yading@10: /*
yading@10:  * AAC coefficients encoder
yading@10:  * Copyright (C) 2008-2009 Konstantin Shishkov
yading@10:  *
yading@10:  * This file is part of FFmpeg.
yading@10:  *
yading@10:  * FFmpeg is free software; you can redistribute it and/or
yading@10:  * modify it under the terms of the GNU Lesser General Public
yading@10:  * License as published by the Free Software Foundation; either
yading@10:  * version 2.1 of the License, or (at your option) any later version.
yading@10:  *
yading@10:  * FFmpeg is distributed in the hope that it will be useful,
yading@10:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
yading@10:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
yading@10:  * Lesser General Public License for more details.
yading@10:  *
yading@10:  * You should have received a copy of the GNU Lesser General Public
yading@10:  * License along with FFmpeg; if not, write to the Free Software
yading@10:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
yading@10:  */
yading@10: 
yading@10: /**
yading@10:  * @file
yading@10:  * AAC coefficients encoder
yading@10:  */
yading@10: 
yading@10: /***********************************
yading@10:  *              TODOs:
yading@10:  * speedup quantizer selection
yading@10:  * add sane pulse detection
yading@10:  ***********************************/
yading@10: 
yading@10: #include "libavutil/libm.h" // brought forward to work around cygwin header breakage
yading@10: 
yading@10: #include <float.h>
yading@10: #include "libavutil/mathematics.h"
yading@10: #include "avcodec.h"
yading@10: #include "put_bits.h"
yading@10: #include "aac.h"
yading@10: #include "aacenc.h"
yading@10: #include "aactab.h"
yading@10: 
yading@10: /** bits needed to code codebook run value for long windows */
yading@10: static const uint8_t run_value_bits_long[64] = {
yading@10:      5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,
yading@10:      5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5, 10,
yading@10:     10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
yading@10:     10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 15
yading@10: };
yading@10: 
yading@10: /** bits needed to code codebook run value for short windows */
yading@10: static const uint8_t run_value_bits_short[16] = {
yading@10:     3, 3, 3, 3, 3, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 9
yading@10: };
yading@10: 
yading@10: static const uint8_t *run_value_bits[2] = {
yading@10:     run_value_bits_long, run_value_bits_short
yading@10: };
yading@10: 
yading@10: 
yading@10: /**
yading@10:  * Quantize one coefficient.
yading@10:  * @return absolute value of the quantized coefficient
yading@10:  * @see 3GPP TS26.403 5.6.2 "Scalefactor determination"
yading@10:  */
yading@10: static av_always_inline int quant(float coef, const float Q)
yading@10: {
yading@10:     float a = coef * Q;
yading@10:     return sqrtf(a * sqrtf(a)) + 0.4054;
yading@10: }
yading@10: 
yading@10: static void quantize_bands(int *out, const float *in, const float *scaled,
yading@10:                            int size, float Q34, int is_signed, int maxval)
yading@10: {
yading@10:     int i;
yading@10:     double qc;
yading@10:     for (i = 0; i < size; i++) {
yading@10:         qc = scaled[i] * Q34;
yading@10:         out[i] = (int)FFMIN(qc + 0.4054, (double)maxval);
yading@10:         if (is_signed && in[i] < 0.0f) {
yading@10:             out[i] = -out[i];
yading@10:         }
yading@10:     }
yading@10: }
yading@10: 
yading@10: static void abs_pow34_v(float *out, const float *in, const int size)
yading@10: {
yading@10: #ifndef USE_REALLY_FULL_SEARCH
yading@10:     int i;
yading@10:     for (i = 0; i < size; i++) {
yading@10:         float a = fabsf(in[i]);
yading@10:         out[i] = sqrtf(a * sqrtf(a));
yading@10:     }
yading@10: #endif /* USE_REALLY_FULL_SEARCH */
yading@10: }
yading@10: 
yading@10: static const uint8_t aac_cb_range [12] = {0, 3, 3, 3, 3, 9, 9, 8, 8, 13, 13, 17};
yading@10: static const uint8_t aac_cb_maxval[12] = {0, 1, 1, 2, 2, 4, 4, 7, 7, 12, 12, 16};
yading@10: 
yading@10: /**
yading@10:  * Calculate rate distortion cost for quantizing with given codebook
yading@10:  *
yading@10:  * @return quantization distortion
yading@10:  */
yading@10: static av_always_inline float quantize_and_encode_band_cost_template(
yading@10:                                 struct AACEncContext *s,
yading@10:                                 PutBitContext *pb, const float *in,
yading@10:                                 const float *scaled, int size, int scale_idx,
yading@10:                                 int cb, const float lambda, const float uplim,
yading@10:                                 int *bits, int BT_ZERO, int BT_UNSIGNED,
yading@10:                                 int BT_PAIR, int BT_ESC)
yading@10: {
yading@10:     const int q_idx = POW_SF2_ZERO - scale_idx + SCALE_ONE_POS - SCALE_DIV_512;
yading@10:     const float Q   = ff_aac_pow2sf_tab [q_idx];
yading@10:     const float Q34 = ff_aac_pow34sf_tab[q_idx];
yading@10:     const float IQ  = ff_aac_pow2sf_tab [POW_SF2_ZERO + scale_idx - SCALE_ONE_POS + SCALE_DIV_512];
yading@10:     const float CLIPPED_ESCAPE = 165140.0f*IQ;
yading@10:     int i, j;
yading@10:     float cost = 0;
yading@10:     const int dim = BT_PAIR ? 2 : 4;
yading@10:     int resbits = 0;
yading@10:     const int range  = aac_cb_range[cb];
yading@10:     const int maxval = aac_cb_maxval[cb];
yading@10:     int off;
yading@10: 
yading@10:     if (BT_ZERO) {
yading@10:         for (i = 0; i < size; i++)
yading@10:             cost += in[i]*in[i];
yading@10:         if (bits)
yading@10:             *bits = 0;
yading@10:         return cost * lambda;
yading@10:     }
yading@10:     if (!scaled) {
yading@10:         abs_pow34_v(s->scoefs, in, size);
yading@10:         scaled = s->scoefs;
yading@10:     }
yading@10:     quantize_bands(s->qcoefs, in, scaled, size, Q34, !BT_UNSIGNED, maxval);
yading@10:     if (BT_UNSIGNED) {
yading@10:         off = 0;
yading@10:     } else {
yading@10:         off = maxval;
yading@10:     }
yading@10:     for (i = 0; i < size; i += dim) {
yading@10:         const float *vec;
yading@10:         int *quants = s->qcoefs + i;
yading@10:         int curidx = 0;
yading@10:         int curbits;
yading@10:         float rd = 0.0f;
yading@10:         for (j = 0; j < dim; j++) {
yading@10:             curidx *= range;
yading@10:             curidx += quants[j] + off;
yading@10:         }
yading@10:         curbits =  ff_aac_spectral_bits[cb-1][curidx];
yading@10:         vec     = &ff_aac_codebook_vectors[cb-1][curidx*dim];
yading@10:         if (BT_UNSIGNED) {
yading@10:             for (j = 0; j < dim; j++) {
yading@10:                 float t = fabsf(in[i+j]);
yading@10:                 float di;
