] = f1;

			//Virtual cycle C
			int2 = __lwbrx( 0, ++src);
			int1 ^= 0x80000000UL;
			transfer.i[1] = int0;
			d1 = transfer.d[1];
			d0 -= dBias;
			f3 = d3;
			(++dest)[0] = f2;

			//Virtual cycle D
			int3 = __lwbrx( 0, ++src);
			int2 ^= 0x80000000UL;
			transfer.i[3] = int1;
			d2 = transfer.d[2];
			d1 -= dBias;
			f0 = d0;
			(++dest)[0] = f3;
		}

		//Virtual cycle 8
		int3 ^= 0x80000000UL;
		transfer.i[5] = int2;
		d3 = transfer.d[3];
		d2 -= dBias;
		f1 = d1;
		(++dest)[0] = f0;

		//Virtual cycle 7
		transfer.i[7] = int3;
		d0 = transfer.d[0];
		d3 -= dBias;
		f2 = d2;
		(++dest)[0] = f1;

		//Virtual cycle 6
		d1 = transfer.d[1];
		d0 -= dBias;
		f3 = d3;
		(++dest)[0] = f2;

		//Virtual cycle 5
		d2 = transfer.d[2];
		d1 -= dBias;
		f0 = d0;
		(++dest)[0] = f3;

		//Virtual cycle 4
		d3 = transfer.d[3];
		d2 -= dBias;
		f1 = d1;
		(++dest)[0] = f0;

		//Virtual cycle 3
		d3 -= dBias;
		f2 = d2;
		(++dest)[0] = f1;

		//Virtual cycle 2
		f3 = d3;
		(++dest)[0] = f2;

		//Virtual cycle 1
		(++dest)[0] = f3;
	}


	while( count-- )
	{
		int0 = __lwbrx( 0, ++src);
		int0 ^= 0x80000000UL;
		transfer.i[1] = int0;

		d0 = transfer.d[0];
		d0 -= dBias;
		f0 = d0;
		(++dest)[0] = f0;
	}
}

void Float32ToInt8( float *src, SInt8 *dst, unsigned int count )
{
	register double		scale = 2147483648.0;
	register double		round = 128.0;
	unsigned long		loopCount = count / 4;
	long				buffer[2];
	register float		startingFloat;
	register double 	scaled;
	register double 	converted;
	register SInt8		copy;

	//
	//	The fastest way to do this is to set up a staggered loop that models a 7 stage virtual pipeline:
	//
	//		stage 1:		load the src value
	//		stage 2:		scale it to LONG_MIN...LONG_MAX and add a rounding value to it
	//		stage 3:		convert it to an integer within the FP register
	//		stage 4:		store that to the stack
	//		stage 5:		 (do nothing to let the store complete)
	//		stage 6:		load the high half word from the stack
	//		stage 7:		store it to the destination
	//
	//	We set it up so that at any given time 7 different pieces of data are being worked on at a time.
	//	Because of the do nothing stage, the inner loop had to be unrolled by one, so in actuality, each
	//	inner loop iteration represents two virtual clock cycles that push data through our virtual pipeline.
	//
	//	The reason why this works is that this allows us to break data dependency chains and insert 5 real 
	//	operations in between every virtual pipeline stage. This means 5 instructions between each data 
	//	dependency, which is just enough to keep all of our real pipelines happy. The data flow follows 
	//	standard pipeline diagrams:
	//
	//					stage1	stage2	stage3	stage4	stage5	stage6	stage7
	//	virtual cycle 1:	data1	-		-		-		-		-		-
	//	virtual cycle 2:	data2	data1	-		-		-		-		-
	//	virtual cycle 3:	data3	data2	data1	-		-		-		-
	//	virtual cycle 4:	data4	data3	data2	data1	-		-		-		
	//	virtual cycle 5:	data5	data4	data3	data2	data1	-		-			   
	//	virtual cycle 6:	data6	data5	data4	data3	data2	data1	-	
	//
	//	inner loop:
	//	  virtual cycle A:	data7	data6	data5	data4	data3	data2	data1					  
	//	  virtual cycle B:	data8	data7	data6	data5	data4	data3	data2	
	//
	//	virtual cycle 7 -		dataF	dataE	dataD	dataC	dataB	dataA
	//	virtual cycle 8 -		-		dataF	dataE	dataD	dataC	dataB	
	//	virtual cycle 9 -		-		-		dataF	dataE	dataD	dataC
	//	virtual cycle 10	-		-		-		-		dataF	dataE	dataD  
	//	virtual cycle 11	-		-		-		-		-		dataF	dataE	
	//	virtual cycle 12	-		-		-		-		-		-		dataF						 
	
	if( count >= 6 )
	{
		//virtual cycle 1
		startingFloat = (src++)[0];
		
		//virtual cycle 2
		scaled = startingFloat * scale + round;
		startingFloat = (src++)[0];
		
		//virtual cycle 3
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (src++)[0];
		
		//virtual cycle 4
		__stfiwx( converted, 0, buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (src++)[0];
		
		//virtual cycle 5
		__stfiwx( converted, sizeof(float), buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (src++)[0];
		
		//virtual cycle 6
		copy = ((SInt8*) buffer)[0];	
		__stfiwx( converted, 0, buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (src++)[0];
		
		count -= 6;
		loopCount = count / 2;
		count &= 1;
		while( loopCount-- )
		{
			register float	startingFloat2;
			register double scaled2;
			register double converted2;
			register SInt8	copy2;
			
			//virtual Cycle A
			(dst++)[0] = copy;
			__asm__ __volatile__ ( "fctiw %0, %1" : "=f" (converted2) : "f" ( scaled ) );
			copy2 = ((SInt8*) buffer)[4];		
			__asm__ __volatile__ ( "fmadd %0, %1, %2, %3" : "=f" (scaled2) : "f" ( startingFloat), "f" (scale), "f" (round) );
			__asm__ __volatile__ ( "stfiwx %0, %1, %2" : : "f" (converted), "b%" (sizeof(float)), "r" (buffer) : "memory" );
			startingFloat2 = (src++)[0];

			//virtual cycle B
			(dst++)[0] = copy2;
			__asm__ __volatile__ ( "fctiw %0, %1" : "=f" (converted) : "f" ( scaled2 ) );
			copy = ((SInt8*) buffer)[0];
			__asm__ __volatile__ ( "fmadd %0, %1, %2, %3" : "=f" (scaled) : "f" ( startingFloat2), "f" (scale), "f" (round) );
			__asm__ __volatile__ ( "stfiwx %0, %1, %2" : : "f" (converted2), "b%" (0), "r" (buffer) : "memory" );
			startingFloat = (src++)[0];
		}
		
		//Virtual Cycle 7
		(dst++)[0] = copy;
		copy = ((SInt8*) buffer)[4];
		__stfiwx( converted, sizeof(float), buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		
		//Virtual Cycle 8
		(dst++)[0] = copy;
		copy = ((SInt8*) buffer)[0];
		__stfiwx( converted, 0, buffer );
		converted = __fctiw( scaled );
	
		//Virtual Cycle 9
		(dst++)[0] = copy;
		copy = ((SInt8*) buffer)[4];
		__stfiwx( converted, sizeof(float), buffer );
		
		//Virtual Cycle 10
		(dst++)[0] = copy;
		copy = ((SInt8*) buffer)[0];
	
