psymodel.c

音频编码

/*
 *	psymodel.c
 *
 *	Copyright (c) 1999-2000 Mark Taylor
 *	Copyright (c) 2001-2002 Naoki Shibata
 *	Copyright (c) 2000-2003 Takehiro Tominaga
 *	Copyright (c) 2000-2005 Robert Hegemann
 *	Copyright (c) 2000-2005 Gabriel Bouvigne
 *	Copyright (c) 2000-2005 Alexander Leidinger
 *
 * This library is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Library General Public
 * License as published by the Free Software Foundation; either
 * version 2 of the License, or (at your option) any later version.
 *
 * This library is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.	 See the GNU
 * Library General Public License for more details.
 *
 * You should have received a copy of the GNU Library General Public
 * License along with this library; if not, write to the
 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
 * Boston, MA 02111-1307, USA.
 */

/* $Id: psymodel.c,v 1.142.2.1 2005/11/20 14:08:25 bouvigne Exp $ */


/*
PSYCHO ACOUSTICS


This routine computes the psycho acoustics, delayed by one granule.  

Input: buffer of PCM data (1024 samples).  

This window should be centered over the 576 sample granule window.
The routine will compute the psycho acoustics for
this granule, but return the psycho acoustics computed
for the *previous* granule.  This is because the block
type of the previous granule can only be determined
after we have computed the psycho acoustics for the following
granule.  

Output:  maskings and energies for each scalefactor band.
block type, PE, and some correlation measures.  
The PE is used by CBR modes to determine if extra bits
from the bit reservoir should be used.  The correlation
measures are used to determine mid/side or regular stereo.
*/
/*
Notation:

barks:  a non-linear frequency scale.  Mapping from frequency to
        barks is given by freq2bark()

scalefactor bands: The spectrum (frequencies) are broken into 
                   SBMAX "scalefactor bands".  Thes bands
                   are determined by the MPEG ISO spec.  In
                   the noise shaping/quantization code, we allocate
                   bits among the partition bands to achieve the
                   best possible quality

partition bands:   The spectrum is also broken into about
                   64 "partition bands".  Each partition 
                   band is about .34 barks wide.  There are about 2-5
                   partition bands for each scalefactor band.

LAME computes all psycho acoustic information for each partition
band.  Then at the end of the computations, this information
is mapped to scalefactor bands.  The energy in each scalefactor
band is taken as the sum of the energy in all partition bands
which overlap the scalefactor band.  The maskings can be computed
in the same way (and thus represent the average masking in that band)
or by taking the minmum value multiplied by the number of
partition bands used (which represents a minimum masking in that band).
*/
/*
The general outline is as follows:

1. compute the energy in each partition band
2. compute the tonality in each partition band
3. compute the strength of each partion band "masker"
4. compute the masking (via the spreading function applied to each masker)
5. Modifications for mid/side masking.  

Each partition band is considiered a "masker".  The strength
of the i'th masker in band j is given by:

    s3(bark(i)-bark(j))*strength(i)

The strength of the masker is a function of the energy and tonality.
The more tonal, the less masking.  LAME uses a simple linear formula
(controlled by NMT and TMN) which says the strength is given by the
energy divided by a linear function of the tonality.
*/
/*
s3() is the "spreading function".  It is given by a formula
determined via listening tests.  

The total masking in the j'th partition band is the sum over
all maskings i.  It is thus given by the convolution of
the strength with s3(), the "spreading function."

masking(j) = sum_over_i  s3(i-j)*strength(i)  = s3 o strength

where "o" = convolution operator.  s3 is given by a formula determined
via listening tests.  It is normalized so that s3 o 1 = 1.

Note: instead of a simple convolution, LAME also has the
option of using "additive masking"

The most critical part is step 2, computing the tonality of each
partition band.  LAME has two tonality estimators.  The first
is based on the ISO spec, and measures how predictiable the
signal is over time.  The more predictable, the more tonal.
The second measure is based on looking at the spectrum of
a single granule.  The more peaky the spectrum, the more
tonal.  By most indications, the latter approach is better.

