sbrhfgen.c

nds上大名鼎鼎的DSOrganize 2.8最新源代码。很值得研究。

	p02im.r.hi32 = accBuf[3];

	/* 64-bit accumulators now have 2*FBITS_OUT_QMFA fraction bits
	 * want to scale them down to integers (32-bit signed, Q0)
	 *   with scale factor of 2^n, n >= 0
	 * leave 1 GB for calculating determinant, so take top 30 non-zero bits
	 */
	gbMask  = ((p02re.r.hi32) ^ (p02re.r.hi32 >> 31)) | ((p02im.r.hi32) ^ (p02im.r.hi32 >> 31));
	if (gbMask == 0) {
		s = p02re.r.hi32 >> 31; gbMask  = (p02re.r.lo32 ^ s) - s;
		s = p02im.r.hi32 >> 31; gbMask |= (p02im.r.lo32 ^ s) - s;
		z = 32 + CLZ(gbMask);
	} else {
		gbMask  = FASTABS(p02re.r.hi32) | FASTABS(p02im.r.hi32);
		z = CLZ(gbMask);
	}
	n = 64 - z;	/* number of non-zero bits in bottom of 64-bit word */

	if (n <= 30) {
		loShift = (30 - n);
		*p02reN = p02re.r.lo32 << loShift;	
		*p02imN = p02im.r.lo32 << loShift;
		return -(loShift + 2*FBITS_OUT_QMFA);
	} else if (n < 32 + 30) {
		loShift = (n - 30);
		hiShift = 32 - loShift;
		*p02reN = (p02re.r.hi32 << hiShift) | (p02re.r.lo32 >> loShift);
		*p02imN = (p02im.r.hi32 << hiShift) | (p02im.r.lo32 >> loShift);
		return (loShift - 2*FBITS_OUT_QMFA);
	} else {
		hiShift = n - (32 + 30);
		*p02reN = p02re.r.hi32 >> hiShift;	
		*p02imN = p02im.r.hi32 >> hiShift;
		return (32 - 2*FBITS_OUT_QMFA - hiShift);
	}

	return 0;
}

/**************************************************************************************
 * Function:    CalcLPCoefs
 *
 * Description: calculate linear prediction coefficients for one subband (4.6.18.6.2)
 *
 * Inputs:      buffer of low-freq samples, starting at time index = 0, 
 *                freq index = patch subband
 *              number of guard bits in input sample buffer
 *
 * Outputs:     complex LP coefficients a0re, a0im, a1re, a1im, format = Q29
 *
 * Return:      none
 *
 * Notes:       output coefficients (a0re, a0im, a1re, a1im) clipped to range (-4, 4)
 *              if the comples coefficients have magnitude >= 4.0, they are all
 *                set to 0 (see spec)
 **************************************************************************************/
static void CalcLPCoefs(int *XBuf, int *a0re, int *a0im, int *a1re, int *a1im, int gb)
{
	int zFlag, n1, n2, nd, d, dInv, tre, tim;
	int p01re, p01im, p02re, p02im, p12re, p12im, p11re, p22re;

	/* pre-scale to avoid overflow - probably never happens in practice (see QMFA)
	 *   max bit growth per accumulator = 38*2 = 76 mul-adds (X * X)
	 *   using 64-bit MADD, so if X has n guard bits, X*X has 2n+1 guard bits
	 *   gain 1 extra sign bit per multiply, so ensure ceil(log2(76/2) / 2) = 3 guard bits on inputs
	 */
	if (gb < 3) {
		nd = 3 - gb;
		for (n1 = (NUM_TIME_SLOTS*SAMPLES_PER_SLOT + 6 + 2); n1 != 0; n1--) {
			XBuf[0] >>= nd;	XBuf[1] >>= nd;
			XBuf += (2*64);
		}
		XBuf -= (2*64*(NUM_TIME_SLOTS*SAMPLES_PER_SLOT + 6 + 2));
	}
	