		//Virtual Cycle 11
		(dst++)[0] = copy;
		copy = ((SInt8*) buffer)[4];

		//Virtual Cycle 11
		(dst++)[0] = copy;
	}

	//clean up any extras
	while( count-- )
	{
		double scaled = src[0] * scale + round;
		double converted = __fctiw( scaled );
		__stfiwx( converted, 0, buffer );
		dst[0] = buffer[0] >> 24;
		src++;
		dst++;
	}
}


void Float32ToNativeInt16( float *src, signed short *dst, unsigned int count )
{
	register double		scale = 2147483648.0;
	register double		round = 32768.0;
	unsigned long		loopCount = count / 4;
	long				buffer[2];
	register float		startingFloat;
	register double 	scaled;
	register double 	converted;
	register short		copy;

	//
	//	The fastest way to do this is to set up a staggered loop that models a 7 stage virtual pipeline:
	//
	//		stage 1:		load the src value
	//		stage 2:		scale it to LONG_MIN...LONG_MAX and add a rounding value to it
	//		stage 3:		convert it to an integer within the FP register
	//		stage 4:		store that to the stack
	//		stage 5:		 (do nothing to let the store complete)
	//		stage 6:		load the high half word from the stack
	//		stage 7:		store it to the destination
	//
	//	We set it up so that at any given time 7 different pieces of data are being worked on at a time.
	//	Because of the do nothing stage, the inner loop had to be unrolled by one, so in actuality, each
	//	inner loop iteration represents two virtual clock cycles that push data through our virtual pipeline.
	//
	//	The reason why this works is that this allows us to break data dependency chains and insert 5 real 
	//	operations in between every virtual pipeline stage. This means 5 instructions between each data 
	//	dependency, which is just enough to keep all of our real pipelines happy. The data flow follows 
	//	standard pipeline diagrams:
	//
	//					stage1	stage2	stage3	stage4	stage5	stage6	stage7
	//	virtual cycle 1:	data1	-		-		-		-		-		-
	//	virtual cycle 2:	data2	data1	-		-		-		-		-
	//	virtual cycle 3:	data3	data2	data1	-		-		-		-
	//	virtual cycle 4:	data4	data3	data2	data1	-		-		-		
	//	virtual cycle 5:	data5	data4	data3	data2	data1	-		-			   
	//	virtual cycle 6:	data6	data5	data4	data3	data2	data1	-	
	//
	//	inner loop:
	//	  virtual cycle A:	data7	data6	data5	data4	data3	data2	data1					  
	//	  virtual cycle B:	data8	data7	data6	data5	data4	data3	data2	
	//
	//	virtual cycle 7 -		dataF	dataE	dataD	dataC	dataB	dataA
	//	virtual cycle 8 -		-		dataF	dataE	dataD	dataC	dataB	
	//	virtual cycle 9 -		-		-		dataF	dataE	dataD	dataC
	//	virtual cycle 10	-		-		-		-		dataF	dataE	dataD  
	//	virtual cycle 11	-		-		-		-		-		dataF	dataE	
	//	virtual cycle 12	-		-		-		-		-		-		dataF						 
	
	if( count >= 6 )
	{
		//virtual cycle 1
		startingFloat = (src++)[0];
		
		//virtual cycle 2
		scaled = startingFloat * scale + round;
		startingFloat = (src++)[0];
		
		//virtual cycle 3
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (src++)[0];
		
		//virtual cycle 4
		__stfiwx( converted, 0, buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (src++)[0];
		
		//virtual cycle 5
		__stfiwx( converted, sizeof(float), buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (src++)[0];
		
		//virtual cycle 6
		copy = ((short*) buffer)[0];
		__stfiwx( converted, 0, buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (src++)[0];
		
		count -= 6;
		loopCount = count / 2;
		count &= 1;
		while( loopCount-- )
		{
			register float	startingFloat2;
			register double scaled2;
			register double converted2;
			register short	copy2;
			
			//virtual Cycle A
			(dst++)[0] = copy;
			__asm__ __volatile__ ( "fctiw %0, %1" : "=f" (converted2) : "f" ( scaled ) );
			copy2 = ((short*) buffer)[2];
			__asm__ __volatile__ ( "fmadd %0, %1, %2, %3" : "=f" (scaled2) : "f" ( startingFloat), "f" (scale), "f" (round) );
			__asm__ __volatile__ ( "stfiwx %0, %1, %2" : : "f" (converted), "b%" (sizeof(float)), "r" (buffer) : "memory" );
			startingFloat2 = (src++)[0];

			//virtual cycle B
			(dst++)[0] = copy2;
			__asm__ __volatile__ ( "fctiw %0, %1" : "=f" (converted) : "f" ( scaled2 ) );
			copy = ((short*) buffer)[0];
			__asm__ __volatile__ ( "fmadd %0, %1, %2, %3" : "=f" (scaled) : "f" ( startingFloat2), "f" (scale), "f" (round) );
			__asm__ __volatile__ ( "stfiwx %0, %1, %2" : : "f" (converted2), "b%" (0), "r" (buffer) : "memory" );
			startingFloat = (src++)[0];
		}
		
		//Virtual Cycle 7
		(dst++)[0] = copy;
		copy = ((short*) buffer)[2];
		__stfiwx( converted, sizeof(float), buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		
		//Virtual Cycle 8
		(dst++)[0] = copy;
		copy = ((short*) buffer)[0];
		__stfiwx( converted, 0, buffer );
		converted = __fctiw( scaled );
	
		//Virtual Cycle 9
		(dst++)[0] = copy;
		copy = ((short*) buffer)[2];
		__stfiwx( converted, sizeof(float), buffer );
		
		//Virtual Cycle 10
		(dst++)[0] = copy;
		copy = ((short*) buffer)[0];
	
		//Virtual Cycle 11
		(dst++)[0] = copy;
		copy = ((short*) buffer)[2];

		//Virtual Cycle 11
		(dst++)[0] = copy;
	}

	//clean up any extras
	while( count-- )
	{
		double scaled = src[0] * scale + round;
		double converted = __fctiw( scaled );
		__stfiwx( converted, 0, buffer );
		dst[0] = buffer[0] >> 16;
		src++;
		dst++;
	}
}

void Float32ToSwapInt16( float *src, signed short *dst, unsigned int count )
{
	register double		scale = 2147483648.0;
	register double		round = 32768.0;
	unsigned long	loopCount = count / 4;
	long		buffer[2];
	register float	startingFloat;
	register double scaled;
	register double converted;
	register short	copy;

