/*

 *      psymodel.c

 *

 *      Copyright (c) 1999-2000 Mark Taylor

 *      Copyright (c) 2001-2002 Naoki Shibata

 *      Copyright (c) 2000-2003 Takehiro Tominaga

 *      Copyright (c) 2000-2012 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 1.216 2017/09/06 19:38:23 aleidinger 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 <float.h>

#include "lame.h"

#include "machine.h"

#include "encoder.h"

#include "util.h"

#include "psymodel.h"

#include "lame_global_flags.h"

#include "fft.h"

#include "lame-analysis.h"

#define NSFIRLEN 21

#ifdef M_LN10

#define  LN_TO_LOG10  (M_LN10/10)

#else

#define  LN_TO_LOG10  0.2302585093

#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 const *energy, FLOAT const *eql_w)

{

    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] * eql_w[i];

    loudness_power *= VO_SCALE;

    return loudness_power;

}

/* mask_add optimization */

/* init the limit values used to avoid computing log in mask_add when it is not necessary */

/* For example, with i = 10*log10(m2/m1)/10*16         (= log10(m2/m1)*16)

 *

 * abs(i)>8 is equivalent (as i is an integer) to

 * abs(i)>=9

 * i>=9 || i<=-9

 * equivalent to (as i is the biggest integer smaller than log10(m2/m1)*16 

 * or the smallest integer bigger than log10(m2/m1)*16 depending on the sign of log10(m2/m1)*16)

 * log10(m2/m1)>=9/16 || log10(m2/m1)<=-9/16

 * exp10 is strictly increasing thus this is equivalent to

 * m2/m1 >= 10^(9/16) || m2/m1<=10^(-9/16) which are comparisons to constants

 */

#define I1LIMIT 8       /* as in if(i>8)  */

#define I2LIMIT 23      /* as in if(i>24) -> changed 23 */

#define MLIMIT  15      /* as in if(m<15) */

/* pow(10, (I1LIMIT + 1) / 16.0); */

static const FLOAT ma_max_i1 = 3.6517412725483771;

/* pow(10, (I2LIMIT + 1) / 16.0); */

static const FLOAT ma_max_i2 = 31.622776601683793;

/* pow(10, (MLIMIT) / 10.0); */

static const FLOAT ma_max_m  = 31.622776601683793;

    /*This is the masking table:

       According to tonality, values are going from 0dB (TMN)

       to 9.3dB (NMT).

       After additive masking computation, 8dB are added, so

       final values are going from 8dB to 17.3dB

     */

static const FLOAT tab[] = {

    1.0 /*pow(10, -0) */ ,

    0.79433 /*pow(10, -0.1) */ ,

    0.63096 /*pow(10, -0.2) */ ,

    0.63096 /*pow(10, -0.2) */ ,

    0.63096 /*pow(10, -0.2) */ ,

    0.63096 /*pow(10, -0.2) */ ,

    0.63096 /*pow(10, -0.2) */ ,

    0.25119 /*pow(10, -0.6) */ ,

    0.11749             /*pow(10, -0.93) */

};

static const int tab_mask_add_delta[] = { 2, 2, 2, 1, 1, 1, 0, 0, -1 };

#define STATIC_ASSERT_EQUAL_DIMENSION(A,B) enum{static_assert_##A=1/((dimension_of(A) == dimension_of(B))?1:0)}

inline static int

mask_add_delta(int i)

{

    STATIC_ASSERT_EQUAL_DIMENSION(tab_mask_add_delta,tab);

    assert(i < (int)dimension_of(tab));

    return tab_mask_add_delta[i];

}

static void

init_mask_add_max_values(void)

{

#ifndef NDEBUG

    FLOAT const _ma_max_i1 = pow(10, (I1LIMIT + 1) / 16.0);

    FLOAT const _ma_max_i2 = pow(10, (I2LIMIT + 1) / 16.0);

    FLOAT const _ma_max_m = pow(10, (MLIMIT) / 10.0);

    assert(fabs(ma_max_i1 - _ma_max_i1) <= FLT_EPSILON);

    assert(fabs(ma_max_i2 - _ma_max_i2) <= FLT_EPSILON);

    assert(fabs(ma_max_m  - _ma_max_m ) <= FLT_EPSILON);