	/* calculate covariance elements */
	n1 = CalcCovariance1(XBuf, &p01re, &p01im, &p12re, &p12im, &p11re, &p22re);
	n2 = CalcCovariance2(XBuf, &p02re, &p02im);

	/* normalize everything to larger power of 2 scalefactor, call it n1 */
	if (n1 < n2) {
		nd = MIN(n2 - n1, 31);
		p01re >>= nd;	p01im >>= nd;
		p12re >>= nd;	p12im >>= nd;
		p11re >>= nd;	p22re >>= nd;
		n1 = n2;
	} else if (n1 > n2) {
		nd = MIN(n1 - n2, 31);
		p02re >>= nd;	p02im >>= nd;
	}

	/* calculate determinant of covariance matrix (at least 1 GB in pXX) */
	d = MULSHIFT32(p12re, p12re) + MULSHIFT32(p12im, p12im);
	d = MULSHIFT32(d, RELAX_COEF) << 1;
	d = MULSHIFT32(p11re, p22re) - d;
	ASSERT(d >= 0);	/* should never be < 0 */

	zFlag = 0;
	*a0re = *a0im = 0;
	*a1re = *a1im = 0;
	if (d > 0) {
		/* input =   Q31  d    = Q(-2*n1 - 32 + nd) = Q31 * 2^(31 + 2*n1 + 32 - nd)
		 * inverse = Q29  dInv = Q29 * 2^(-31 - 2*n1 - 32 + nd) = Q(29 + 31 + 2*n1 + 32 - nd)
		 *
		 * numerator has same Q format as d, since it's sum of normalized squares
		 * so num * inverse = Q(-2*n1 - 32) * Q(29 + 31 + 2*n1 + 32 - nd)
		 *                  = Q(29 + 31 - nd), drop low 32 in MULSHIFT32
		 *                  = Q(29 + 31 - 32 - nd) = Q(28 - nd)
		 */
		nd = CLZ(d) - 1;
		d <<= nd;
		dInv = InvRNormalized(d);

		/* 1 GB in pXX */
		tre = MULSHIFT32(p01re, p12re) - MULSHIFT32(p01im, p12im) - MULSHIFT32(p02re, p11re);
		tre = MULSHIFT32(tre, dInv);
		tim = MULSHIFT32(p01re, p12im) + MULSHIFT32(p01im, p12re) - MULSHIFT32(p02im, p11re);
		tim = MULSHIFT32(tim, dInv);

		/* if d is extremely small, just set coefs to 0 (would have poor precision anyway) */
		if (nd > 28 || (FASTABS(tre) >> (28 - nd)) >= 4 || (FASTABS(tim) >> (28 - nd)) >= 4) {
			zFlag = 1;
		} else {
			*a1re = tre << (FBITS_LPCOEFS - 28 + nd);	/* i.e. convert Q(28 - nd) to Q(29) */
			*a1im = tim << (FBITS_LPCOEFS - 28 + nd);
		}
	}

	if (p11re) {
		/* input =   Q31  p11re = Q(-n1 + nd) = Q31 * 2^(31 + n1 - nd)
		 * inverse = Q29  dInv  = Q29 * 2^(-31 - n1 + nd) = Q(29 + 31 + n1 - nd)
		 *
		 * numerator is Q(-n1 - 3)
		 * so num * inverse = Q(-n1 - 3) * Q(29 + 31 + n1 - nd)
		 *                  = Q(29 + 31 - 3 - nd), drop low 32 in MULSHIFT32
		 *                  = Q(29 + 31 - 3 - 32 - nd) = Q(25 - nd)
		 */
		nd = CLZ(p11re) - 1;	/* assume positive */
		p11re <<= nd;
		dInv = InvRNormalized(p11re);