	//
	//	The fastest way to do this is to set up a staggered loop that models a 7 stage virtual pipeline:
	//
	//		stage 1:		load the src value
	//		stage 2:		scale it to LONG_MIN...LONG_MAX and add a rounding value to it
	//		stage 3:		convert it to an integer within the FP register
	//		stage 4:		store that to the stack
	//		stage 5:		 (do nothing to let the store complete)
	//		stage 6:		load the high half word from the stack
	//		stage 7:		store it to the destination
	//
	//	We set it up so that at any given time 7 different pieces of data are being worked on at a time.
	//	Because of the do nothing stage, the inner loop had to be unrolled by one, so in actuality, each
	//	inner loop iteration represents two virtual clock cycles that push data through our virtual pipeline.
	//
	//	The reason why this works is that this allows us to break data dependency chains and insert 5 real 
	//	operations in between every virtual pipeline stage. This means 5 instructions between each data 
	//	dependency, which is just enough to keep all of our real pipelines happy. The data flow follows 
	//	standard pipeline diagrams:
	//
	//					stage1	stage2	stage3	stage4	stage5	stage6	stage7
	//	virtual cycle 1:	data1	-		-		-		-		-		-
	//	virtual cycle 2:	data2	data1	-		-		-		-		-
	//	virtual cycle 3:	data3	data2	data1	-		-		-		-
	//	virtual cycle 4:	data4	data3	data2	data1	-		-		-		
	//	virtual cycle 5:	data5	data4	data3	data2	data1	-		-			   
	//	virtual cycle 6:	data6	data5	data4	data3	data2	data1	-	
	//
	//	inner loop:
	//	  virtual cycle A:	data7	data6	data5	data4	data3	data2	data1					  
	//	  virtual cycle B:	data8	data7	data6	data5	data4	data3	data2	
	//
	//	virtual cycle 7 -		dataF	dataE	dataD	dataC	dataB	dataA
	//	virtual cycle 8 -		-		dataF	dataE	dataD	dataC	dataB	
	//	virtual cycle 9 -		-		-		dataF	dataE	dataD	dataC
	//	virtual cycle 10	-		-		-		-		dataF	dataE	dataD  
	//	virtual cycle 11	-		-		-		-		-		dataF	dataE	
	//	virtual cycle 12	-		-		-		-		-		-		dataF						 
	
	if( count >= 6 )
	{
		//virtual cycle 1
		startingFloat = (src++)[0];
		
		//virtual cycle 2
		scaled = startingFloat * scale + round;
		startingFloat = (src++)[0];
		
		//virtual cycle 3
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (src++)[0];
		
		//virtual cycle 4
		__stfiwx( converted, 0, buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (src++)[0];
		
		//virtual cycle 5
		__stfiwx( converted, sizeof(float), buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (src++)[0];
		
		//virtual cycle 6
		copy = ((short*) buffer)[0];
		__stfiwx( converted, 0, buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (src++)[0];
		
		count -= 6;
		loopCount = count / 2;
		count &= 1;
		while( loopCount-- )
		{
			register float	startingFloat2;
			register double scaled2;
			register double converted2;
			register short	copy2;
			
			//virtual Cycle A
//			  (dst++)[0] = copy;
			__asm__ __volatile__ ( "sthbrx %0, 0, %1" : : "r" (copy), "r" (dst) );
			__asm__ __volatile__ ( "fctiw %0, %1" : "=f" (converted2) : "f" ( scaled ) );
			copy2 = ((short*) buffer)[2];
			__asm__ __volatile__ ( "fmadd %0, %1, %2, %3" : "=f" (scaled2) : "f" ( startingFloat), "f" (scale), "f" (round) );
			__asm__ __volatile__ ( "stfiwx %0, %1, %2" : : "f" (converted), "b%" (sizeof(float)), "r" (buffer) : "memory" );
			startingFloat2 = (src)[0];	src+=2;

			//virtual cycle B
//			  (dst++)[0] = copy2;
			dst+=2;
			__asm__ __volatile__ ( "sthbrx %0, %1, %2" : : "r" (copy2), "r" (-2), "r" (dst) );	
			__asm__ __volatile__ ( "fctiw %0, %1" : "=f" (converted) : "f" ( scaled2 ) );
			copy = ((short*) buffer)[0];
			__asm__ __volatile__ ( "fmadd %0, %1, %2, %3" : "=f" (scaled) : "f" ( startingFloat2), "f" (scale), "f" (round) );
			__asm__ __volatile__ ( "stfiwx %0, %1, %2" : : "f" (converted2), "b%" (0), "r" (buffer) : "memory" );
			startingFloat = (src)[-1];	
		}
		
		//Virtual Cycle 7
		__asm__ __volatile__ ( "sthbrx %0, 0, %1" : : "r" (copy),  "r" (dst) );	 dst++;
		copy = ((short*) buffer)[2];
		__stfiwx( converted, sizeof(float), buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		
		//Virtual Cycle 8
		__asm__ __volatile__ ( "sthbrx %0, 0, %1" : : "r" (copy),  "r" (dst) );	 dst++;
		copy = ((short*) buffer)[0];
		__stfiwx( converted, 0, buffer );
		converted = __fctiw( scaled );
	
		//Virtual Cycle 9
		__asm__ __volatile__ ( "sthbrx %0, 0, %1" : : "r" (copy),  "r" (dst) );	 dst++;
		copy = ((short*) buffer)[2];
		__stfiwx( converted, sizeof(float), buffer );
		
		//Virtual Cycle 10
		__asm__ __volatile__ ( "sthbrx %0, 0, %1" : : "r" (copy),  "r" (dst) );	 dst++;
		copy = ((short*) buffer)[0];
	
		//Virtual Cycle 11
		__asm__ __volatile__ ( "sthbrx %0, 0, %1" : : "r" (copy),  "r" (dst) );	 dst++;
		copy = ((short*) buffer)[2];

		//Virtual Cycle 11
		__asm__ __volatile__ ( "sthbrx %0, 0, %1" : : "r" (copy),  "r" (dst) );	 dst++;
	}

	//clean up any extras
	while( count-- )
	{
		double scaled = src[0] * scale + round;
		double converted = __fctiw( scaled );
		__stfiwx( converted, 0, buffer );
		copy = buffer[0] >> 16;
		__asm__ __volatile__ ( "sthbrx %0, 0, %1" : : "r" (copy),  "r" (dst) );	 
		src++;
		dst++;
	}
}


void Float32ToNativeInt24( float *src, signed long *dst, unsigned int count )
{
	register double		scale = 2147483648.0;
	register double		round = 0.5 * 256.0;
	unsigned long	loopCount = count / 4;
	long		buffer[4];
	register float	startingFloat, startingFloat2;
	register double scaled, scaled2;
	register double converted, converted2;
	register long	copy1;//, merge1, rotate1;
	register long	copy2;//, merge2, rotate2;
	register long	copy3;//, merge3, rotate3;
	register long	copy4;//, merge4, rotate4;
	register double		oldSetting;


	//Set the FPSCR to round to -Inf mode
	{
		union
		{
			double	d;
			int		i[2];
		}setting;
		register double newSetting;

		//Read the the current FPSCR value
		asm volatile ( "mffs %0" : "=f" ( oldSetting ) );

		//Store it to the stack
		setting.d = oldSetting;

		//Read in the low 32 bits and mask off the last two bits so they are zero
		//in the integer unit. These two bits set to zero means round to nearest mode.
		//Finally, then store the result back
		setting.i[1] |= 3;

		//Load the new FPSCR setting into the FP register file again
		newSetting = setting.d;

		//Change the FPSCR to the new setting
		asm volatile( "mtfsf 7, %0" : : "f" (newSetting ) );
	}