Finally, in step 5, the maskings for the mid and side
channel are possibly increased.  Under certain circumstances,
noise in the mid & side channels is assumed to also
be masked by strong maskers in the L or R channels.


Other data computed by the psy-model:

ms_ratio        side-channel / mid-channel masking ratio (for previous granule)
ms_ratio_next   side-channel / mid-channel masking ratio for this granule

percep_entropy[2]     L and R values (prev granule) of PE - A measure of how 
                      much pre-echo is in the previous granule
percep_entropy_MS[2]  mid and side channel values (prev granule) of percep_entropy
energy[4]             L,R,M,S energy in each channel, prev granule
blocktype_d[2]        block type to use for previous granule
*/




#ifdef HAVE_CONFIG_H
# include <config.h>
#endif

#include "util.h"
#include "encoder.h"
#include "psymodel.h"
#include "l3side.h"
#include <assert.h>
#include "tables.h"
#include "fft.h"
#include "machine.h"

#ifdef WITH_DMALLOC
#include <dmalloc.h>
#endif

#define NSFIRLEN 21

#ifdef M_LN10
#define		LN_TO_LOG10		(M_LN10/10)
#else
#define         LN_TO_LOG10             0.2302585093
#endif

#ifdef NON_LINEAR_PSY

static const float non_linear_psy_constant = .3;

#define NON_LINEAR_SCALE_ITEM(x)   pow((x), non_linear_psy_constant)
#define NON_LINEAR_SCALE_SUM(x)    pow((x), 1/non_linear_psy_constant)

#if 0
#define NON_LINEAR_SCALE_ENERGY(x) pow(10, (x)/10)
#else
#define NON_LINEAR_SCALE_ENERGY(x) (x)
#endif

#else

#define NON_LINEAR_SCALE_ITEM(x)   (x)
#define NON_LINEAR_SCALE_SUM(x)    (x)
#define NON_LINEAR_SCALE_ENERGY(x) (x)

#endif


/*
   L3psycho_anal.  Compute psycho acoustics.

   Data returned to the calling program must be delayed by one 
   granule. 

   This is done in two places.  
   If we do not need to know the blocktype, the copying
   can be done here at the top of the program: we copy the data for
   the last granule (computed during the last call) before it is
   overwritten with the new data.  It looks like this:
  
   0. static psymodel_data 
   1. calling_program_data = psymodel_data
   2. compute psymodel_data
    
   For data which needs to know the blocktype, the copying must be
   done at the end of this loop, and the old values must be saved:
   
   0. static psymodel_data_old 
   1. compute psymodel_data
   2. compute possible block type of this granule
   3. compute final block type of previous granule based on #2.
   4. calling_program_data = psymodel_data_old
   5. psymodel_data_old = psymodel_data
*/





/* psycho_loudness_approx
   jd - 2001 mar 12
in:  energy   - BLKSIZE/2 elements of frequency magnitudes ^ 2
     gfp      - uses out_samplerate, ATHtype (also needed for ATHformula)
returns: loudness^2 approximation, a positive value roughly tuned for a value
         of 1.0 for signals near clipping.
notes:   When calibrated, feeding this function binary white noise at sample
         values +32767 or -32768 should return values that approach 3.
         ATHformula is used to approximate an equal loudness curve.
future:  Data indicates that the shape of the equal loudness curve varies
         with intensity.  This function might be improved by using an equal
         loudness curve shaped for typical playback levels (instead of the
         ATH, that is shaped for the threshold).  A flexible realization might
         simply bend the existing ATH curve to achieve the desired shape.
         However, the potential gain may not be enough to justify an effort.
*/
static FLOAT
psycho_loudness_approx( FLOAT *energy, lame_internal_flags *gfc )
{
    int i;
    FLOAT loudness_power;