yading@10:                 if (BT_ESC && vec[j] == 64.0f) { //FIXME: slow
yading@10:                     if (t >= CLIPPED_ESCAPE) {
yading@10:                         di = t - CLIPPED_ESCAPE;
yading@10:                         curbits += 21;
yading@10:                     } else {
yading@10:                         int c = av_clip(quant(t, Q), 0, 8191);
yading@10:                         di = t - c*cbrtf(c)*IQ;
yading@10:                         curbits += av_log2(c)*2 - 4 + 1;
yading@10:                     }
yading@10:                 } else {
yading@10:                     di = t - vec[j]*IQ;
yading@10:                 }
yading@10:                 if (vec[j] != 0.0f)
yading@10:                     curbits++;
yading@10:                 rd += di*di;
yading@10:             }
yading@10:         } else {
yading@10:             for (j = 0; j < dim; j++) {
yading@10:                 float di = in[i+j] - vec[j]*IQ;
yading@10:                 rd += di*di;
yading@10:             }
yading@10:         }
yading@10:         cost    += rd * lambda + curbits;
yading@10:         resbits += curbits;
yading@10:         if (cost >= uplim)
yading@10:             return uplim;
yading@10:         if (pb) {
yading@10:             put_bits(pb, ff_aac_spectral_bits[cb-1][curidx], ff_aac_spectral_codes[cb-1][curidx]);
yading@10:             if (BT_UNSIGNED)
yading@10:                 for (j = 0; j < dim; j++)
yading@10:                     if (ff_aac_codebook_vectors[cb-1][curidx*dim+j] != 0.0f)
yading@10:                         put_bits(pb, 1, in[i+j] < 0.0f);
yading@10:             if (BT_ESC) {
yading@10:                 for (j = 0; j < 2; j++) {
yading@10:                     if (ff_aac_codebook_vectors[cb-1][curidx*2+j] == 64.0f) {
yading@10:                         int coef = av_clip(quant(fabsf(in[i+j]), Q), 0, 8191);
yading@10:                         int len = av_log2(coef);
yading@10: 
yading@10:                         put_bits(pb, len - 4 + 1, (1 << (len - 4 + 1)) - 2);
yading@10:                         put_bits(pb, len, coef & ((1 << len) - 1));
yading@10:                     }
yading@10:                 }
yading@10:             }
yading@10:         }
yading@10:     }
yading@10: 
yading@10:     if (bits)
yading@10:         *bits = resbits;
yading@10:     return cost;
yading@10: }
yading@10: 
yading@10: #define QUANTIZE_AND_ENCODE_BAND_COST_FUNC(NAME, BT_ZERO, BT_UNSIGNED, BT_PAIR, BT_ESC) \
yading@10: static float quantize_and_encode_band_cost_ ## NAME(                                        \
yading@10:                                 struct AACEncContext *s,                                \
yading@10:                                 PutBitContext *pb, const float *in,                     \
yading@10:                                 const float *scaled, int size, int scale_idx,           \
yading@10:                                 int cb, const float lambda, const float uplim,          \
yading@10:                                 int *bits) {                                            \
yading@10:     return quantize_and_encode_band_cost_template(                                      \
yading@10:                                 s, pb, in, scaled, size, scale_idx,                     \
yading@10:                                 BT_ESC ? ESC_BT : cb, lambda, uplim, bits,              \
yading@10:                                 BT_ZERO, BT_UNSIGNED, BT_PAIR, BT_ESC);                 \
yading@10: }
yading@10: 
yading@10: QUANTIZE_AND_ENCODE_BAND_COST_FUNC(ZERO,  1, 0, 0, 0)
yading@10: QUANTIZE_AND_ENCODE_BAND_COST_FUNC(SQUAD, 0, 0, 0, 0)
yading@10: QUANTIZE_AND_ENCODE_BAND_COST_FUNC(UQUAD, 0, 1, 0, 0)
yading@10: QUANTIZE_AND_ENCODE_BAND_COST_FUNC(SPAIR, 0, 0, 1, 0)
yading@10: QUANTIZE_AND_ENCODE_BAND_COST_FUNC(UPAIR, 0, 1, 1, 0)
yading@10: QUANTIZE_AND_ENCODE_BAND_COST_FUNC(ESC,   0, 1, 1, 1)
yading@10: 
yading@10: static float (*const quantize_and_encode_band_cost_arr[])(
yading@10:                                 struct AACEncContext *s,
yading@10:                                 PutBitContext *pb, const float *in,
yading@10:                                 const float *scaled, int size, int scale_idx,
yading@10:                                 int cb, const float lambda, const float uplim,
yading@10:                                 int *bits) = {
yading@10:     quantize_and_encode_band_cost_ZERO,
yading@10:     quantize_and_encode_band_cost_SQUAD,
yading@10:     quantize_and_encode_band_cost_SQUAD,
yading@10:     quantize_and_encode_band_cost_UQUAD,
yading@10:     quantize_and_encode_band_cost_UQUAD,
yading@10:     quantize_and_encode_band_cost_SPAIR,
yading@10:     quantize_and_encode_band_cost_SPAIR,
yading@10:     quantize_and_encode_band_cost_UPAIR,
yading@10:     quantize_and_encode_band_cost_UPAIR,
yading@10:     quantize_and_encode_band_cost_UPAIR,
yading@10:     quantize_and_encode_band_cost_UPAIR,
yading@10:     quantize_and_encode_band_cost_ESC,
yading@10: };
yading@10: 
yading@10: #define quantize_and_encode_band_cost(                                  \
yading@10:                                 s, pb, in, scaled, size, scale_idx, cb, \
yading@10:                                 lambda, uplim, bits)                    \
yading@10:     quantize_and_encode_band_cost_arr[cb](                              \
yading@10:                                 s, pb, in, scaled, size, scale_idx, cb, \
yading@10:                                 lambda, uplim, bits)
yading@10: 
yading@10: static float quantize_band_cost(struct AACEncContext *s, const float *in,
yading@10:                                 const float *scaled, int size, int scale_idx,
yading@10:                                 int cb, const float lambda, const float uplim,
yading@10:                                 int *bits)
yading@10: {
yading@10:     return quantize_and_encode_band_cost(s, NULL, in, scaled, size, scale_idx,
yading@10:                                          cb, lambda, uplim, bits);
yading@10: }
yading@10: 
yading@10: static void quantize_and_encode_band(struct AACEncContext *s, PutBitContext *pb,
yading@10:                                      const float *in, int size, int scale_idx,
yading@10:                                      int cb, const float lambda)
yading@10: {
yading@10:     quantize_and_encode_band_cost(s, pb, in, NULL, size, scale_idx, cb, lambda,
yading@10:                                   INFINITY, NULL);
yading@10: }
yading@10: 
yading@10: static float find_max_val(int group_len, int swb_size, const float *scaled) {
yading@10:     float maxval = 0.0f;
yading@10:     int w2, i;
yading@10:     for (w2 = 0; w2 < group_len; w2++) {
yading@10:         for (i = 0; i < swb_size; i++) {
yading@10:             maxval = FFMAX(maxval, scaled[w2*128+i]);
