/* ***** BEGIN LICENSE BLOCK *****
* Source last modified: $Id: pns.c,v 1.2 2005/03/10 17:01:56 jrecker Exp $
*
* Portions Copyright (c) 1995-2005 RealNetworks, Inc. All Rights Reserved.
*
* The contents of this file, and the files included with this file,
* are subject to the current version of the RealNetworks Public
* Source License (the "RPSL") available at
* http://www.helixcommunity.org/content/rpsl unless you have licensed
* the file under the current version of the RealNetworks Community
* Source License (the "RCSL") available at
* http://www.helixcommunity.org/content/rcsl, in which case the RCSL
* will apply. You may also obtain the license terms directly from
* RealNetworks. You may not use this file except in compliance with
* the RPSL or, if you have a valid RCSL with RealNetworks applicable
* to this file, the RCSL. Please see the applicable RPSL or RCSL for
* the rights, obligations and limitations governing use of the
* contents of the file.
*
* This file is part of the Helix DNA Technology. RealNetworks is the
* developer of the Original Code and owns the copyrights in the
* portions it created.
*
* This file, and the files included with this file, is distributed
* and made available on an 'AS IS' basis, WITHOUT WARRANTY OF ANY
* KIND, EITHER EXPRESS OR IMPLIED, AND REALNETWORKS HEREBY DISCLAIMS
* ALL SUCH WARRANTIES, INCLUDING WITHOUT LIMITATION, ANY WARRANTIES
* OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, QUIET
* ENJOYMENT OR NON-INFRINGEMENT.
*
* Technology Compatibility Kit Test Suite(s) Location:
* http://www.helixcommunity.org/content/tck
*
* Contributor(s):
*
* ***** END LICENSE BLOCK ***** */
/**************************************************************************************
* Fixed-point HE-AAC decoder
* Jon Recker (jrecker@real.com)
* February 2005
*
* pns.c - perceptual noise substitution
**************************************************************************************/
#include "coder.h"
#include "assembly.h"
/**************************************************************************************
* Function: Get32BitVal
*
* Description: generate 32-bit unsigned random number
*
* Inputs: last number calculated (seed, first time through)
*
* Outputs: new number, saved in *last
*
* Return: 32-bit number, uniformly distributed between [0, 2^32)
*
* Notes: uses simple linear congruential generator
**************************************************************************************/
static unsigned int Get32BitVal(unsigned int *last)
{
unsigned int r = *last;
/* use same coefs as MPEG reference code (classic LCG)
* use unsigned multiply to force reliable wraparound behavior in C (mod 2^32)
*/
r = (1664525U * r) + 1013904223U;
*last = r;
return r;
}
#define NUM_ITER_INVSQRT 4
#define X0_COEF_2 0xc0000000 /* Q29: -2.0 */
#define X0_OFF_2 0x60000000 /* Q29: 3.0 */
#define Q26_3 0x0c000000 /* Q26: 3.0 */
/**************************************************************************************
* Function: InvRootR
*
* Description: use Newton's method to solve for x = 1/sqrt(r)
*
* Inputs: r in Q30 format, range = [0.25, 1] (normalize inputs to this range)
*
* Outputs: none
*
* Return: x = Q29, range = (1, 2)
*
* Notes: guaranteed to converge and not overflow for any r in this range
*
* xn+1 = xn - f(xn)/f'(xn)
* f(x) = 1/sqrt(r) - x = 0 (find root)
* = 1/x^2 - r
* f'(x) = -2/x^3
*
* so xn+1 = xn/2 * (3 - r*xn^2)
*
* NUM_ITER_INVSQRT = 3, maxDiff = 1.3747e-02
* NUM_ITER_INVSQRT = 4, maxDiff = 3.9832e-04
**************************************************************************************/
static int InvRootR(int r)
{
int i, xn, t;
/* use linear equation for initial guess
* x0 = -2*r + 3 (so x0 always >= correct answer in range [0.25, 1))
* xn = Q29 (at every step)
*/
xn = (MULSHIFT32(r, X0_COEF_2) << 2) + X0_OFF_2;