#endif

}

/* addition of simultaneous masking   Naoki Shibata 2000/7 */

inline static FLOAT

vbrpsy_mask_add(FLOAT m1, FLOAT m2, int b, int delta)

{

    static const FLOAT table2[] = {

        1.33352 * 1.33352, 1.35879 * 1.35879, 1.38454 * 1.38454, 1.39497 * 1.39497,

        1.40548 * 1.40548, 1.3537 * 1.3537, 1.30382 * 1.30382, 1.22321 * 1.22321,

        1.14758 * 1.14758,

        1

    };

    FLOAT   ratio;

    if (m1 < 0) {

        m1 = 0;

    }

    if (m2 < 0) {

        m2 = 0;

    }

    if (m1 <= 0) {

        return m2;

    }

    if (m2 <= 0) {

        return m1;

    }

    if (m2 > m1) {

        ratio = m2 / m1;

    }

    else {

        ratio = m1 / m2;

    }

    if (abs(b) <= delta) {       /* approximately, 1 bark = 3 partitions */

        /* originally 'if(i > 8)' */

        if (ratio >= ma_max_i1) {

            return m1 + m2;

        }

        else {

            int     i = (int) (FAST_LOG10_X(ratio, 16.0f));

            return (m1 + m2) * table2[i];

        }

    }

    if (ratio < ma_max_i2) {

        return m1 + m2;

    }

    if (m1 < m2) {

        m1 = m2;

    }

    return m1;

}

/* short block threshold calculation (part 2)

    partition band bo_s[sfb] is at the transition from scalefactor

    band sfb to the next one sfb+1; enn and thmm have to be split

    between them

*/

static void

convert_partition2scalefac(PsyConst_CB2SB_t const *const gd, FLOAT const *eb, FLOAT const *thr,

                           FLOAT enn_out[], FLOAT thm_out[])

{

    FLOAT   enn, thmm;

    int     sb, b, n = gd->n_sb;

    enn = thmm = 0.0f;

    for (sb = b = 0; sb < n; ++b, ++sb) {

        int const bo_sb = gd->bo[sb];

        int const npart = gd->npart;

        int const b_lim = bo_sb < npart ? bo_sb : npart;

        while (b < b_lim) {

            assert(eb[b] >= 0); /* iff failed, it may indicate some index error elsewhere */

            assert(thr[b] >= 0);

            enn += eb[b];

            thmm += thr[b];

            b++;

        }

        if (b >= npart) {

            enn_out[sb] = enn;

            thm_out[sb] = thmm;

            ++sb;

            break;

        }

        assert(eb[b] >= 0); /* iff failed, it may indicate some index error elsewhere */

        assert(thr[b] >= 0);

        {

            /* at transition sfb -> sfb+1 */

            FLOAT const w_curr = gd->bo_weight[sb];

            FLOAT const w_next = 1.0f - w_curr;

            enn += w_curr * eb[b];

            thmm += w_curr * thr[b];

            enn_out[sb] = enn;

            thm_out[sb] = thmm;

            enn = w_next * eb[b];

            thmm = w_next * thr[b];

        }

    }

    /* zero initialize the rest */

    for (; sb < n; ++sb) {

        enn_out[sb] = 0;

        thm_out[sb] = 0;

    }

}

static void

convert_partition2scalefac_s(lame_internal_flags * gfc, FLOAT const *eb, FLOAT const *thr, int chn,

                             int sblock)

{

    PsyStateVar_t *const psv = &gfc->sv_psy;

    PsyConst_CB2SB_t const *const gds = &gfc->cd_psy->s;

    FLOAT   enn[SBMAX_s], thm[SBMAX_s];

    int     sb;

    convert_partition2scalefac(gds, eb, thr, enn, thm);

    for (sb = 0; sb < SBMAX_s; ++sb) {

        psv->en[chn].s[sb][sblock] = enn[sb];

        psv->thm[chn].s[sb][sblock] = thm[sb];

    }

}

/* longblock threshold calculation (part 2) */

static void

convert_partition2scalefac_l(lame_internal_flags * gfc, FLOAT const *eb, FLOAT const *thr, int chn)

{

    PsyStateVar_t *const psv = &gfc->sv_psy;

    PsyConst_CB2SB_t const *const gdl = &gfc->cd_psy->l;

    FLOAT  *enn = &psv->en[chn].l[0];

    FLOAT  *thm = &psv->thm[chn].l[0];

    convert_partition2scalefac(gdl, eb, thr, enn, thm);