yading@10:         }
yading@10:     }
yading@10:     return maxval;
yading@10: }
yading@10: 
yading@10: static int find_min_book(float maxval, int sf) {
yading@10:     float Q = ff_aac_pow2sf_tab[POW_SF2_ZERO - sf + SCALE_ONE_POS - SCALE_DIV_512];
yading@10:     float Q34 = sqrtf(Q * sqrtf(Q));
yading@10:     int qmaxval, cb;
yading@10:     qmaxval = maxval * Q34 + 0.4054f;
yading@10:     if      (qmaxval ==  0) cb = 0;
yading@10:     else if (qmaxval ==  1) cb = 1;
yading@10:     else if (qmaxval ==  2) cb = 3;
yading@10:     else if (qmaxval <=  4) cb = 5;
yading@10:     else if (qmaxval <=  7) cb = 7;
yading@10:     else if (qmaxval <= 12) cb = 9;
yading@10:     else                    cb = 11;
yading@10:     return cb;
yading@10: }
yading@10: 
yading@10: /**
yading@10:  * structure used in optimal codebook search
yading@10:  */
yading@10: typedef struct BandCodingPath {
yading@10:     int prev_idx; ///< pointer to the previous path point
yading@10:     float cost;   ///< path cost
yading@10:     int run;
yading@10: } BandCodingPath;
yading@10: 
yading@10: /**
yading@10:  * Encode band info for single window group bands.
yading@10:  */
yading@10: static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce,
yading@10:                                      int win, int group_len, const float lambda)
yading@10: {
yading@10:     BandCodingPath path[120][12];
yading@10:     int w, swb, cb, start, size;
yading@10:     int i, j;
yading@10:     const int max_sfb  = sce->ics.max_sfb;
yading@10:     const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
yading@10:     const int run_esc  = (1 << run_bits) - 1;
yading@10:     int idx, ppos, count;
yading@10:     int stackrun[120], stackcb[120], stack_len;
yading@10:     float next_minrd = INFINITY;
yading@10:     int next_mincb = 0;
yading@10: 
yading@10:     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
yading@10:     start = win*128;
yading@10:     for (cb = 0; cb < 12; cb++) {
yading@10:         path[0][cb].cost     = 0.0f;
yading@10:         path[0][cb].prev_idx = -1;
yading@10:         path[0][cb].run      = 0;
yading@10:     }
yading@10:     for (swb = 0; swb < max_sfb; swb++) {
yading@10:         size = sce->ics.swb_sizes[swb];
yading@10:         if (sce->zeroes[win*16 + swb]) {
yading@10:             for (cb = 0; cb < 12; cb++) {
yading@10:                 path[swb+1][cb].prev_idx = cb;
yading@10:                 path[swb+1][cb].cost     = path[swb][cb].cost;
yading@10:                 path[swb+1][cb].run      = path[swb][cb].run + 1;
yading@10:             }
yading@10:         } else {
yading@10:             float minrd = next_minrd;
yading@10:             int mincb = next_mincb;
yading@10:             next_minrd = INFINITY;
yading@10:             next_mincb = 0;
yading@10:             for (cb = 0; cb < 12; cb++) {
yading@10:                 float cost_stay_here, cost_get_here;
yading@10:                 float rd = 0.0f;
yading@10:                 for (w = 0; w < group_len; w++) {
yading@10:                     FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(win+w)*16+swb];
yading@10:                     rd += quantize_band_cost(s, sce->coeffs + start + w*128,
yading@10:                                              s->scoefs + start + w*128, size,
yading@10:                                              sce->sf_idx[(win+w)*16+swb], cb,
yading@10:                                              lambda / band->threshold, INFINITY, NULL);
yading@10:                 }
yading@10:                 cost_stay_here = path[swb][cb].cost + rd;
yading@10:                 cost_get_here  = minrd              + rd + run_bits + 4;
yading@10:                 if (   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
yading@10:                     != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
yading@10:                     cost_stay_here += run_bits;
yading@10:                 if (cost_get_here < cost_stay_here) {
yading@10:                     path[swb+1][cb].prev_idx = mincb;
yading@10:                     path[swb+1][cb].cost     = cost_get_here;
yading@10:                     path[swb+1][cb].run      = 1;
yading@10:                 } else {
yading@10:                     path[swb+1][cb].prev_idx = cb;
yading@10:                     path[swb+1][cb].cost     = cost_stay_here;
yading@10:                     path[swb+1][cb].run      = path[swb][cb].run + 1;
yading@10:                 }
yading@10:                 if (path[swb+1][cb].cost < next_minrd) {
yading@10:                     next_minrd = path[swb+1][cb].cost;
yading@10:                     next_mincb = cb;
yading@10:                 }
yading@10:             }
yading@10:         }
yading@10:         start += sce->ics.swb_sizes[swb];
yading@10:     }
yading@10: 
yading@10:     //convert resulting path from backward-linked list
yading@10:     stack_len = 0;
yading@10:     idx       = 0;
yading@10:     for (cb = 1; cb < 12; cb++)
yading@10:         if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
yading@10:             idx = cb;
yading@10:     ppos = max_sfb;
yading@10:     while (ppos > 0) {
yading@10:         cb = idx;
yading@10:         stackrun[stack_len] = path[ppos][cb].run;
yading@10:         stackcb [stack_len] = cb;
yading@10:         idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
yading@10:         ppos -= path[ppos][cb].run;
yading@10:         stack_len++;
yading@10:     }
yading@10:     //perform actual band info encoding
yading@10:     start = 0;
yading@10:     for (i = stack_len - 1; i >= 0; i--) {
yading@10:         put_bits(&s->pb, 4, stackcb[i]);
yading@10:         count = stackrun[i];
yading@10:         memset(sce->zeroes + win*16 + start, !stackcb[i], count);
yading@10:         //XXX: memset when band_type is also uint8_t
yading@10:         for (j = 0; j < count; j++) {
yading@10:             sce->band_type[win*16 + start] =  stackcb[i];
yading@10:             start++;
yading@10:         }
yading@10:         while (count >= run_esc) {
yading@10:             put_bits(&s->pb, run_bits, run_esc);
yading@10:             count -= run_esc;
yading@10:         }
yading@10:         put_bits(&s->pb, run_bits, count);
yading@10:     }
yading@10: }
yading@10: 
yading@10: static void codebook_trellis_rate(AACEncContext *s, SingleChannelElement *sce,
yading@10:                                   int win, int group_len, const float lambda)
yading@10: {
yading@10:     BandCodingPath path[120][12];
yading@10:     int w, swb, cb, start, size;
yading@10:     int i, j;
yading@10:     const int max_sfb  = sce->ics.max_sfb;
yading@10:     const int run_bits = sce->ics.num_windows == 1 ? 5 : 3;
yading@10:     const int run_esc  = (1 << run_bits) - 1;
yading@10:     int idx, ppos, count;
yading@10:     int stackrun[120], stackcb[120], stack_len;