for (i = 0; i < NUM_ITER_INVSQRT; i++) {
t = MULSHIFT32(xn, xn); /* Q26 = Q29*Q29 */
t = Q26_3 - (MULSHIFT32(r, t) << 2); /* Q26 = Q26 - (Q31*Q26 << 1) */
xn = MULSHIFT32(xn, t) << (6 - 1); /* Q29 = (Q29*Q26 << 6), and -1 for division by 2 */
}
/* clip to range (1.0, 2.0)
* (because of rounding, this can converge to xn slightly > 2.0 when r is near 0.25)
*/
if (xn >> 30)
xn = (1 << 30) - 1;
return xn;
}
/**************************************************************************************
* Function: ScaleNoiseVector
*
* Description: apply scaling to vector of noise coefficients for one scalefactor band
*
* Inputs: unscaled coefficients
* number of coefficients in vector (one scalefactor band of coefs)
* scalefactor for this band (i.e. noise energy)
*
* Outputs: nVals coefficients in Q(FBITS_OUT_DQ_OFF)
*
* Return: guard bit mask (OR of abs value of all noise coefs)
**************************************************************************************/
static int ScaleNoiseVector(int *coef, int nVals, int sf)
{
/* pow(2, i/4.0) for i = [0,1,2,3], format = Q30 */
static const int pow14[4] PROGMEM = {
0x40000000, 0x4c1bf829, 0x5a82799a, 0x6ba27e65
};
int i, c, spec, energy, sq, scalef, scalei, invSqrtEnergy, z, gbMask;
energy = 0;
for (i = 0; i < nVals; i++) {
spec = coef[i];
/* max nVals = max SFB width = 96, so energy can gain < 2^7 bits in accumulation */
sq = (spec * spec) >> 8; /* spec*spec range = (-2^30, 2^30) */
energy += sq;
}
/* unless nVals == 1 (or the number generator is broken...), this should not happen */
if (energy == 0)
return 0; /* coef[i] must = 0 for i = [0, nVals-1], so gbMask = 0 */
/* pow(2, sf/4) * pow(2, FBITS_OUT_DQ_OFF) */
scalef = pow14[sf & 0x3];
scalei = (sf >> 2) + FBITS_OUT_DQ_OFF;
/* energy has implied factor of 2^-8 since we shifted the accumulator
* normalize energy to range [0.25, 1.0), calculate 1/sqrt(1), and denormalize
* i.e. divide input by 2^(30-z) and convert to Q30
* output of 1/sqrt(i) now has extra factor of 2^((30-z)/2)
* for energy > 0, z is an even number between 0 and 28
* final scaling of invSqrtEnergy:
* 2^(15 - z/2) to compensate for implicit 2^(30-z) factor in input
* +4 to compensate for implicit 2^-8 factor in input
*/
z = CLZ(energy) - 2; /* energy has at least 2 leading zeros (see acc loop) */
z &= 0xfffffffe; /* force even */
invSqrtEnergy = InvRootR(energy << z); /* energy << z must be in range [0x10000000, 0x40000000] */
scalei -= (15 - z/2 + 4); /* nInt = 1/sqrt(energy) in Q29 */
/* normalize for final scaling */
z = CLZ(invSqrtEnergy) - 1;
invSqrtEnergy <<= z;
scalei -= (z - 3 - 2); /* -2 for scalef, z-3 for invSqrtEnergy */
scalef = MULSHIFT32(scalef, invSqrtEnergy); /* scalef (input) = Q30, invSqrtEnergy = Q29 * 2^z */
gbMask = 0;
if (scalei < 0) {
scalei = -scalei;
if (scalei > 31)
scalei = 31;
for (i = 0; i < nVals; i++) {
c = MULSHIFT32(coef[i], scalef) >> scalei;
gbMask |= FASTABS(c);
coef[i] = c;
}
} else {
/* for scalei <= 16, no clipping possible (coef[i] is < 2^15 before scaling)
* for scalei > 16, just saturate exponent (rare)
* scalef is close to full-scale (since we normalized invSqrtEnergy)
* remember, we are just producing noise here
*/
if (scalei > 16)
scalei = 16;
for (i = 0; i < nVals; i++) {
c = MULSHIFT32(coef[i] << scalei, scalef);
coef[i] = c;
gbMask |= FASTABS(c);
}
}
return gbMask;
}
/**************************************************************************************
* Function: GenerateNoiseVector
*
* Description: create vector of noise coefficients for one scalefactor band
*
* Inputs: seed for number generator
* number of coefficients to generate
*
* Outputs: buffer of nVals coefficients, range = [-2^15, 2^15)
* updated seed for number generator
*
* Return: none