	//
	//	The fastest way to do this is to set up a staggered loop that models a 7 stage virtual pipeline:
	//
	//		stage 1:		load the src value
	//		stage 2:		scale it to LONG_MIN...LONG_MAX and add a rounding value to it
	//		stage 3:		convert it to an integer within the FP register
	//		stage 4:		store that to the stack
	//		stage 5:		 (do nothing to let the store complete)
	//		stage 6:		load the high half word from the stack
		//		stage 7:		merge with later data to form a 32 bit word
		//		stage 8:		possible rotate to correct byte order
	//		stage 9:		store it to the destination
	//
	//	We set it up so that at any given time 7 different pieces of data are being worked on at a time.
	//	Because of the do nothing stage, the inner loop had to be unrolled by one, so in actuality, each
	//	inner loop iteration represents two virtual clock cycles that push data through our virtual pipeline.
	//
	//	The reason why this works is that this allows us to break data dependency chains and insert 5 real
	//	operations in between every virtual pipeline stage. This means 5 instructions between each data
	//	dependency, which is just enough to keep all of our real pipelines happy. The data flow follows
	//	standard pipeline diagrams:
	//
	//				stage1	stage2	stage3	stage4	stage5	stage6	stage7	stage8	stage9
	//	virtual cycle 1:	data1	-	-	-		-		-		-		-		-
	//	virtual cycle 2:	data2	data1	-	-		-		-		-		-		-
	//	virtual cycle 3:	data3	data2	data1	-		-		-		-		-		-
	//	virtual cycle 4:	data4	data3	data2	data1	-		-		-		-		-
	//	virtual cycle 5:	data5	data4	data3	data2	data1	-		-		-		-
	//	virtual cycle 6:	data6	data5	data4	data3	data2	data1	-		-		-
	//	virtual cycle 7:	data7	data6	data5	data4	data3	data2	data1	-		-
	//	virtual cycle 8:	data8	data7	data6	data5	data4	data3	data2	data1	-
	//
	//	inner loop:
	//	  virtual cycle A:	data9	data8	data7	data6	data5	data4	data3	data2	data1
	//	  virtual cycle B:	data10	data9	data8	data7	data6	data5	data4	data3	data2
	//	  virtual cycle C:	data11	data10	data9	data8	data7	data6	data5	data4	data3
	//	  virtual cycle D:	data12	data11	data10	data9	data8	data7	data6	data5	data4
	//
	//	virtual cycle 9		-	dataH	dataG	dataF	dataE	dataD	dataC	dataB	dataA
	//	virtual cycle 10	-	-	dataH	dataG	dataF	dataE	dataD	dataC	dataB
	//	virtual cycle 11	-	-	-	dataH	dataG	dataF	dataE	dataD	dataC
	//	virtual cycle 12	-	-	-	-		dataH	dataG	dataF	dataE	dataD
	//	virtual cycle 13	-	-	-	-		-		dataH	dataG	dataF	dataE
	//	virtual cycle 14	-	-	-	-		-		-		dataH	dataG	dataF
	//	virtual cycle 15	-	-	-	-		-		-		-	dataH	dataG	
	//	virtual cycle 16	-	-	-	-		-		-		-	-	dataH

	src--;
	dst--;

	if( count >= 8 )
	{
		//virtual cycle 1
		startingFloat = (++src)[0];

		//virtual cycle 2
		scaled = startingFloat * scale + round;
		startingFloat = (++src)[0];

		//virtual cycle 3
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (++src)[0];

		//virtual cycle 4
		__stfiwx( converted, 0, buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (++src)[0];

		//virtual cycle 5
		__stfiwx( converted, sizeof(float), buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (++src)[0];

		//virtual cycle 6
		copy1 = buffer[0];
		__stfiwx( converted, 2 * sizeof( float), buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (++src)[0];

		//virtual cycle 7
		copy2 = buffer[1];
		__stfiwx( converted, 3 * sizeof(float), buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (++src)[0];

		//virtual cycle 8
		copy1 = __rlwimi( copy1, copy2, 8, 24, 31 );
		copy3 = buffer[2];
		__stfiwx( converted, 0, buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (++src)[0];

		count -= 8;
		loopCount = count / 4;
		count &= 3;
		while( loopCount-- )
		{
			//virtual cycle A
			//no store yet						//store
			//no rotation needed for copy1,				//rotate
			__asm__ __volatile__( "fmadds %0, %1, %2, %3" : "=f"(scaled2) : "f" (startingFloat), "f" ( scale ), "f" ( round ));		//scale for clip and add rounding
			startingFloat2 = (++src)[0];				//load the float
			__asm__ __volatile__( "fctiw %0, %1" : "=f" (converted2) : "f" ( scaled ) );				//convert to int and clip
			 copy4 = buffer[3];						//load clipped int back in
			copy2 = __rlwimi_volatile( copy2, copy3, 8, 24, 7 );			//merge
			__stfiwx( converted, 1 * sizeof(float), buffer );		//store clipped int

			//virtual Cycle B
			__asm__ __volatile__( "fmadds %0, %1, %2, %3" : "=f"(scaled) : "f" (startingFloat2), "f" ( scale ), "f" ( round ));		//scale for clip and add rounding
			startingFloat = (++src)[0];					//load the float
			 __asm__ __volatile__( "fctiw %0, %1" : "=f" (converted) : "f" ( scaled2 ) );				//convert to int and clip
			(++dst)[0] = copy1;						//store
			copy3 = __rlwimi_volatile( copy3, copy4, 8, 24, 15 );	//merge with adjacent pixel
			copy1 = buffer[0];						//load clipped int back in
			copy2 = __rlwimi_volatile( copy2, copy2, 8, 0, 31 );	//rotate
			__stfiwx( converted2, 2 * sizeof(float), buffer );		//store clipped int

			//virtual Cycle C
			__asm__ __volatile__( "fmadds %0, %1, %2, %3" : "=f"(scaled2) : "f" (startingFloat), "f" ( scale ), "f" ( round ));		//scale for clip and add rounding
			startingFloat2 = (++src)[0];				//load the float
			//We dont store copy 4 so no merge needs to be done to it	//merge with adjacent pixel
			converted2 = __fctiw( scaled );				//convert to int and clip
			(++dst)[0] = copy2;						//store
			copy3 = __rlwimi_volatile( copy3, copy3, 16, 0, 31 );		//rotate
			copy2 = buffer[1];						//load clipped int back in
			__stfiwx( converted, 3 * sizeof(float), buffer );		//store clipped int

			//virtual Cycle D
			__asm__ ( "fmadds %0, %1, %2, %3" : "=f"(scaled) : "f" (startingFloat2), "f" ( scale ), "f" ( round ));		//scale for clip and add rounding
			startingFloat = (++src)[0];					//load the float
			converted = __fctiw( scaled2 );				//convert to int and clip
			//We dont store copy 4 so no rotation needs to be done to it//rotate
			(++dst)[0] = copy3;						//store
			copy1 = __rlwimi_volatile( copy1, copy2, 8, 24, 31 );		//merge with adjacent pixel
			 __stfiwx( converted2, 0 * sizeof(float), buffer );		//store clipped int
			copy3 = buffer[2];						//load clipped int back in
		}

		//virtual cycle 9
		//no store yet						//store
		//no rotation needed for copy1,				//rotate
		copy2 = __rlwimi( copy2, copy3, 8, 24, 7 );		//merge
		copy4 = buffer[3];					//load clipped int back in
		__stfiwx( converted, 1 * sizeof(float), buffer );	//store clipped int
		converted2 = __fctiw( scaled );				//convert to int and clip
		scaled2 = startingFloat * scale + round;		//scale for clip and add rounding

		//virtual Cycle 10
		(++dst)[0] = copy1;						//store
		copy2 = __rlwimi( copy2, copy2, 8, 0, 31 );			//rotate
		copy3 = __rlwimi( copy3, copy4, 8, 24, 15 );		//merge with adjacent pixel
		copy1 = buffer[0];					//load clipped int back in
		__stfiwx( converted2, 2 * sizeof(float), buffer );	//store clipped int
		converted = __fctiw( scaled2 );				//convert to int and clip