yading@10:     float next_minbits = INFINITY;
yading@10:     int next_mincb = 0;
yading@10: 
yading@10:     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
yading@10:     start = win*128;
yading@10:     for (cb = 0; cb < 12; cb++) {
yading@10:         path[0][cb].cost     = run_bits+4;
yading@10:         path[0][cb].prev_idx = -1;
yading@10:         path[0][cb].run      = 0;
yading@10:     }
yading@10:     for (swb = 0; swb < max_sfb; swb++) {
yading@10:         size = sce->ics.swb_sizes[swb];
yading@10:         if (sce->zeroes[win*16 + swb]) {
yading@10:             float cost_stay_here = path[swb][0].cost;
yading@10:             float cost_get_here  = next_minbits + run_bits + 4;
yading@10:             if (   run_value_bits[sce->ics.num_windows == 8][path[swb][0].run]
yading@10:                 != run_value_bits[sce->ics.num_windows == 8][path[swb][0].run+1])
yading@10:                 cost_stay_here += run_bits;
yading@10:             if (cost_get_here < cost_stay_here) {
yading@10:                 path[swb+1][0].prev_idx = next_mincb;
yading@10:                 path[swb+1][0].cost     = cost_get_here;
yading@10:                 path[swb+1][0].run      = 1;
yading@10:             } else {
yading@10:                 path[swb+1][0].prev_idx = 0;
yading@10:                 path[swb+1][0].cost     = cost_stay_here;
yading@10:                 path[swb+1][0].run      = path[swb][0].run + 1;
yading@10:             }
yading@10:             next_minbits = path[swb+1][0].cost;
yading@10:             next_mincb = 0;
yading@10:             for (cb = 1; cb < 12; cb++) {
yading@10:                 path[swb+1][cb].cost = 61450;
yading@10:                 path[swb+1][cb].prev_idx = -1;
yading@10:                 path[swb+1][cb].run = 0;
yading@10:             }
yading@10:         } else {
yading@10:             float minbits = next_minbits;
yading@10:             int mincb = next_mincb;
yading@10:             int startcb = sce->band_type[win*16+swb];
yading@10:             next_minbits = INFINITY;
yading@10:             next_mincb = 0;
yading@10:             for (cb = 0; cb < startcb; cb++) {
yading@10:                 path[swb+1][cb].cost = 61450;
yading@10:                 path[swb+1][cb].prev_idx = -1;
yading@10:                 path[swb+1][cb].run = 0;
yading@10:             }
yading@10:             for (cb = startcb; cb < 12; cb++) {
yading@10:                 float cost_stay_here, cost_get_here;
yading@10:                 float bits = 0.0f;
yading@10:                 for (w = 0; w < group_len; w++) {
yading@10:                     bits += quantize_band_cost(s, sce->coeffs + start + w*128,
yading@10:                                                s->scoefs + start + w*128, size,
yading@10:                                                sce->sf_idx[(win+w)*16+swb], cb,
yading@10:                                                0, INFINITY, NULL);
yading@10:                 }
yading@10:                 cost_stay_here = path[swb][cb].cost + bits;
yading@10:                 cost_get_here  = minbits            + bits + run_bits + 4;
yading@10:                 if (   run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run]
yading@10:                     != run_value_bits[sce->ics.num_windows == 8][path[swb][cb].run+1])
yading@10:                     cost_stay_here += run_bits;
yading@10:                 if (cost_get_here < cost_stay_here) {
yading@10:                     path[swb+1][cb].prev_idx = mincb;
yading@10:                     path[swb+1][cb].cost     = cost_get_here;
yading@10:                     path[swb+1][cb].run      = 1;
yading@10:                 } else {
yading@10:                     path[swb+1][cb].prev_idx = cb;
yading@10:                     path[swb+1][cb].cost     = cost_stay_here;
yading@10:                     path[swb+1][cb].run      = path[swb][cb].run + 1;
yading@10:                 }
yading@10:                 if (path[swb+1][cb].cost < next_minbits) {
yading@10:                     next_minbits = path[swb+1][cb].cost;
yading@10:                     next_mincb = cb;
yading@10:                 }
yading@10:             }
yading@10:         }
yading@10:         start += sce->ics.swb_sizes[swb];
yading@10:     }
yading@10: 
yading@10:     //convert resulting path from backward-linked list
yading@10:     stack_len = 0;
yading@10:     idx       = 0;
yading@10:     for (cb = 1; cb < 12; cb++)
yading@10:         if (path[max_sfb][cb].cost < path[max_sfb][idx].cost)
yading@10:             idx = cb;
yading@10:     ppos = max_sfb;
yading@10:     while (ppos > 0) {
yading@10:         av_assert1(idx >= 0);
yading@10:         cb = idx;
yading@10:         stackrun[stack_len] = path[ppos][cb].run;
yading@10:         stackcb [stack_len] = cb;
yading@10:         idx = path[ppos-path[ppos][cb].run+1][cb].prev_idx;
yading@10:         ppos -= path[ppos][cb].run;
yading@10:         stack_len++;
yading@10:     }
yading@10:     //perform actual band info encoding
yading@10:     start = 0;
yading@10:     for (i = stack_len - 1; i >= 0; i--) {
yading@10:         put_bits(&s->pb, 4, stackcb[i]);
yading@10:         count = stackrun[i];
yading@10:         memset(sce->zeroes + win*16 + start, !stackcb[i], count);
yading@10:         //XXX: memset when band_type is also uint8_t
yading@10:         for (j = 0; j < count; j++) {
yading@10:             sce->band_type[win*16 + start] =  stackcb[i];
yading@10:             start++;
yading@10:         }
yading@10:         while (count >= run_esc) {
yading@10:             put_bits(&s->pb, run_bits, run_esc);
yading@10:             count -= run_esc;
yading@10:         }
yading@10:         put_bits(&s->pb, run_bits, count);
yading@10:     }
yading@10: }
yading@10: 
yading@10: /** Return the minimum scalefactor where the quantized coef does not clip. */
yading@10: static av_always_inline uint8_t coef2minsf(float coef) {
yading@10:     return av_clip_uint8(log2f(coef)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
yading@10: }
yading@10: 
yading@10: /** Return the maximum scalefactor where the quantized coef is not zero. */
yading@10: static av_always_inline uint8_t coef2maxsf(float coef) {
yading@10:     return av_clip_uint8(log2f(coef)*4 +  6 + SCALE_ONE_POS - SCALE_DIV_512);
yading@10: }
yading@10: 
yading@10: typedef struct TrellisPath {
yading@10:     float cost;
yading@10:     int prev;
yading@10: } TrellisPath;
yading@10: 
yading@10: #define TRELLIS_STAGES 121
yading@10: #define TRELLIS_STATES (SCALE_MAX_DIFF+1)
yading@10: 
yading@10: static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
yading@10:                                        SingleChannelElement *sce,
yading@10:                                        const float lambda)
yading@10: {
yading@10:     int q, w, w2, g, start = 0;