    loudness_power = 0.0;
    /* apply weights to power in freq. bands*/
    for( i = 0; i < BLKSIZE/2; ++i )
	loudness_power += energy[i] * gfc->ATH->eql_w[i];
    loudness_power *= VO_SCALE;

    return loudness_power;
}

static void
compute_ffts(
    lame_global_flags *gfp,
    FLOAT fftenergy[HBLKSIZE],
    FLOAT (*fftenergy_s)[HBLKSIZE_s],
    FLOAT (*wsamp_l)[BLKSIZE],
    FLOAT (*wsamp_s)[3][BLKSIZE_s],
    int gr_out,
    int chn,
    const sample_t *buffer[2]
    )
{
    int b, j;
    lame_internal_flags *gfc=gfp->internal_flags;
    if (chn<2) {
	fft_long ( gfc, *wsamp_l, chn, buffer);
	fft_short( gfc, *wsamp_s, chn, buffer);
    }
    /* FFT data for mid and side channel is derived from L & R */
    else if (chn == 2) {
	for (j = BLKSIZE-1; j >=0 ; --j) {
	    FLOAT l = wsamp_l[0][j];
	    FLOAT r = wsamp_l[1][j];
	    wsamp_l[0][j] = (l+r)*(FLOAT)(SQRT2*0.5);
	    wsamp_l[1][j] = (l-r)*(FLOAT)(SQRT2*0.5);
	}
	for (b = 2; b >= 0; --b) {
	    for (j = BLKSIZE_s-1; j >= 0 ; --j) {
		FLOAT l = wsamp_s[0][b][j];
		FLOAT r = wsamp_s[1][b][j];
		wsamp_s[0][b][j] = (l+r)*(FLOAT)(SQRT2*0.5);
		wsamp_s[1][b][j] = (l-r)*(FLOAT)(SQRT2*0.5);
	    }
	}
    }
	
    /*********************************************************************
     *  compute energies
     *********************************************************************/
    fftenergy[0]  = NON_LINEAR_SCALE_ENERGY(wsamp_l[0][0]);
    fftenergy[0] *= fftenergy[0];

    for (j=BLKSIZE/2-1; j >= 0; --j) {
	FLOAT re = (*wsamp_l)[BLKSIZE/2-j];
	FLOAT im = (*wsamp_l)[BLKSIZE/2+j];
	fftenergy[BLKSIZE/2-j] = NON_LINEAR_SCALE_ENERGY((re * re + im * im) * 0.5f);
    }
    for (b = 2; b >= 0; --b) {
	fftenergy_s[b][0]  = (*wsamp_s)[b][0];
	fftenergy_s[b][0] *=  fftenergy_s [b][0];
	for (j=BLKSIZE_s/2-1; j >= 0; --j) {
	    FLOAT re = (*wsamp_s)[b][BLKSIZE_s/2-j];
	    FLOAT im = (*wsamp_s)[b][BLKSIZE_s/2+j];
	    fftenergy_s[b][BLKSIZE_s/2-j] = NON_LINEAR_SCALE_ENERGY((re * re + im * im) * 0.5f);
	}
    }
    /* total energy */
    {FLOAT totalenergy=0.0;
    for (j=11;j < HBLKSIZE; j++)
	totalenergy += fftenergy[j];

    gfc->tot_ener[chn] = totalenergy;
    }

#if defined(HAVE_GTK)
    if (gfp->analysis) {
	for (j=0; j<HBLKSIZE ; j++) {
	    gfc->pinfo->energy[gr_out][chn][j]=gfc->energy_save[chn][j];
	    gfc->energy_save[chn][j]=fftenergy[j];
	}
 	gfc->pinfo->pe[gr_out][chn]=gfc->pe[chn];
    }
#endif
    /*********************************************************************
     * compute loudness approximation (used for ATH auto-level adjustment) 
     *********************************************************************/
    if (gfp->athaa_loudapprox == 2 && chn < 2) {/*no loudness for mid/side ch*/
	gfc->loudness_sq[gr_out][chn] = gfc->loudness_sq_save[chn];
	gfc->loudness_sq_save[chn]
	    = psycho_loudness_approx(fftenergy, gfc);
    }
}