		//virtual Cycle 11
		(++dst)[0] = copy2;						//store
		copy3 = __rlwimi( copy3, copy3, 16, 0, 31 );		//rotate
		//We dont store copy 4 so no merge needs to be done to it//merge with adjacent pixel
		copy2 = buffer[1];					//load clipped int back in
		__stfiwx( converted, 3 * sizeof(float), buffer );	//store clipped int

		//virtual Cycle 12
		(++dst)[0] = copy3;						//store
		//We dont store copy 4 so no rotation needs to be done to it//rotate
		copy1 = __rlwimi( copy1, copy2, 8, 24, 31 );		//merge with adjacent pixel
		copy3 = buffer[2];						//load clipped int back in

		//virtual cycle 13
		//no store yet						//store
		//no rotation needed for copy1,				//rotate
		copy2 = __rlwimi( copy2, copy3, 8, 24, 7 );		//merge
		copy4 = buffer[3];					//load clipped int back in

		//virtual Cycle 14
		(++dst)[0] = copy1;						//store
		copy2 = __rlwimi( copy2, copy2, 8, 0, 31 );			//rotate
		copy3 = __rlwimi( copy3, copy4, 8, 24, 15 );		//merge with adjacent pixel

		//virtual Cycle 15
		(++dst)[0] = copy2;						//store
		copy3 = __rlwimi( copy3, copy3, 16, 0, 31 );		//rotate

		//virtual Cycle 16
		(++dst)[0] = copy3;						//store
	}

	//clean up any extras
	dst++;
	while( count-- )
	{
		startingFloat = (++src)[0];				//load the float
		scaled = startingFloat * scale + round;			//scale for clip and add rounding
		converted = __fctiw( scaled );				//convert to int and clip
		__stfiwx( converted, 0, buffer );			//store clipped int
		copy1 = buffer[0];					//load clipped int back in
		((signed char*) dst)[0] = copy1 >> 24;
		dst = (signed long*) ((signed char*) dst + 1 );
		((unsigned short*) dst)[0] = copy1 >> 8;
		dst = (signed long*) ((unsigned short*) dst + 1 );
	}

	//restore the old FPSCR setting
	__asm__ __volatile__ ( "mtfsf 7, %0" : : "f" (oldSetting) );
}

void Float32ToSwapInt24( float *src, signed long *dst, unsigned int count )
{
	register double		scale = 2147483648.0;
	register double		round = 0.5 * 256.0;
	unsigned long	loopCount = count / 4;
	long		buffer[4];
	register float	startingFloat, startingFloat2;
	register double scaled, scaled2;
	register double converted, converted2;
	register long	copy1;
	register long	copy2;
	register long	copy3;
	register long	copy4;
	register double		oldSetting;


	//Set the FPSCR to round to -Inf mode
	{
		union
		{
			double	d;
			int		i[2];
		}setting;
		register double newSetting;

		//Read the the current FPSCR value
		asm volatile ( "mffs %0" : "=f" ( oldSetting ) );

		//Store it to the stack
		setting.d = oldSetting;

		//Read in the low 32 bits and mask off the last two bits so they are zero
		//in the integer unit. These two bits set to zero means round to nearest mode.
		//Finally, then store the result back
		setting.i[1] |= 3;

		//Load the new FPSCR setting into the FP register file again
		newSetting = setting.d;

		//Change the FPSCR to the new setting
		asm volatile( "mtfsf 7, %0" : : "f" (newSetting ) );
	}


	//
	//	The fastest way to do this is to set up a staggered loop that models a 7 stage virtual pipeline:
	//
	//		stage 1:		load the src value
	//		stage 2:		scale it to LONG_MIN...LONG_MAX and add a rounding value to it
	//		stage 3:		convert it to an integer within the FP register
	//		stage 4:		store that to the stack
	//		stage 5:		 (do nothing to let the store complete)
	//		stage 6:		load the high half word from the stack
		//		stage 7:		merge with later data to form a 32 bit word
		//		stage 8:		possible rotate to correct byte order
	//		stage 9:		store it to the destination
	//
	//	We set it up so that at any given time 7 different pieces of data are being worked on at a time.
	//	Because of the do nothing stage, the inner loop had to be unrolled by one, so in actuality, each
	//	inner loop iteration represents two virtual clock cycles that push data through our virtual pipeline.
	//
	//	The reason why this works is that this allows us to break data dependency chains and insert 5 real
	//	operations in between every virtual pipeline stage. This means 5 instructions between each data
	//	dependency, which is just enough to keep all of our real pipelines happy. The data flow follows
	//	standard pipeline diagrams:
	//
	//				stage1	stage2	stage3	stage4	stage5	stage6	stage7	stage8	stage9
	//	virtual cycle 1:	data1	-	-	-		-		-		-		-		-
	//	virtual cycle 2:	data2	data1	-	-		-		-		-		-		-
	//	virtual cycle 3:	data3	data2	data1	-		-		-		-		-		-
	//	virtual cycle 4:	data4	data3	data2	data1	-		-		-		-		-
	//	virtual cycle 5:	data5	data4	data3	data2	data1	-		-		-		-
	//	virtual cycle 6:	data6	data5	data4	data3	data2	data1	-		-		-
	//	virtual cycle 7:	data7	data6	data5	data4	data3	data2	data1	-		-
	//	virtual cycle 8:	data8	data7	data6	data5	data4	data3	data2	data1	-
	//
	//	inner loop:
	//	  virtual cycle A:	data9	data8	data7	data6	data5	data4	data3	data2	data1
	//	  virtual cycle B:	data10	data9	data8	data7	data6	data5	data4	data3	data2
	//	  virtual cycle C:	data11	data10	data9	data8	data7	data6	data5	data4	data3
	//	  virtual cycle D:	data12	data11	data10	data9	data8	data7	data6	data5	data4
	//
	//	virtual cycle 9		-	dataH	dataG	dataF	dataE	dataD	dataC	dataB	dataA
	//	virtual cycle 10	-	-	dataH	dataG	dataF	dataE	dataD	dataC	dataB
	//	virtual cycle 11	-	-	-	dataH	dataG	dataF	dataE	dataD	dataC
	//	virtual cycle 12	-	-	-	-		dataH	dataG	dataF	dataE	dataD
	//	virtual cycle 13	-	-	-	-		-		dataH	dataG	dataF	dataE
	//	virtual cycle 14	-	-	-	-		-		-		dataH	dataG	dataF
	//	virtual cycle 15	-	-	-	-		-		-		-	dataH	dataG	
	//	virtual cycle 16	-	-	-	-		-		-		-	-	dataH

	src--;
	dst--;

	if( count >= 8 )
	{
		//virtual cycle 1
		startingFloat = (++src)[0];

		//virtual cycle 2
		scaled = startingFloat * scale + round;
		startingFloat = (++src)[0];

		//virtual cycle 3
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (++src)[0];

		//virtual cycle 4
		__stfiwx( converted, 0, buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (++src)[0];

		//virtual cycle 5
		__stfiwx( converted, sizeof(float), buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (++src)[0];

		//virtual cycle 6
		copy1 = __lwbrx( 0, buffer );
		__stfiwx( converted, 2 * sizeof( float), buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (++src)[0];