yading@10:     int i, j;
yading@10:     int idx;
yading@10:     TrellisPath paths[TRELLIS_STAGES][TRELLIS_STATES];
yading@10:     int bandaddr[TRELLIS_STAGES];
yading@10:     int minq;
yading@10:     float mincost;
yading@10:     float q0f = FLT_MAX, q1f = 0.0f, qnrgf = 0.0f;
yading@10:     int q0, q1, qcnt = 0;
yading@10: 
yading@10:     for (i = 0; i < 1024; i++) {
yading@10:         float t = fabsf(sce->coeffs[i]);
yading@10:         if (t > 0.0f) {
yading@10:             q0f = FFMIN(q0f, t);
yading@10:             q1f = FFMAX(q1f, t);
yading@10:             qnrgf += t*t;
yading@10:             qcnt++;
yading@10:         }
yading@10:     }
yading@10: 
yading@10:     if (!qcnt) {
yading@10:         memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
yading@10:         memset(sce->zeroes, 1, sizeof(sce->zeroes));
yading@10:         return;
yading@10:     }
yading@10: 
yading@10:     //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
yading@10:     q0 = coef2minsf(q0f);
yading@10:     //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
yading@10:     q1 = coef2maxsf(q1f);
yading@10:     if (q1 - q0 > 60) {
yading@10:         int q0low  = q0;
yading@10:         int q1high = q1;
yading@10:         //minimum scalefactor index is when maximum nonzero coefficient after quantizing is not clipped
yading@10:         int qnrg = av_clip_uint8(log2f(sqrtf(qnrgf/qcnt))*4 - 31 + SCALE_ONE_POS - SCALE_DIV_512);
yading@10:         q1 = qnrg + 30;
yading@10:         q0 = qnrg - 30;
yading@10:         if (q0 < q0low) {
yading@10:             q1 += q0low - q0;
yading@10:             q0  = q0low;
yading@10:         } else if (q1 > q1high) {
yading@10:             q0 -= q1 - q1high;
yading@10:             q1  = q1high;
yading@10:         }
yading@10:     }
yading@10: 
yading@10:     for (i = 0; i < TRELLIS_STATES; i++) {
yading@10:         paths[0][i].cost    = 0.0f;
yading@10:         paths[0][i].prev    = -1;
yading@10:     }
yading@10:     for (j = 1; j < TRELLIS_STAGES; j++) {
yading@10:         for (i = 0; i < TRELLIS_STATES; i++) {
yading@10:             paths[j][i].cost    = INFINITY;
yading@10:             paths[j][i].prev    = -2;
yading@10:         }
yading@10:     }
yading@10:     idx = 1;
yading@10:     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
yading@10:     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
yading@10:         start = w*128;
yading@10:         for (g = 0; g < sce->ics.num_swb; g++) {
yading@10:             const float *coefs = sce->coeffs + start;
yading@10:             float qmin, qmax;
yading@10:             int nz = 0;
yading@10: 
yading@10:             bandaddr[idx] = w * 16 + g;
yading@10:             qmin = INT_MAX;
yading@10:             qmax = 0.0f;
yading@10:             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
yading@10:                 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
yading@10:                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
yading@10:                     sce->zeroes[(w+w2)*16+g] = 1;
yading@10:                     continue;
yading@10:                 }
yading@10:                 sce->zeroes[(w+w2)*16+g] = 0;
yading@10:                 nz = 1;
yading@10:                 for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
yading@10:                     float t = fabsf(coefs[w2*128+i]);
yading@10:                     if (t > 0.0f)
yading@10:                         qmin = FFMIN(qmin, t);
yading@10:                     qmax = FFMAX(qmax, t);
yading@10:                 }
yading@10:             }
yading@10:             if (nz) {
yading@10:                 int minscale, maxscale;
yading@10:                 float minrd = INFINITY;
yading@10:                 float maxval;
yading@10:                 //minimum scalefactor index is when minimum nonzero coefficient after quantizing is not clipped
yading@10:                 minscale = coef2minsf(qmin);
yading@10:                 //maximum scalefactor index is when maximum coefficient after quantizing is still not zero
yading@10:                 maxscale = coef2maxsf(qmax);
yading@10:                 minscale = av_clip(minscale - q0, 0, TRELLIS_STATES - 1);
yading@10:                 maxscale = av_clip(maxscale - q0, 0, TRELLIS_STATES);
yading@10:                 maxval = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], s->scoefs+start);
yading@10:                 for (q = minscale; q < maxscale; q++) {
yading@10:                     float dist = 0;
yading@10:                     int cb = find_min_book(maxval, sce->sf_idx[w*16+g]);
yading@10:                     for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
yading@10:                         FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
yading@10:                         dist += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
yading@10:                                                    q + q0, cb, lambda / band->threshold, INFINITY, NULL);
yading@10:                     }
yading@10:                     minrd = FFMIN(minrd, dist);
yading@10: 
yading@10:                     for (i = 0; i < q1 - q0; i++) {
yading@10:                         float cost;
yading@10:                         cost = paths[idx - 1][i].cost + dist
yading@10:                                + ff_aac_scalefactor_bits[q - i + SCALE_DIFF_ZERO];
yading@10:                         if (cost < paths[idx][q].cost) {
yading@10:                             paths[idx][q].cost    = cost;
yading@10:                             paths[idx][q].prev    = i;
yading@10:                         }
yading@10:                     }
yading@10:                 }
yading@10:             } else {
yading@10:                 for (q = 0; q < q1 - q0; q++) {
yading@10:                     paths[idx][q].cost = paths[idx - 1][q].cost + 1;
yading@10:                     paths[idx][q].prev = q;
yading@10:                 }
yading@10:             }
yading@10:             sce->zeroes[w*16+g] = !nz;
yading@10:             start += sce->ics.swb_sizes[g];
yading@10:             idx++;
yading@10:         }
yading@10:     }
yading@10:     idx--;
yading@10:     mincost = paths[idx][0].cost;
yading@10:     minq    = 0;
yading@10:     for (i = 1; i < TRELLIS_STATES; i++) {
yading@10:         if (paths[idx][i].cost < mincost) {
yading@10:             mincost = paths[idx][i].cost;
yading@10:             minq = i;
yading@10:         }
yading@10:     }
yading@10:     while (idx) {
yading@10:         sce->sf_idx[bandaddr[idx]] = minq + q0;
yading@10:         minq = paths[idx][minq].prev;
yading@10:         idx--;
yading@10:     }
yading@10:     //set the same quantizers inside window groups
yading@10:     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