/*************************************************************** 
 * compute interchannel masking effects
 ***************************************************************/
static void
calc_interchannel_masking(
    lame_global_flags * gfp,
    FLOAT ratio
    )
{
    lame_internal_flags *gfc=gfp->internal_flags;
    int sb, sblock;
    FLOAT l, r;
    if (gfc->channels_out > 1){
        for ( sb = 0; sb < SBMAX_l; sb++ ) {
	        l = gfc->thm[0].l[sb];
	        r = gfc->thm[1].l[sb];
	        gfc->thm[0].l[sb] += r*ratio;
	        gfc->thm[1].l[sb] += l*ratio;
        }
        for ( sb = 0; sb < SBMAX_s; sb++ ) {
	        for ( sblock = 0; sblock < 3; sblock++ ) {
	            l = gfc->thm[0].s[sb][sblock];
	            r = gfc->thm[1].s[sb][sblock];
	            gfc->thm[0].s[sb][sblock] += r*ratio;
	            gfc->thm[1].s[sb][sblock] += l*ratio;
	        }
        }
    }
}



/*************************************************************** 
 * compute M/S thresholds from Johnston & Ferreira 1992 ICASSP paper
 ***************************************************************/
static void
msfix1(
    lame_internal_flags *gfc
    )
{
    int sb, sblock;
    FLOAT rside,rmid,mld;
    for ( sb = 0; sb < SBMAX_l; sb++ ) {
	/* use this fix if L & R masking differs by 2db or less */
	/* if db = 10*log10(x2/x1) < 2 */
	/* if (x2 < 1.58*x1) { */
	if (gfc->thm[0].l[sb] > 1.58*gfc->thm[1].l[sb]
	 || gfc->thm[1].l[sb] > 1.58*gfc->thm[0].l[sb])
	    continue;

	mld = gfc->mld_l[sb]*gfc->en[3].l[sb];
	rmid = Max(gfc->thm[2].l[sb], Min(gfc->thm[3].l[sb],mld));

	mld = gfc->mld_l[sb]*gfc->en[2].l[sb];
	rside = Max(gfc->thm[3].l[sb], Min(gfc->thm[2].l[sb],mld));
	gfc->thm[2].l[sb]=rmid;
	gfc->thm[3].l[sb]=rside;
    }

    for ( sb = 0; sb < SBMAX_s; sb++ ) {
	for ( sblock = 0; sblock < 3; sblock++ ) {
	    if (gfc->thm[0].s[sb][sblock] > 1.58*gfc->thm[1].s[sb][sblock]
	     || gfc->thm[1].s[sb][sblock] > 1.58*gfc->thm[0].s[sb][sblock])
		continue;

	    mld = gfc->mld_s[sb]*gfc->en[3].s[sb][sblock];
	    rmid = Max(gfc->thm[2].s[sb][sblock],
		       Min(gfc->thm[3].s[sb][sblock],mld));

	    mld = gfc->mld_s[sb]*gfc->en[2].s[sb][sblock];
	    rside = Max(gfc->thm[3].s[sb][sblock],
			Min(gfc->thm[2].s[sb][sblock],mld));

	    gfc->thm[2].s[sb][sblock]=rmid;
	    gfc->thm[3].s[sb][sblock]=rside;
	}
    }
}

/*************************************************************** 
 * Adjust M/S maskings if user set "msfix"
 ***************************************************************/
/* Naoki Shibata 2000 */
static void
ns_msfix(
    lame_internal_flags *gfc,
    FLOAT msfix,
    FLOAT athadjust
    )
{
    int sb, sblock;
    FLOAT msfix2 = msfix;
    FLOAT athlower = pow(10, athadjust);