}

static void

convert_partition2scalefac_l_to_s(lame_internal_flags * gfc, FLOAT const *eb, FLOAT const *thr,

                                  int chn)

{

    PsyStateVar_t *const psv = &gfc->sv_psy;

    PsyConst_CB2SB_t const *const gds = &gfc->cd_psy->l_to_s;

    FLOAT   enn[SBMAX_s], thm[SBMAX_s];

    int     sb, sblock;

    convert_partition2scalefac(gds, eb, thr, enn, thm);

    for (sb = 0; sb < SBMAX_s; ++sb) {

        FLOAT const scale = 1. / 64.f;

        FLOAT const tmp_enn = enn[sb];

        FLOAT const tmp_thm = thm[sb] * scale;

        for (sblock = 0; sblock < 3; ++sblock) {

            psv->en[chn].s[sb][sblock] = tmp_enn;

            psv->thm[chn].s[sb][sblock] = tmp_thm;

        }

    }

}

static inline FLOAT

NS_INTERP(FLOAT x, FLOAT y, FLOAT r)

{

    /* was pow((x),(r))*pow((y),1-(r)) */

    if (r >= 1.0f)

        return x;       /* 99.7% of the time */

    if (r <= 0.0f)

        return y;

    if (y > 0.0f)

        return powf(x / y, r) * y; /* rest of the time */

    return 0.0f;        /* never happens */

}

static  FLOAT

pecalc_s(III_psy_ratio const *mr, FLOAT masking_lower)

{

    FLOAT   pe_s;

    static const FLOAT regcoef_s[] = {

        11.8,           /* these values are tuned only for 44.1kHz... */

        13.6,

        17.2,

        32,

        46.5,

        51.3,

        57.5,

        67.1,

        71.5,

        84.6,

        97.6,

        130,

/*      255.8 */

    };

    unsigned int sb, sblock;

    pe_s = 1236.28f / 4;

    for (sb = 0; sb < SBMAX_s - 1; sb++) {

        for (sblock = 0; sblock < 3; sblock++) {

            FLOAT const thm = mr->thm.s[sb][sblock];

            assert(sb < dimension_of(regcoef_s));

            if (thm > 0.0f) {

                FLOAT const x = thm * masking_lower;

                FLOAT const en = mr->en.s[sb][sblock];

                if (en > x) {

                    if (en > x * 1e10f) {

                        pe_s += regcoef_s[sb] * (10.0f * LOG10);

                    }

                    else {

                        assert(x > 0);

                        pe_s += regcoef_s[sb] * FAST_LOG10(en / x);

                    }

                }

            }

        }

    }

    return pe_s;

}

static  FLOAT

pecalc_l(III_psy_ratio const *mr, FLOAT masking_lower)

{

    FLOAT   pe_l;

    static const FLOAT regcoef_l[] = {

        6.8,            /* these values are tuned only for 44.1kHz... */

        5.8,

        5.8,

        6.4,

        6.5,

        9.9,

        12.1,

        14.4,

        15,

        18.9,

        21.6,

        26.9,

        34.2,

        40.2,

        46.8,

        56.5,

        60.7,

        73.9,

        85.7,

        93.4,

        126.1,

/*      241.3 */

    };

    unsigned int sb;

    pe_l = 1124.23f / 4;

    for (sb = 0; sb < SBMAX_l - 1; sb++) {

        FLOAT const thm = mr->thm.l[sb];

        assert(sb < dimension_of(regcoef_l));

        if (thm > 0.0f) {

            FLOAT const x = thm * masking_lower;

            FLOAT const en = mr->en.l[sb];

            if (en > x) {

                if (en > x * 1e10f) {

                    pe_l += regcoef_l[sb] * (10.0f * LOG10);

                }

                else {

                    assert(x > 0);

                    pe_l += regcoef_l[sb] * FAST_LOG10(en / x);

                }

            }

        }

    }

    return pe_l;

}

static void

calc_energy(PsyConst_CB2SB_t const *l, FLOAT const *fftenergy, FLOAT * eb, FLOAT * max, FLOAT * avg)

{

    int     b, j;

    for (b = j = 0; b < l->npart; ++b) {

        FLOAT   ebb = 0, m = 0;

        int     i;

        for (i = 0; i < l->numlines[b]; ++i, ++j) {

            FLOAT const el = fftenergy[j];

            assert(el >= 0);

            ebb += el;

            if (m < el)

                m = el;