yading@10:         for (g = 0;  g < sce->ics.num_swb; g++)
yading@10:             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
yading@10:                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
yading@10: }
yading@10: 
yading@10: /**
yading@10:  * two-loop quantizers search taken from ISO 13818-7 Appendix C
yading@10:  */
yading@10: static void search_for_quantizers_twoloop(AVCodecContext *avctx,
yading@10:                                           AACEncContext *s,
yading@10:                                           SingleChannelElement *sce,
yading@10:                                           const float lambda)
yading@10: {
yading@10:     int start = 0, i, w, w2, g;
yading@10:     int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels;
yading@10:     float dists[128] = { 0 }, uplims[128];
yading@10:     float maxvals[128];
yading@10:     int fflag, minscaler;
yading@10:     int its  = 0;
yading@10:     int allz = 0;
yading@10:     float minthr = INFINITY;
yading@10: 
yading@10:     // for values above this the decoder might end up in an endless loop
yading@10:     // due to always having more bits than what can be encoded.
yading@10:     destbits = FFMIN(destbits, 5800);
yading@10:     //XXX: some heuristic to determine initial quantizers will reduce search time
yading@10:     //determine zero bands and upper limits
yading@10:     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
yading@10:         for (g = 0;  g < sce->ics.num_swb; g++) {
yading@10:             int nz = 0;
yading@10:             float uplim = 0.0f;
yading@10:             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
yading@10:                 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
yading@10:                 uplim += band->threshold;
yading@10:                 if (band->energy <= band->threshold || band->threshold == 0.0f) {
yading@10:                     sce->zeroes[(w+w2)*16+g] = 1;
yading@10:                     continue;
yading@10:                 }
yading@10:                 nz = 1;
yading@10:             }
yading@10:             uplims[w*16+g] = uplim *512;
yading@10:             sce->zeroes[w*16+g] = !nz;
yading@10:             if (nz)
yading@10:                 minthr = FFMIN(minthr, uplim);
yading@10:             allz |= nz;
yading@10:         }
yading@10:     }
yading@10:     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
yading@10:         for (g = 0;  g < sce->ics.num_swb; g++) {
yading@10:             if (sce->zeroes[w*16+g]) {
yading@10:                 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
yading@10:                 continue;
yading@10:             }
yading@10:             sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2f(uplims[w*16+g]/minthr)*4,59);
yading@10:         }
yading@10:     }
yading@10: 
yading@10:     if (!allz)
yading@10:         return;
yading@10:     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
yading@10: 
yading@10:     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
yading@10:         start = w*128;
yading@10:         for (g = 0;  g < sce->ics.num_swb; g++) {
yading@10:             const float *scaled = s->scoefs + start;
yading@10:             maxvals[w*16+g] = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], scaled);
yading@10:             start += sce->ics.swb_sizes[g];
yading@10:         }
yading@10:     }
yading@10: 
yading@10:     //perform two-loop search
yading@10:     //outer loop - improve quality
yading@10:     do {
yading@10:         int tbits, qstep;
yading@10:         minscaler = sce->sf_idx[0];
yading@10:         //inner loop - quantize spectrum to fit into given number of bits
yading@10:         qstep = its ? 1 : 32;
yading@10:         do {
yading@10:             int prev = -1;
yading@10:             tbits = 0;
yading@10:             fflag = 0;
yading@10:             for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
yading@10:                 start = w*128;
yading@10:                 for (g = 0;  g < sce->ics.num_swb; g++) {
yading@10:                     const float *coefs = sce->coeffs + start;
yading@10:                     const float *scaled = s->scoefs + start;
yading@10:                     int bits = 0;
yading@10:                     int cb;
yading@10:                     float dist = 0.0f;
yading@10: 
yading@10:                     if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
yading@10:                         start += sce->ics.swb_sizes[g];
yading@10:                         continue;
yading@10:                     }
yading@10:                     minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
yading@10:                     cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
yading@10:                     for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
yading@10:                         int b;
yading@10:                         dist += quantize_band_cost(s, coefs + w2*128,
yading@10:                                                    scaled + w2*128,
yading@10:                                                    sce->ics.swb_sizes[g],
yading@10:                                                    sce->sf_idx[w*16+g],
yading@10:                                                    cb,
yading@10:                                                    1.0f,
yading@10:                                                    INFINITY,
yading@10:                                                    &b);
yading@10:                         bits += b;
yading@10:                     }
yading@10:                     dists[w*16+g] = dist - bits;
yading@10:                     if (prev != -1) {
yading@10:                         bits += ff_aac_scalefactor_bits[sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO];
yading@10:                     }
yading@10:                     tbits += bits;
yading@10:                     start += sce->ics.swb_sizes[g];
yading@10:                     prev = sce->sf_idx[w*16+g];
yading@10:                 }
yading@10:             }
yading@10:             if (tbits > destbits) {
yading@10:                 for (i = 0; i < 128; i++)
yading@10:                     if (sce->sf_idx[i] < 218 - qstep)
yading@10:                         sce->sf_idx[i] += qstep;
yading@10:             } else {
yading@10:                 for (i = 0; i < 128; i++)
yading@10:                     if (sce->sf_idx[i] > 60 - qstep)
yading@10:                         sce->sf_idx[i] -= qstep;
yading@10:             }
yading@10:             qstep >>= 1;
yading@10:             if (!qstep && tbits > destbits*1.02 && sce->sf_idx[0] < 217)
yading@10:                 qstep = 1;
yading@10:         } while (qstep);
yading@10: 
yading@10:         fflag = 0;
yading@10:         minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