		/* a1re, a1im = Q29, so scaled by (n1 + 3) */
		tre = (p01re >> 3) + MULSHIFT32(p12re, *a1re) + MULSHIFT32(p12im, *a1im);
		tre = -MULSHIFT32(tre, dInv);
		tim = (p01im >> 3) - MULSHIFT32(p12im, *a1re) + MULSHIFT32(p12re, *a1im);
		tim = -MULSHIFT32(tim, dInv);

		if (nd > 25 || (FASTABS(tre) >> (25 - nd)) >= 4 || (FASTABS(tim) >> (25 - nd)) >= 4) {
			zFlag = 1;
		} else {
			*a0re = tre << (FBITS_LPCOEFS - 25 + nd);	/* i.e. convert Q(25 - nd) to Q(29) */
			*a0im = tim << (FBITS_LPCOEFS - 25 + nd);
		}
	} 

	/* see 4.6.18.6.2 - if magnitude of a0 or a1 >= 4 then a0 = a1 = 0 
	 * i.e. a0re < 4, a0im < 4, a1re < 4, a1im < 4
	 * Q29*Q29 = Q26
	 */
	if (zFlag || MULSHIFT32(*a0re, *a0re) + MULSHIFT32(*a0im, *a0im) >= MAG_16 || MULSHIFT32(*a1re, *a1re) + MULSHIFT32(*a1im, *a1im) >= MAG_16) {
		*a0re = *a0im = 0;
		*a1re = *a1im = 0;
	}

	/* no need to clip - we never changed the XBuf data, just used it to calculate a0 and a1 */
	if (gb < 3) {
		nd = 3 - gb;
		for (n1 = (NUM_TIME_SLOTS*SAMPLES_PER_SLOT + 6 + 2); n1 != 0; n1--) {
			XBuf[0] <<= nd;	XBuf[1] <<= nd;
			XBuf += (2*64);
		}
	}
}

/**************************************************************************************
 * Function:    GenerateHighFreq
 *
 * Description: generate high frequencies with SBR (4.6.18.6)
 *
 * Inputs:      initialized PSInfoSBR struct
 *              initialized SBRGrid struct for this channel
 *              initialized SBRFreq struct for this SCE/CPE block
 *              initialized SBRChan struct for this channel
 *              index of current channel (0 for SCE, 0 or 1 for CPE)
 *
 * Outputs:     new high frequency samples starting at frequency kStart
 *
 * Return:      none
 **************************************************************************************/
void GenerateHighFreq(PSInfoSBR *psi, SBRGrid *sbrGrid, SBRFreq *sbrFreq, SBRChan *sbrChan, int ch)
{
	int band, newBW, c, t, gb, gbMask, gbIdx;
	int currPatch, p, x, k, g, i, iStart, iEnd, bw, bwsq;
	int a0re, a0im, a1re, a1im;
	int x1re, x1im, x2re, x2im;
	int ACCre, ACCim;
	int *XBufLo, *XBufHi;

	/* calculate array of chirp factors */
	for (band = 0; band < sbrFreq->numNoiseFloorBands; band++) {
		c = sbrChan->chirpFact[band];	/* previous (bwArray') */
		newBW = newBWTab[sbrChan->invfMode[0][band]][sbrChan->invfMode[1][band]];

		/* weighted average of new and old (can't overflow - total gain = 1.0) */
		if (newBW < c)
			t = MULSHIFT32(newBW, 0x60000000) + MULSHIFT32(0x20000000, c);	/* new is smaller: 0.75*new + 0.25*old */
		else
			t = MULSHIFT32(newBW, 0x74000000) + MULSHIFT32(0x0c000000, c);	/* new is larger: 0.90625*new + 0.09375*old */
		t <<= 1;

		if (t < 0x02000000)	/* below 0.015625, clip to 0 */
			t = 0;
		if (t > 0x7f800000)	/* clip to 0.99609375 */  
			t = 0x7f800000;

		/* save curr as prev for next time */
		sbrChan->chirpFact[band] = t;
		sbrChan->invfMode[0][band] = sbrChan->invfMode[1][band];
	}

	iStart = sbrGrid->envTimeBorder[0] + HF_ADJ;
	iEnd =   sbrGrid->envTimeBorder[sbrGrid->numEnv] + HF_ADJ;