        }

        eb[b] = ebb;

        max[b] = m;

        avg[b] = ebb * l->rnumlines[b];

        assert(l->rnumlines[b] >= 0);

        assert(ebb >= 0);

        assert(eb[b] >= 0);

        assert(max[b] >= 0);

        assert(avg[b] >= 0);

    }

}

static void

calc_mask_index_l(lame_internal_flags const *gfc, FLOAT const *max,

                  FLOAT const *avg, unsigned char *mask_idx)

{

    PsyConst_CB2SB_t const *const gdl = &gfc->cd_psy->l;

    FLOAT   m, a;

    int     b, k;

    int const last_tab_entry = sizeof(tab) / sizeof(tab[0]) - 1;

    b = 0;

    a = avg[b] + avg[b + 1];

    assert(a >= 0);

    if (a > 0.0f) {

        m = max[b];

        if (m < max[b + 1])

            m = max[b + 1];

        assert((gdl->numlines[b] + gdl->numlines[b + 1] - 1) > 0);

        a = 20.0f * (m * 2.0f - a)

            / (a * (gdl->numlines[b] + gdl->numlines[b + 1] - 1));

        k = (int) a;

        if (k > last_tab_entry)

            k = last_tab_entry;

        mask_idx[b] = k;

    }

    else {

        mask_idx[b] = 0;

    }

    for (b = 1; b < gdl->npart - 1; b++) {

        a = avg[b - 1] + avg[b] + avg[b + 1];

        assert(a >= 0);

        if (a > 0.0f) {

            m = max[b - 1];

            if (m < max[b])

                m = max[b];

            if (m < max[b + 1])

                m = max[b + 1];

            assert((gdl->numlines[b - 1] + gdl->numlines[b] + gdl->numlines[b + 1] - 1) > 0);

            a = 20.0f * (m * 3.0f - a)

                / (a * (gdl->numlines[b - 1] + gdl->numlines[b] + gdl->numlines[b + 1] - 1));

            k = (int) a;

            if (k > last_tab_entry)

                k = last_tab_entry;

            mask_idx[b] = k;

        }

        else {

            mask_idx[b] = 0;

        }

    }

    assert(b > 0);

    assert(b == gdl->npart - 1);

    a = avg[b - 1] + avg[b];

    assert(a >= 0);

    if (a > 0.0f) {

        m = max[b - 1];

        if (m < max[b])

            m = max[b];

        assert((gdl->numlines[b - 1] + gdl->numlines[b] - 1) > 0);

        a = 20.0f * (m * 2.0f - a)

            / (a * (gdl->numlines[b - 1] + gdl->numlines[b] - 1));

        k = (int) a;

        if (k > last_tab_entry)

            k = last_tab_entry;

        mask_idx[b] = k;

    }

    else {

        mask_idx[b] = 0;

    }

    assert(b == (gdl->npart - 1));

}

static void

vbrpsy_compute_fft_l(lame_internal_flags * gfc, const sample_t * const buffer[2], int chn,

                     int gr_out, FLOAT fftenergy[HBLKSIZE], FLOAT(*wsamp_l)[BLKSIZE])

{

    SessionConfig_t const *const cfg = &gfc->cfg;

    PsyStateVar_t *psv = &gfc->sv_psy;

    plotting_data *plt = cfg->analysis ? gfc->pinfo : 0;

    int     j;

    if (chn < 2) {

        fft_long(gfc, *wsamp_l, chn, buffer);

    }

    else if (chn == 2) {

        FLOAT const sqrt2_half = SQRT2 * 0.5f;

        /* FFT data for mid and side channel is derived from L & R */

        for (j = BLKSIZE - 1; j >= 0; --j) {

            FLOAT const l = wsamp_l[0][j];

            FLOAT const r = wsamp_l[1][j];

            wsamp_l[0][j] = (l + r) * sqrt2_half;

            wsamp_l[1][j] = (l - r) * sqrt2_half;

        }

    }

    /*********************************************************************

    *  compute energies

    *********************************************************************/

    fftenergy[0] = wsamp_l[0][0];

    fftenergy[0] *= fftenergy[0];

    for (j = BLKSIZE / 2 - 1; j >= 0; --j) {

        FLOAT const re = (*wsamp_l)[BLKSIZE / 2 - j];