		//virtual cycle 7
		copy2 = __lwbrx( 4, buffer );
		__stfiwx( converted, 3 * sizeof(float), buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (++src)[0];

		//virtual cycle 8
		copy1 = __rlwimi( copy1, copy2, 8, 0, 7 );
		copy3 = __lwbrx( 8, buffer );;
		__stfiwx( converted, 0, buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale + round;
		startingFloat = (++src)[0];

		count -= 8;
		loopCount = count / 4;
		count &= 3;
		while( loopCount-- )
		{
			//virtual cycle A
			//no store yet						//store
			copy1 = __rlwimi( copy1, copy1, 8, 0, 31 );			//rotate
			__asm__ __volatile__( "fmadds %0, %1, %2, %3" : "=f"(scaled2) : "f" (startingFloat), "f" ( scale ), "f" ( round ));		//scale for clip and add rounding
			startingFloat2 = (++src)[0];				//load the float
			__asm__ __volatile__( "fctiw %0, %1" : "=f" (converted2) : "f" ( scaled ) );				//convert to int and clip
			 copy4 = __lwbrx( 12, buffer );						//load clipped int back in
			copy2 = __rlwimi_volatile( copy2, copy3, 8, 0, 15 );			//merge
			__stfiwx( converted, 1 * sizeof(float), buffer );		//store clipped int

			//virtual Cycle B
			__asm__ __volatile__( "fmadds %0, %1, %2, %3" : "=f"(scaled) : "f" (startingFloat2), "f" ( scale ), "f" ( round ));		//scale for clip and add rounding
			startingFloat = (++src)[0];					//load the float
			 __asm__ __volatile__( "fctiw %0, %1" : "=f" (converted) : "f" ( scaled2 ) );				//convert to int and clip
			(++dst)[0] = copy1;						//store
			copy4 = __rlwimi_volatile( copy4, copy3, 24, 0, 7 );	//merge with adjacent pixel
			copy1 = __lwbrx( 0, buffer );						//load clipped int back in
			copy2 = __rlwimi_volatile( copy2, copy2, 16, 0, 31 );	//rotate
			__stfiwx( converted2, 2 * sizeof(float), buffer );		//store clipped int

			//virtual Cycle C
			__asm__ __volatile__( "fmadds %0, %1, %2, %3" : "=f"(scaled2) : "f" (startingFloat), "f" ( scale ), "f" ( round ));		//scale for clip and add rounding
			startingFloat2 = (++src)[0];				//load the float
			converted2 = __fctiw( scaled );				//convert to int and clip
			(++dst)[0] = copy2;						//store
			copy2 = __lwbrx( 4, buffer );						//load clipped int back in
			__stfiwx( converted, 3 * sizeof(float), buffer );		//store clipped int


			//virtual Cycle D
			__asm__ ( "fmadds %0, %1, %2, %3" : "=f"(scaled) : "f" (startingFloat2), "f" ( scale ), "f" ( round ));		//scale for clip and add rounding
			startingFloat = (++src)[0];					//load the float
			converted = __fctiw( scaled2 );				//convert to int and clip
			(++dst)[0] = copy4;						//store
			copy1 = __rlwimi_volatile( copy1, copy2, 8, 0, 7 );		//merge with adjacent pixel
			 __stfiwx( converted2, 0 * sizeof(float), buffer );		//store clipped int
			copy3 = __lwbrx( 8, buffer );						//load clipped int back in
		}

		//virtual cycle A
		//no store yet						//store
		copy1 = __rlwimi( copy1, copy1, 8, 0, 31 );			//rotate
		__asm__ __volatile__( "fmadds %0, %1, %2, %3" : "=f"(scaled2) : "f" (startingFloat), "f" ( scale ), "f" ( round ));		//scale for clip and add rounding
		__asm__ __volatile__( "fctiw %0, %1" : "=f" (converted2) : "f" ( scaled ) );				//convert to int and clip
			copy4 = __lwbrx( 12, buffer );						//load clipped int back in
		copy2 = __rlwimi_volatile( copy2, copy3, 8, 0, 15 );			//merge
		__stfiwx( converted, 1 * sizeof(float), buffer );		//store clipped int

		//virtual Cycle B
			__asm__ __volatile__( "fctiw %0, %1" : "=f" (converted) : "f" ( scaled2 ) );				//convert to int and clip
		(++dst)[0] = copy1;						//store
		copy4 = __rlwimi_volatile( copy4, copy3, 24, 0, 7 );	//merge with adjacent pixel
		copy1 = __lwbrx( 0, buffer );						//load clipped int back in
		copy2 = __rlwimi_volatile( copy2, copy2, 16, 0, 31 );	//rotate
		__stfiwx( converted2, 2 * sizeof(float), buffer );		//store clipped int

		//virtual Cycle C
		(++dst)[0] = copy2;						//store
		copy2 = __lwbrx( 4, buffer );						//load clipped int back in
		__stfiwx( converted, 3 * sizeof(float), buffer );		//store clipped int


		//virtual Cycle D
		(++dst)[0] = copy4;						//store
		copy1 = __rlwimi_volatile( copy1, copy2, 8, 0, 7 );		//merge with adjacent pixel
		copy3 = __lwbrx( 8, buffer );						//load clipped int back in

		//virtual cycle A
		//no store yet						//store
		copy1 = __rlwimi( copy1, copy1, 8, 0, 31 );			//rotate
		copy4 = __lwbrx( 12, buffer );						//load clipped int back in
		copy2 = __rlwimi_volatile( copy2, copy3, 8, 0, 15 );			//merge

		//virtual Cycle B
		(++dst)[0] = copy1;						//store
		copy4 = __rlwimi_volatile( copy4, copy3, 24, 0, 7 );	//merge with adjacent pixel
		copy2 = __rlwimi_volatile( copy2, copy2, 16, 0, 31 );	//rotate

		//virtual Cycle C
		(++dst)[0] = copy2;						//store


		//virtual Cycle D
		(++dst)[0] = copy4;						//store
	}

	//clean up any extras
	dst++;
	while( count-- )
	{
		startingFloat = (++src)[0];				//load the float
		scaled = startingFloat * scale + round;			//scale for clip and add rounding
		converted = __fctiw( scaled );				//convert to int and clip
		__stfiwx( converted, 0, buffer );			//store clipped int
		copy1 = __lwbrx( 0, buffer);					//load clipped int back in
		((signed char*) dst)[0] = copy1 >> 16;
		dst = (signed long*) ((signed char*) dst + 1 );
		((unsigned short*) dst)[0] = copy1;
		dst = (signed long*) ((unsigned short*) dst + 1 );
	}

	//restore the old FPSCR setting
	__asm__ __volatile__ ( "mtfsf 7, %0" : : "f" (oldSetting) );
}


void Float32ToSwapInt32( float *src, signed long *dst, unsigned int count )
{
	register double		scale = 2147483648.0;
	unsigned long	loopCount = count / 4;
	long			buffer[2];
	register float		startingFloat;
	register double scaled;
	register double converted;
	register long		copy;
	register double		oldSetting;

	//Set the FPSCR to round to -Inf mode
	{
		union
		{
			double	d;
			int		i[2];
		}setting;
		register double newSetting;
		
		//Read the the current FPSCR value
		asm volatile ( "mffs %0" : "=f" ( oldSetting ) );
		
		//Store it to the stack
		setting.d = oldSetting;
		