yading@10:         for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
yading@10:             for (g = 0; g < sce->ics.num_swb; g++) {
yading@10:                 int prevsc = sce->sf_idx[w*16+g];
yading@10:                 if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60) {
yading@10:                     if (find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]-1))
yading@10:                         sce->sf_idx[w*16+g]--;
yading@10:                     else //Try to make sure there is some energy in every band
yading@10:                         sce->sf_idx[w*16+g]-=2;
yading@10:                 }
yading@10:                 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
yading@10:                 sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
yading@10:                 if (sce->sf_idx[w*16+g] != prevsc)
yading@10:                     fflag = 1;
yading@10:                 sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
yading@10:             }
yading@10:         }
yading@10:         its++;
yading@10:     } while (fflag && its < 10);
yading@10: }
yading@10: 
yading@10: static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
yading@10:                                        SingleChannelElement *sce,
yading@10:                                        const float lambda)
yading@10: {
yading@10:     int start = 0, i, w, w2, g;
yading@10:     float uplim[128], maxq[128];
yading@10:     int minq, maxsf;
yading@10:     float distfact = ((sce->ics.num_windows > 1) ? 85.80 : 147.84) / lambda;
yading@10:     int last = 0, lastband = 0, curband = 0;
yading@10:     float avg_energy = 0.0;
yading@10:     if (sce->ics.num_windows == 1) {
yading@10:         start = 0;
yading@10:         for (i = 0; i < 1024; i++) {
yading@10:             if (i - start >= sce->ics.swb_sizes[curband]) {
yading@10:                 start += sce->ics.swb_sizes[curband];
yading@10:                 curband++;
yading@10:             }
yading@10:             if (sce->coeffs[i]) {
yading@10:                 avg_energy += sce->coeffs[i] * sce->coeffs[i];
yading@10:                 last = i;
yading@10:                 lastband = curband;
yading@10:             }
yading@10:         }
yading@10:     } else {
yading@10:         for (w = 0; w < 8; w++) {
yading@10:             const float *coeffs = sce->coeffs + w*128;
yading@10:             curband = start = 0;
yading@10:             for (i = 0; i < 128; i++) {
yading@10:                 if (i - start >= sce->ics.swb_sizes[curband]) {
yading@10:                     start += sce->ics.swb_sizes[curband];
yading@10:                     curband++;
yading@10:                 }
yading@10:                 if (coeffs[i]) {
yading@10:                     avg_energy += coeffs[i] * coeffs[i];
yading@10:                     last = FFMAX(last, i);
yading@10:                     lastband = FFMAX(lastband, curband);
yading@10:                 }
yading@10:             }
yading@10:         }
yading@10:     }
yading@10:     last++;
yading@10:     avg_energy /= last;
yading@10:     if (avg_energy == 0.0f) {
yading@10:         for (i = 0; i < FF_ARRAY_ELEMS(sce->sf_idx); i++)
yading@10:             sce->sf_idx[i] = SCALE_ONE_POS;
yading@10:         return;
yading@10:     }
yading@10:     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
yading@10:         start = w*128;
yading@10:         for (g = 0; g < sce->ics.num_swb; g++) {
yading@10:             float *coefs   = sce->coeffs + start;
yading@10:             const int size = sce->ics.swb_sizes[g];
yading@10:             int start2 = start, end2 = start + size, peakpos = start;
yading@10:             float maxval = -1, thr = 0.0f, t;
yading@10:             maxq[w*16+g] = 0.0f;
yading@10:             if (g > lastband) {
yading@10:                 maxq[w*16+g] = 0.0f;
yading@10:                 start += size;
yading@10:                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
yading@10:                     memset(coefs + w2*128, 0, sizeof(coefs[0])*size);
yading@10:                 continue;
yading@10:             }
yading@10:             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
yading@10:                 for (i = 0; i < size; i++) {
yading@10:                     float t = coefs[w2*128+i]*coefs[w2*128+i];
yading@10:                     maxq[w*16+g] = FFMAX(maxq[w*16+g], fabsf(coefs[w2*128 + i]));
yading@10:                     thr += t;
yading@10:                     if (sce->ics.num_windows == 1 && maxval < t) {
yading@10:                         maxval  = t;
yading@10:                         peakpos = start+i;
yading@10:                     }
yading@10:                 }
yading@10:             }
yading@10:             if (sce->ics.num_windows == 1) {
yading@10:                 start2 = FFMAX(peakpos - 2, start2);
yading@10:                 end2   = FFMIN(peakpos + 3, end2);
yading@10:             } else {
yading@10:                 start2 -= start;
yading@10:                 end2   -= start;
yading@10:             }
yading@10:             start += size;
yading@10:             thr = pow(thr / (avg_energy * (end2 - start2)), 0.3 + 0.1*(lastband - g) / lastband);
yading@10:             t   = 1.0 - (1.0 * start2 / last);
yading@10:             uplim[w*16+g] = distfact / (1.4 * thr + t*t*t + 0.075);
yading@10:         }
yading@10:     }
yading@10:     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
yading@10:     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
yading@10:     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
yading@10:         start = w*128;
yading@10:         for (g = 0;  g < sce->ics.num_swb; g++) {
yading@10:             const float *coefs  = sce->coeffs + start;
yading@10:             const float *scaled = s->scoefs   + start;
yading@10:             const int size      = sce->ics.swb_sizes[g];
yading@10:             int scf, prev_scf, step;
yading@10:             int min_scf = -1, max_scf = 256;
yading@10:             float curdiff;
yading@10:             if (maxq[w*16+g] < 21.544) {
yading@10:                 sce->zeroes[w*16+g] = 1;
yading@10:                 start += size;
yading@10:                 continue;
yading@10:             }
yading@10:             sce->zeroes[w*16+g] = 0;
yading@10:             scf  = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2f(1/maxq[w*16+g])*16/3, 60, 218);
yading@10:             step = 16;
yading@10:             for (;;) {
yading@10:                 float dist = 0.0f;
yading@10:                 int quant_max;
yading@10: 
yading@10:                 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
yading@10:                     int b;
yading@10:                     dist += quantize_band_cost(s, coefs + w2*128,
yading@10:                                                scaled + w2*128,