        FLOAT const im = (*wsamp_l)[BLKSIZE / 2 + j];

        fftenergy[BLKSIZE / 2 - j] = (re * re + im * im) * 0.5f;

    }

    /* total energy */

    {

        FLOAT   totalenergy = 0.0f;

        for (j = 11; j < HBLKSIZE; j++)

            totalenergy += fftenergy[j];

        psv->tot_ener[chn] = totalenergy;

    }

    if (plt) {

        for (j = 0; j < HBLKSIZE; j++) {

            plt->energy[gr_out][chn][j] = plt->energy_save[chn][j];

            plt->energy_save[chn][j] = fftenergy[j];

        }

    }

}

static void

vbrpsy_compute_fft_s(lame_internal_flags const *gfc, const sample_t * const buffer[2], int chn,

                     int sblock, FLOAT(*fftenergy_s)[HBLKSIZE_s], FLOAT(*wsamp_s)[3][BLKSIZE_s])

{

    int     j;

    if (sblock == 0 && chn < 2) {

        fft_short(gfc, *wsamp_s, chn, buffer);

    }

    if (chn == 2) {

        FLOAT const sqrt2_half = SQRT2 * 0.5f;

        /* FFT data for mid and side channel is derived from L & R */

        for (j = BLKSIZE_s - 1; j >= 0; --j) {

            FLOAT const l = wsamp_s[0][sblock][j];

            FLOAT const r = wsamp_s[1][sblock][j];

            wsamp_s[0][sblock][j] = (l + r) * sqrt2_half;

            wsamp_s[1][sblock][j] = (l - r) * sqrt2_half;

        }

    }

    /*********************************************************************

    *  compute energies

    *********************************************************************/

    fftenergy_s[sblock][0] = (*wsamp_s)[sblock][0];

    fftenergy_s[sblock][0] *= fftenergy_s[sblock][0];

    for (j = BLKSIZE_s / 2 - 1; j >= 0; --j) {

        FLOAT const re = (*wsamp_s)[sblock][BLKSIZE_s / 2 - j];

        FLOAT const im = (*wsamp_s)[sblock][BLKSIZE_s / 2 + j];

        fftenergy_s[sblock][BLKSIZE_s / 2 - j] = (re * re + im * im) * 0.5f;

    }

}

    /*********************************************************************

    * compute loudness approximation (used for ATH auto-level adjustment) 

    *********************************************************************/

static void

vbrpsy_compute_loudness_approximation_l(lame_internal_flags * gfc, int gr_out, int chn,

                                        const FLOAT fftenergy[HBLKSIZE])

{

    PsyStateVar_t *psv = &gfc->sv_psy;

    if (chn < 2) {      /*no loudness for mid/side ch */

        gfc->ov_psy.loudness_sq[gr_out][chn] = psv->loudness_sq_save[chn];

        psv->loudness_sq_save[chn] = psycho_loudness_approx(fftenergy, gfc->ATH->eql_w);

    }

}

    /**********************************************************************

    *  Apply HPF of fs/4 to the input signal.

    *  This is used for attack detection / handling.

    **********************************************************************/

static void

vbrpsy_attack_detection(lame_internal_flags * gfc, const sample_t * const buffer[2], int gr_out,

                        III_psy_ratio masking_ratio[2][2], III_psy_ratio masking_MS_ratio[2][2],

                        FLOAT energy[4], FLOAT sub_short_factor[4][3], int ns_attacks[4][4],

                        int uselongblock[2])

{

    FLOAT   ns_hpfsmpl[2][576];

    SessionConfig_t const *const cfg = &gfc->cfg;

    PsyStateVar_t *const psv = &gfc->sv_psy;

    plotting_data *plt = cfg->analysis ? gfc->pinfo : 0;

    int const n_chn_out = cfg->channels_out;

    /* chn=2 and 3 = Mid and Side channels */

    int const n_chn_psy = (cfg->mode == JOINT_STEREO) ? 4 : n_chn_out;

    int     chn, i, j;

    memset(&ns_hpfsmpl[0][0], 0, sizeof(ns_hpfsmpl));