		//Read in the low 32 bits and mask off the last two bits so they are zero 
		//in the integer unit. These two bits set to zero means round to nearest mode.
		//Finally, then store the result back
		setting.i[1] &= 0xFFFFFFFC;
		
		//Load the new FPSCR setting into the FP register file again
		newSetting = setting.d;
		
		//Change the FPSCR to the new setting
		asm volatile( "mtfsf 7, %0" : : "f" (newSetting ) );
	}


	//
	//	The fastest way to do this is to set up a staggered loop that models a 7 stage virtual pipeline:
	//
	//		stage 1:		load the src value
	//		stage 2:		scale it to LONG_MIN...LONG_MAX and add a rounding value to it
	//		stage 3:		convert it to an integer within the FP register
	//		stage 4:		store that to the stack
	//		stage 5:		 (do nothing to let the store complete)
	//		stage 6:		load the high half word from the stack
	//		stage 7:		store it to the destination
	//
	//	We set it up so that at any given time 7 different pieces of data are being worked on at a time.
	//	Because of the do nothing stage, the inner loop had to be unrolled by one, so in actuality, each
	//	inner loop iteration represents two virtual clock cycles that push data through our virtual pipeline.
	//
	//	The reason why this works is that this allows us to break data dependency chains and insert 5 real 
	//	operations in between every virtual pipeline stage. This means 5 instructions between each data 
	//	dependency, which is just enough to keep all of our real pipelines happy. The data flow follows 
	//	standard pipeline diagrams:
	//
	//				stage1	stage2	stage3	stage4	stage5	stage6	stage7
	//	virtual cycle 1:	data1	-	-	-		-		-		-
	//	virtual cycle 2:	data2	data1	-	-		-		-		-
	//	virtual cycle 3:	data3	data2	data1	-		-		-		-
	//	virtual cycle 4:	data4	data3	data2	data1	-		-		-		
	//	virtual cycle 5:	data5	data4	data3	data2	data1	-		-			   
	//	virtual cycle 6:	data6	data5	data4	data3	data2	data1	-	
	//
	//	inner loop:
	//	  virtual cycle A:	data7	data6	data5	data4	data3	data2	data1					  
	//	  virtual cycle B:	data8	data7	data6	data5	data4	data3	data2	
	//
	//	virtual cycle 7		-	dataF	dataE	dataD	dataC	dataB	dataA
	//	virtual cycle 8		-	-	dataF	dataE	dataD	dataC	dataB	
	//	virtual cycle 9		-	-	-	dataF	dataE	dataD	dataC
	//	virtual cycle 10	-	-	-	-		dataF	dataE	dataD  
	//	virtual cycle 11	-	-	-	-		-		dataF	dataE	
	//	virtual cycle 12	-	-	-	-		-		-		dataF						 
	
	if( count >= 6 )
	{
		//virtual cycle 1
		startingFloat = (src++)[0];
		
		//virtual cycle 2
		scaled = startingFloat * scale;
		startingFloat = (src++)[0];
		
		//virtual cycle 3
		converted = __fctiw( scaled );
		scaled = startingFloat * scale;
		startingFloat = (src++)[0];
		
		//virtual cycle 4
		__stfiwx( converted, 0, buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale;
		startingFloat = (src++)[0];
		
		//virtual cycle 5
		__stfiwx( converted, sizeof(float), buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale;
		startingFloat = (src++)[0];
		
		//virtual cycle 6
		copy = buffer[0];
		__stfiwx( converted, 0, buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale;
		startingFloat = (src++)[0];
		
		count -= 6;
		loopCount = count / 2;
		count &= 1;
		while( loopCount-- )
		{
			register float	startingFloat2;
			register double scaled2;
			register double converted2;
			register long	copy2;
			
			//virtual Cycle A
//			  (dst++)[0] = copy;
			__asm__ __volatile__ ( "stwbrx %0, 0, %1" : : "r" (copy), "r" (dst) );
			__asm__ __volatile__ ( "fctiw %0, %1" : "=f" (converted2) : "f" ( scaled ) );
			copy2 = buffer[1];
			__asm__ __volatile__ ( "fmuls %0, %1, %2" : "=f" (scaled2) : "f" ( startingFloat), "f" (scale) );
			__asm__ __volatile__ ( "stfiwx %0, %1, %2" : : "f" (converted), "b%" (sizeof(*buffer)), "r" (buffer) : "memory" );
			startingFloat2 = (src)[0];	src+=2;

			//virtual cycle B
//			  (dst++)[0] = copy2;
			dst+=2;
			__asm__ __volatile__ ( "stwbrx %0, %1, %2" : : "r" (copy2), "r" (-sizeof(dst[0])), "r" (dst) );	 
			__asm__ __volatile__ ( "fctiw %0, %1" : "=f" (converted) : "f" ( scaled2 ) );
			copy = buffer[0];
			__asm__ __volatile__ ( "fmuls %0, %1, %2" : "=f" (scaled) : "f" ( startingFloat2), "f" (scale) );
			__asm__ __volatile__ ( "stfiwx %0, %1, %2" : : "f" (converted2), "b%" (0), "r" (buffer) : "memory" );
			startingFloat = (src)[-1];	
		}
		
		//Virtual Cycle 7
		__asm__ __volatile__ ( "stwbrx %0, 0, %1" : : "r" (copy),  "r" (dst) );	 dst++;
		copy = buffer[1];
		__stfiwx( converted, sizeof(float), buffer );
		converted = __fctiw( scaled );
		scaled = startingFloat * scale;
		
		//Virtual Cycle 8
		__asm__ __volatile__ ( "stwbrx %0, 0, %1" : : "r" (copy),  "r" (dst) );	 dst++;
		copy =	buffer[0];
		__stfiwx( converted, 0, buffer );
		converted = __fctiw( scaled );
	
		//Virtual Cycle 9
		__asm__ __volatile__ ( "stwbrx %0, 0, %1" : : "r" (copy),  "r" (dst) );	 dst++;
		copy = buffer[1];
		__stfiwx( converted, sizeof(float), buffer );
		
		//Virtual Cycle 10
		__asm__ __volatile__ ( "stwbrx %0, 0, %1" : : "r" (copy),  "r" (dst) );	 dst++;
		copy = buffer[0];
	
		//Virtual Cycle 11
		__asm__ __volatile__ ( "stwbrx %0, 0, %1" : : "r" (copy),  "r" (dst) );	 dst++;
		copy = buffer[1];

		//Virtual Cycle 11
		__asm__ __volatile__ ( "stwbrx %0, 0, %1" : : "r" (copy),  "r" (dst) );	 dst++;
	}

	//clean up any extras
	while( count-- )
	{
		double scaled = src[0] * scale;
		double converted = __fctiw( scaled );
		__stfiwx( converted, 0, buffer );
		copy = buffer[0];
		__asm__ __volatile__ ( "stwbrx %0, 0, %1" : : "r" (copy),  "r" (dst) );	 
		src++;
		dst++;
	}
	
	//restore the old FPSCR setting
	__asm__ __volatile__ ( "mtfsf 7, %0" : : "f" (oldSetting) );
}

void Float32ToNativeInt32( float *src, signed long *dst, unsigned int count )
{
	register double		scale = 2147483648.0;
	unsigned long	loopCount;
	register float	startingFloat;
	register double scaled;
	register double converted;
	register double		oldSetting;