**************************************************************************************/
static void GenerateNoiseVector(int *coef, int *last, int nVals)
{
int i;
for (i = 0; i < nVals; i++)
coef[i] = ((signed int)Get32BitVal((unsigned int *)last)) >> 16;
}
/**************************************************************************************
* Function: CopyNoiseVector
*
* Description: copy vector of noise coefficients for one scalefactor band from L to R
*
* Inputs: buffer of left coefficients
* number of coefficients to copy
*
* Outputs: buffer of right coefficients
*
* Return: none
**************************************************************************************/
static void CopyNoiseVector(int *coefL, int *coefR, int nVals)
{
int i;
for (i = 0; i < nVals; i++)
coefR[i] = coefL[i];
}
/**************************************************************************************
* Function: PNS
*
* Description: apply perceptual noise substitution, if enabled (MPEG-4 only)
*
* Inputs: valid AACDecInfo struct
* index of current channel
*
* Outputs: shaped noise in scalefactor bands where PNS is active
* updated minimum guard bit count for this channel
*
* Return: 0 if successful, -1 if error
**************************************************************************************/
int PNS(AACDecInfo *aacDecInfo, int ch)
{
int gp, sfb, win, width, nSamps, gb, gbMask;
int *coef;
const /*short*/ int *sfbTab;
unsigned char *sfbCodeBook;
short *scaleFactors;
int msMaskOffset, checkCorr, genNew;
unsigned char msMask;
unsigned char *msMaskPtr;
PSInfoBase *psi;
ICSInfo *icsInfo;
/* validate pointers */
if (!aacDecInfo || !aacDecInfo->psInfoBase)
return -1;
psi = (PSInfoBase *)(aacDecInfo->psInfoBase);
icsInfo = (ch == 1 && psi->commonWin == 1) ? &(psi->icsInfo[0]) : &(psi->icsInfo[ch]);
if (!psi->pnsUsed[ch])
return 0;
if (icsInfo->winSequence == 2) {
sfbTab = sfBandTabShort + sfBandTabShortOffset[psi->sampRateIdx];
nSamps = NSAMPS_SHORT;
} else {
sfbTab = sfBandTabLong + sfBandTabLongOffset[psi->sampRateIdx];
nSamps = NSAMPS_LONG;
}
coef = psi->coef[ch];
sfbCodeBook = psi->sfbCodeBook[ch];
scaleFactors = psi->scaleFactors[ch];
checkCorr = (aacDecInfo->currBlockID == AAC_ID_CPE && psi->commonWin == 1 ? 1 : 0);
gbMask = 0;
for (gp = 0; gp < icsInfo->numWinGroup; gp++) {
for (win = 0; win < icsInfo->winGroupLen[gp]; win++) {
msMaskPtr = psi->msMaskBits + ((gp*icsInfo->maxSFB) >> 3);
msMaskOffset = ((gp*icsInfo->maxSFB) & 0x07);
msMask = (*msMaskPtr++) >> msMaskOffset;
for (sfb = 0; sfb < icsInfo->maxSFB; sfb++) {
width = sfbTab[sfb+1] - sfbTab[sfb];
if (sfbCodeBook[sfb] == 13) {
if (ch == 0) {
/* generate new vector, copy into ch 1 if it's possible that the channels will be correlated
* if ch 1 has PNS enabled for this SFB but it's uncorrelated (i.e. ms_used == 0),
* the copied values will be overwritten when we process ch 1
*/
GenerateNoiseVector(coef, &psi->pnsLastVal, width);
if (checkCorr && psi->sfbCodeBook[1][gp*icsInfo->maxSFB + sfb] == 13)
CopyNoiseVector(coef, psi->coef[1] + (coef - psi->coef[0]), width);
} else {
/* generate new vector if no correlation between channels */
genNew = 1;
if (checkCorr && psi->sfbCodeBook[0][gp*icsInfo->maxSFB + sfb] == 13) {
if ( (psi->msMaskPresent == 1 && (msMask & 0x01)) || psi->msMaskPresent == 2 )
genNew = 0;
}
if (genNew)
GenerateNoiseVector(coef, &psi->pnsLastVal, width);
}
gbMask |= ScaleNoiseVector(coef, width, psi->scaleFactors[ch][gp*icsInfo->maxSFB + sfb]);
}
coef += width;
/* get next mask bit (should be branchless on ARM) */
msMask >>= 1;
if (++msMaskOffset == 8) {
msMask = *msMaskPtr++;
msMaskOffset = 0;
}
}
coef += (nSamps - sfbTab[icsInfo->maxSFB]);
}
sfbCodeBook += icsInfo->maxSFB;
scaleFactors += icsInfo->maxSFB;
}
/* update guard bit count if necessary */
gb = CLZ(gbMask) - 1;
if (psi->gbCurrent[ch] > gb)
psi->gbCurrent[ch] = gb;
return 0;
}