    /* Don't copy the input buffer into a temporary buffer */

    /* unroll the loop 2 times */

    for (chn = 0; chn < n_chn_out; chn++) {

        static const FLOAT fircoef[] = {

            -8.65163e-18 * 2, -0.00851586 * 2, -6.74764e-18 * 2, 0.0209036 * 2,

            -3.36639e-17 * 2, -0.0438162 * 2, -1.54175e-17 * 2, 0.0931738 * 2,

            -5.52212e-17 * 2, -0.313819 * 2

        };

        /* apply high pass filter of fs/4 */

        const sample_t *const firbuf = &buffer[chn][576 - 350 - NSFIRLEN + 192];

        assert(dimension_of(fircoef) == ((NSFIRLEN - 1) / 2));

        for (i = 0; i < 576; i++) {

            FLOAT   sum1, sum2;

            sum1 = firbuf[i + 10];

            sum2 = 0.0;

            for (j = 0; j < ((NSFIRLEN - 1) / 2) - 1; j += 2) {

                sum1 += fircoef[j] * (firbuf[i + j] + firbuf[i + NSFIRLEN - j]);

                sum2 += fircoef[j + 1] * (firbuf[i + j + 1] + firbuf[i + NSFIRLEN - j - 1]);

            }

            ns_hpfsmpl[chn][i] = sum1 + sum2;

        }

        masking_ratio[gr_out][chn].en = psv->en[chn];

        masking_ratio[gr_out][chn].thm = psv->thm[chn];

        if (n_chn_psy > 2) {

            /* MS maskings  */

            /*percep_MS_entropy         [chn-2]     = gfc -> pe  [chn];  */

            masking_MS_ratio[gr_out][chn].en = psv->en[chn + 2];

            masking_MS_ratio[gr_out][chn].thm = psv->thm[chn + 2];

        }

    }

    for (chn = 0; chn < n_chn_psy; chn++) {

        FLOAT   attack_intensity[12];

        FLOAT   en_subshort[12];

        FLOAT   en_short[4] = { 0, 0, 0, 0 };

        FLOAT const *pf = ns_hpfsmpl[chn & 1];

        int     ns_uselongblock = 1;

        if (chn == 2) {

            for (i = 0, j = 576; j > 0; ++i, --j) {

                FLOAT const l = ns_hpfsmpl[0][i];

                FLOAT const r = ns_hpfsmpl[1][i];

                ns_hpfsmpl[0][i] = l + r;

                ns_hpfsmpl[1][i] = l - r;

            }

        }

        /*************************************************************** 

        * determine the block type (window type)

        ***************************************************************/

        /* calculate energies of each sub-shortblocks */

        for (i = 0; i < 3; i++) {

            en_subshort[i] = psv->last_en_subshort[chn][i + 6];

            assert(psv->last_en_subshort[chn][i + 4] > 0);

            attack_intensity[i] = en_subshort[i] / psv->last_en_subshort[chn][i + 4];

            en_short[0] += en_subshort[i];

        }

        for (i = 0; i < 9; i++) {

            FLOAT const *const pfe = pf + 576 / 9;

            FLOAT   p = 1.;

            for (; pf < pfe; pf++)

                if (p < fabs(*pf))

                    p = fabs(*pf);

            psv->last_en_subshort[chn][i] = en_subshort[i + 3] = p;

            en_short[1 + i / 3] += p;

            if (p > en_subshort[i + 3 - 2]) {

                assert(en_subshort[i + 3 - 2] > 0);

                p = p / en_subshort[i + 3 - 2];

            }

            else if (en_subshort[i + 3 - 2] > p * 10.0f) {

                assert(p > 0);

                p = en_subshort[i + 3 - 2] / (p * 10.0f);

            }

            else {

                p = 0.0;

            }

            attack_intensity[i + 3] = p;

        }

        /* pulse like signal detection for fatboy.wav and so on */

        for (i = 0; i < 3; ++i) {

            FLOAT const enn =

                en_subshort[i * 3 + 3] + en_subshort[i * 3 + 4] + en_subshort[i * 3 + 5];

            FLOAT   factor = 1.f;

            if (en_subshort[i * 3 + 5] * 6 < enn) {

                factor *= 0.5f;

                if (en_subshort[i * 3 + 4] * 6 < enn) {

                    factor *= 0.5f;

                }

            }

            sub_short_factor[chn][i] = factor;

        }

        if (plt) {

            FLOAT   x = attack_intensity[0];

            for (i = 1; i < 12; i++) {

                if (x < attack_intensity[i]) {