	//Set the FPSCR to round to -Inf mode
	{
		union
		{
			double	d;
			int		i[2];
		}setting;
		register double newSetting;
		
		//Read the the current FPSCR value
		asm volatile ( "mffs %0" : "=f" ( oldSetting ) );
		
		//Store it to the stack
		setting.d = oldSetting;
		
		//Read in the low 32 bits and mask off the last two bits so they are zero 
		//in the integer unit. These two bits set to zero means round to -infinity mode.
		//Finally, then store the result back
		setting.i[1] &= 0xFFFFFFFC;
		
		//Load the new FPSCR setting into the FP register file again
		newSetting = setting.d;
		
		//Change the FPSCR to the new setting
		asm volatile( "mtfsf 7, %0" : : "f" (newSetting ) );
	}

	//
	//	The fastest way to do this is to set up a staggered loop that models a 7 stage virtual pipeline:
	//
	//		stage 1:		load the src value
	//		stage 2:		scale it to LONG_MIN...LONG_MAX and add a rounding value to it
	//		stage 3:		convert it to an integer within the FP register
	//		stage 4:		store that to the destination
	//
	//	We set it up so that at any given time 7 different pieces of data are being worked on at a time.
	//	Because of the do nothing stage, the inner loop had to be unrolled by one, so in actuality, each
	//	inner loop iteration represents two virtual clock cycles that push data through our virtual pipeline.
	//
		//	The data flow follows standard pipeline diagrams:
	//
	//				stage1	stage2	stage3	stage4	
	//	virtual cycle 1:	data1	-	-	-	   
	//	virtual cycle 2:	data2	data1	-	-	   
	//	virtual cycle 3:	data3	data2	data1	-			 
	//
	//	inner loop:
	//	  virtual cycle A:	data4	data3	data2	data1					  
	//	  virtual cycle B:	data5	data4	data3	data2	
	//	  ...
	//	virtual cycle 4		-	dataD	dataC	dataB	
	//	virtual cycle 5		-	-		dataD	dataC
	//	virtual cycle 6		-	-	-	dataD  
	
	if( count >= 3 )
	{
		//virtual cycle 1
		startingFloat = (src++)[0];
		
		//virtual cycle 2
		scaled = startingFloat * scale;
		startingFloat = (src++)[0];
		
		//virtual cycle 3
		converted = __fctiw( scaled );
		scaled = startingFloat * scale;
		startingFloat = (src++)[0];
				
		count -= 3;
		loopCount = count / 2;
		count &= 1;
		while( loopCount-- )
		{
			register float	startingFloat2;
			register double scaled2;
			register double converted2;
			//register short	copy2;
			
			//virtual Cycle A
			startingFloat2 = (src)[0];
			__asm__ __volatile__ ( "fmul %0, %1, %2" : "=f" (scaled2) : "f" ( startingFloat), "f" (scale) );
			__asm__ __volatile__ ( "stfiwx %0, %1, %2" : : "f" (converted), "b%" (0), "r" (dst) : "memory" );
			__asm__ __volatile__ ( "fctiw %0, %1" : "=f" (converted2) : "f" ( scaled ) );

			//virtual cycle B
		   startingFloat = (src)[1];	 src+=2; 
			__asm__ __volatile__ ( "fmul %0, %1, %2" : "=f" (scaled) : "f" ( startingFloat2), "f" (scale) );
			__asm__ __volatile__ ( "stfiwx %0, %1, %2" : : "f" (converted2), "b%" (4), "r" (dst) : "memory" );
			__asm__ __volatile__ ( "fctiw %0, %1" : "=f" (converted) : "f" ( scaled2 ) );
			dst+=2;
		}
		
		//Virtual Cycle 4
		__stfiwx( converted, 0, dst++ );
		converted = __fctiw( scaled );
		__asm__ __volatile__ ( "fmul %0, %1, %2" : "=f" (scaled) : "f" ( startingFloat), "f" (scale) );
		
		//Virtual Cycle 5
		__stfiwx( converted, 0, dst++ );
		converted = __fctiw( scaled );
	
		//Virtual Cycle 6
		__stfiwx( converted, 0, dst++ );
	}

	//clean up any extras
	while( count-- )
	{
		double scaled = src[0] * scale;
		double converted = __fctiw( scaled );
		__stfiwx( converted, 0, dst );
		dst++;
		src++;
	}

	//restore the old FPSCR setting
	asm volatile( "mtfsf 7, %0" : : "f" (oldSetting) );
}

#endif


#pragma mark ------------------------ 
#pragma mark ••• Utility Routines
#pragma mark ------------------------ 

UInt32 CalculateOffset (UInt64 nanoseconds, UInt32 sampleRate) {
	return (UInt32)((double)sampleRate * kOneOver1000000000) * nanoseconds;
}

void dBfixed2float(UInt32 indBfixed, float* ioGainPtr) {
    float out, temp, frac;
    // get integer part
    int index = (SInt16)(indBfixed >> 16);		
    // if we're out of bounds, saturate both integer and fraction		
    if (index >= kMaxZeroGain) {
        index = kMaxZeroGain;
        indBfixed = 0;
    } else if (index <= -kMinZeroGain) {
        index = -kMinZeroGain;
        indBfixed = 0;
    }    

    // get fractional part
    frac = ((float)((UInt32)(indBfixed & 0x0000FFFF)))*kOneOver65535;		

    // get the base dB converted value
    out = zeroGaindBConvTable[index + kZeroGaindBConvTableOffset];
    // if we have a fractional part, do linear interpolation on our table
    // this is accurate to about 2 decimal places, which is okay but not great
    if (frac > 0.0f) {
        if (index >= 0) {
            temp = zeroGaindBConvTable[index + kZeroGaindBConvTableOffset + 1];
            out = out + frac*(temp - out);			
        } else {
            temp = zeroGaindBConvTable[index + kZeroGaindBConvTableOffset - 1];
            out = out + frac*(temp - out);
        }
    }

    *ioGainPtr = out;
    
    return;
}

void inputGainConverter(UInt32 inGainIndex, float* ioGainPtr) {
    float out = 1.0f;
    // check bounds		
    if (inGainIndex > (UInt32)(2*kInputGaindBConvTableOffset)) {
        inGainIndex = 2*kInputGaindBConvTableOffset;
    }

    // get the base dB converted value
    out = inputGaindBConvTable[inGainIndex];

    *ioGainPtr = out;
    
    return;
}

void convertToFourDotTwenty(FourDotTwenty* ioFourDotTwenty, float* inFloatPtr)
{
    float 				scale, floatValue;
    SInt32 				result_int;
    
	scale = kFourDotTwentyScaleFactor;
    ioFourDotTwenty->integerAndFraction1 = 0;
    ioFourDotTwenty->fraction2 = 0;
    ioFourDotTwenty->fraction3 = 0;
    floatValue = *inFloatPtr;
	
	if(floatValue > 8.0) // clip to make sure number fits in 4.20 range
        floatValue = 8.0;
    else if(floatValue < -8.0)
        floatValue = -8.0;

	result_int = (SInt32)(floatValue*scale);
	
    ioFourDotTwenty->fraction3 = (UInt8)(result_int & 0x000000FF); // mask off result and store in result
    ioFourDotTwenty->fraction2 = (UInt8)((result_int & 0x0000FF00) >> 8);
    ioFourDotTwenty->integerAndFraction1 = (UInt8)((result_int & 0x00FF0000) >> 16);
	
    return;
}