    msfix *= 2.0;
    msfix2 *= 2.0;
    for ( sb = 0; sb < SBMAX_l; sb++ ) {
	FLOAT thmLR,thmM,thmS,ath;
	ath  = (gfc->ATH->cb[gfc->bm_l[sb]])*athlower;
	thmLR = Min(Max(gfc->thm[0].l[sb],ath), Max(gfc->thm[1].l[sb],ath));
	thmM = Max(gfc->thm[2].l[sb],ath);
	thmS = Max(gfc->thm[3].l[sb],ath);

	if (thmLR*msfix < thmM+thmS) {
	    FLOAT f = thmLR * msfix2 / (thmM+thmS);
	    thmM *= f;
	    thmS *= f;
	}
	gfc->thm[2].l[sb] = Min(thmM,gfc->thm[2].l[sb]);
	gfc->thm[3].l[sb] = Min(thmS,gfc->thm[3].l[sb]);
    }

    athlower *= ((FLOAT)BLKSIZE_s / BLKSIZE);
    for ( sb = 0; sb < SBMAX_s; sb++ ) {
	for ( sblock = 0; sblock < 3; sblock++ ) {
	    FLOAT thmLR,thmM,thmS,ath;
	    ath  = (gfc->ATH->cb[gfc->bm_s[sb]])*athlower;
	    thmLR = Min(Max(gfc->thm[0].s[sb][sblock],ath),
			Max(gfc->thm[1].s[sb][sblock],ath));
	    thmM = Max(gfc->thm[2].s[sb][sblock],ath);
	    thmS = Max(gfc->thm[3].s[sb][sblock],ath);

	    if (thmLR*msfix < thmM+thmS) {
		FLOAT f = thmLR*msfix / (thmM+thmS);
		thmM *= f;
		thmS *= f;
	    }
	    gfc->thm[2].s[sb][sblock] = Min(gfc->thm[2].s[sb][sblock],thmM);
	    gfc->thm[3].s[sb][sblock] = Min(gfc->thm[3].s[sb][sblock],thmS);
	}
    }
}

/* longblock threshold calculation (part 2) */
static void convert_partition2scalefac_l(
    lame_internal_flags *gfc,
    FLOAT *eb,
    FLOAT *thr,
    int chn
    )
{
    FLOAT enn, thmm;
    int sb, b;
    enn = thmm = 0.0;
    sb = b = 0;
    for (;;) {
	while (b < gfc->bo_l[sb]) {
	    enn  += eb[b];
	    thmm += thr[b];
	    b++;
	}

	if (sb == SBMAX_l - 1)
	    break;

    assert( enn >= 0 );
    assert( thmm >= 0 );
	gfc->en [chn].l[sb] = enn  + 0.5 * eb [b];
	gfc->thm[chn].l[sb] = thmm + 0.5 * thr[b];

	enn  = 0.5 *  eb[b];
	thmm = 0.5 * thr[b];
    assert( enn >= 0 );
    assert( thmm >= 0 );
	b++;
	sb++;
    }

    gfc->en [chn].l[SBMAX_l-1] = enn;
    gfc->thm[chn].l[SBMAX_l-1] = thmm;
}

static void
compute_masking_s(
    lame_internal_flags *gfc,
    FLOAT (*fftenergy_s)[HBLKSIZE_s],
    FLOAT *eb,
    FLOAT *thr,
    int chn,
    int sblock,
    FLOAT athlower
    )
{
    int j, b;
    athlower *= ((FLOAT)BLKSIZE_s / BLKSIZE);
    for (j = b = 0; b < gfc->npart_s; b++) {
	FLOAT ecb = fftenergy_s[sblock][j++];
	int kk = gfc->numlines_s[b];