	/* generate new high freqs from low freqs, patches, and chirp factors */
	k = sbrFreq->kStart;
	g = 0;
	bw = sbrChan->chirpFact[g];
	bwsq = MULSHIFT32(bw, bw) << 1;
	
	gbMask = (sbrChan->gbMask[0] | sbrChan->gbMask[1]);	/* older 32 | newer 8 */
	gb = CLZ(gbMask) - 1;

	for (currPatch = 0; currPatch < sbrFreq->numPatches; currPatch++) {
		for (x = 0; x < sbrFreq->patchNumSubbands[currPatch]; x++) {
			/* map k to corresponding noise floor band */
			if (k >= sbrFreq->freqNoise[g+1]) {
				g++;
				bw = sbrChan->chirpFact[g];		/* Q31 */
				bwsq = MULSHIFT32(bw, bw) << 1;	/* Q31 */
			}
		
			p = sbrFreq->patchStartSubband[currPatch] + x;	/* low QMF band */
			XBufHi = psi->XBuf[iStart][k];
			if (bw) {
				CalcLPCoefs(psi->XBuf[0][p], &a0re, &a0im, &a1re, &a1im, gb);

				a0re = MULSHIFT32(bw, a0re);	/* Q31 * Q29 = Q28 */
				a0im = MULSHIFT32(bw, a0im);
				a1re = MULSHIFT32(bwsq, a1re);
				a1im = MULSHIFT32(bwsq, a1im);

				XBufLo = psi->XBuf[iStart-2][p];

				x2re = XBufLo[0];	/* RE{XBuf[n-2]} */
				x2im = XBufLo[1];	/* IM{XBuf[n-2]} */
				XBufLo += (64*2);

				x1re = XBufLo[0];	/* RE{XBuf[n-1]} */
				x1im = XBufLo[1];	/* IM{XBuf[n-1]} */
				XBufLo += (64*2);

				for (i = iStart; i < iEnd; i++) {
					/* a0re/im, a1re/im are Q28 with at least 1 GB, 
					 *   so the summing for AACre/im is fine (1 GB in, plus 1 from MULSHIFT32) 
					 */
					ACCre = MULSHIFT32(x2re, a1re) - MULSHIFT32(x2im, a1im);
					ACCim = MULSHIFT32(x2re, a1im) + MULSHIFT32(x2im, a1re);
					x2re = x1re;
					x2im = x1im;
					
					ACCre += MULSHIFT32(x1re, a0re) - MULSHIFT32(x1im, a0im);
					ACCim += MULSHIFT32(x1re, a0im) + MULSHIFT32(x1im, a0re);
					x1re = XBufLo[0];	/* RE{XBuf[n]} */
					x1im = XBufLo[1];	/* IM{XBuf[n]} */
					XBufLo += (64*2);

					/* lost 4 fbits when scaling by a0re/im, a1re/im (Q28) */
					CLIP_2N_SHIFT30(ACCre, 4);
					ACCre += x1re;
					CLIP_2N_SHIFT30(ACCim, 4);
					ACCim += x1im;

					XBufHi[0] = ACCre;
					XBufHi[1] = ACCim;
					XBufHi += (64*2);

					/* update guard bit masks */
					gbMask  = FASTABS(ACCre);
					gbMask |= FASTABS(ACCim);
					gbIdx = (i >> 5) & 0x01;	/* 0 if i < 32, 1 if i >= 32 */
					sbrChan->gbMask[gbIdx] |= gbMask;
				}
			} else {
				XBufLo = (int *)psi->XBuf[iStart][p];
				for (i = iStart; i < iEnd; i++) {
					XBufHi[0] = XBufLo[0];
					XBufHi[1] = XBufLo[1];
					XBufLo += (64*2); 
					XBufHi += (64*2);
				}
			}
			k++;	/* high QMF band */
		}
	}
}