yading@10:                                                sce->ics.swb_sizes[g],
yading@10:                                                scf,
yading@10:                                                ESC_BT,
yading@10:                                                lambda,
yading@10:                                                INFINITY,
yading@10:                                                &b);
yading@10:                     dist -= b;
yading@10:                 }
yading@10:                 dist *= 1.0f / 512.0f / lambda;
yading@10:                 quant_max = quant(maxq[w*16+g], ff_aac_pow2sf_tab[POW_SF2_ZERO - scf + SCALE_ONE_POS - SCALE_DIV_512]);
yading@10:                 if (quant_max >= 8191) { // too much, return to the previous quantizer
yading@10:                     sce->sf_idx[w*16+g] = prev_scf;
yading@10:                     break;
yading@10:                 }
yading@10:                 prev_scf = scf;
yading@10:                 curdiff = fabsf(dist - uplim[w*16+g]);
yading@10:                 if (curdiff <= 1.0f)
yading@10:                     step = 0;
yading@10:                 else
yading@10:                     step = log2f(curdiff);
yading@10:                 if (dist > uplim[w*16+g])
yading@10:                     step = -step;
yading@10:                 scf += step;
yading@10:                 scf = av_clip_uint8(scf);
yading@10:                 step = scf - prev_scf;
yading@10:                 if (FFABS(step) <= 1 || (step > 0 && scf >= max_scf) || (step < 0 && scf <= min_scf)) {
yading@10:                     sce->sf_idx[w*16+g] = av_clip(scf, min_scf, max_scf);
yading@10:                     break;
yading@10:                 }
yading@10:                 if (step > 0)
yading@10:                     min_scf = prev_scf;
yading@10:                 else
yading@10:                     max_scf = prev_scf;
yading@10:             }
yading@10:             start += size;
yading@10:         }
yading@10:     }
yading@10:     minq = sce->sf_idx[0] ? sce->sf_idx[0] : INT_MAX;
yading@10:     for (i = 1; i < 128; i++) {
yading@10:         if (!sce->sf_idx[i])
yading@10:             sce->sf_idx[i] = sce->sf_idx[i-1];
yading@10:         else
yading@10:             minq = FFMIN(minq, sce->sf_idx[i]);
yading@10:     }
yading@10:     if (minq == INT_MAX)
yading@10:         minq = 0;
yading@10:     minq = FFMIN(minq, SCALE_MAX_POS);
yading@10:     maxsf = FFMIN(minq + SCALE_MAX_DIFF, SCALE_MAX_POS);
yading@10:     for (i = 126; i >= 0; i--) {
yading@10:         if (!sce->sf_idx[i])
yading@10:             sce->sf_idx[i] = sce->sf_idx[i+1];
yading@10:         sce->sf_idx[i] = av_clip(sce->sf_idx[i], minq, maxsf);
yading@10:     }
yading@10: }
yading@10: 
yading@10: static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
yading@10:                                        SingleChannelElement *sce,
yading@10:                                        const float lambda)
yading@10: {
yading@10:     int i, w, w2, g;
yading@10:     int minq = 255;
yading@10: 
yading@10:     memset(sce->sf_idx, 0, sizeof(sce->sf_idx));
yading@10:     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
yading@10:         for (g = 0; g < sce->ics.num_swb; g++) {
yading@10:             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
yading@10:                 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
yading@10:                 if (band->energy <= band->threshold) {
yading@10:                     sce->sf_idx[(w+w2)*16+g] = 218;
yading@10:                     sce->zeroes[(w+w2)*16+g] = 1;
yading@10:                 } else {
yading@10:                     sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2f(band->threshold), 80, 218);
yading@10:                     sce->zeroes[(w+w2)*16+g] = 0;
yading@10:                 }
yading@10:                 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);
yading@10:             }
yading@10:         }
yading@10:     }
yading@10:     for (i = 0; i < 128; i++) {
yading@10:         sce->sf_idx[i] = 140;
yading@10:         //av_clip(sce->sf_idx[i], minq, minq + SCALE_MAX_DIFF - 1);
yading@10:     }
yading@10:     //set the same quantizers inside window groups
yading@10:     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
yading@10:         for (g = 0;  g < sce->ics.num_swb; g++)
yading@10:             for (w2 = 1; w2 < sce->ics.group_len[w]; w2++)
yading@10:                 sce->sf_idx[(w+w2)*16+g] = sce->sf_idx[w*16+g];
yading@10: }
yading@10: 
yading@10: static void search_for_ms(AACEncContext *s, ChannelElement *cpe,
yading@10:                           const float lambda)
yading@10: {
yading@10:     int start = 0, i, w, w2, g;
yading@10:     float M[128], S[128];
yading@10:     float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
yading@10:     SingleChannelElement *sce0 = &cpe->ch[0];
yading@10:     SingleChannelElement *sce1 = &cpe->ch[1];
yading@10:     if (!cpe->common_window)
yading@10:         return;
yading@10:     for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
yading@10:         for (g = 0;  g < sce0->ics.num_swb; g++) {
yading@10:             if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
yading@10:                 float dist1 = 0.0f, dist2 = 0.0f;
yading@10:                 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
yading@10:                     FFPsyBand *band0 = &s->psy.ch[s->cur_channel+0].psy_bands[(w+w2)*16+g];
yading@10:                     FFPsyBand *band1 = &s->psy.ch[s->cur_channel+1].psy_bands[(w+w2)*16+g];
yading@10:                     float minthr = FFMIN(band0->threshold, band1->threshold);
yading@10:                     float maxthr = FFMAX(band0->threshold, band1->threshold);
yading@10:                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
yading@10:                         M[i] = (sce0->coeffs[start+w2*128+i]
yading@10:                               + sce1->coeffs[start+w2*128+i]) * 0.5;
yading@10:                         S[i] =  M[i]
yading@10:                               - sce1->coeffs[start+w2*128+i];
yading@10:                     }
yading@10:                     abs_pow34_v(L34, sce0->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
yading@10:                     abs_pow34_v(R34, sce1->coeffs+start+w2*128, sce0->ics.swb_sizes[g]);
yading@10:                     abs_pow34_v(M34, M,                         sce0->ics.swb_sizes[g]);
yading@10:                     abs_pow34_v(S34, S,                         sce0->ics.swb_sizes[g]);
yading@10:                     dist1 += quantize_band_cost(s, sce0->coeffs + start + w2*128,
yading@10:                                                 L34,
yading@10:                                                 sce0->ics.swb_sizes[g],
yading@10:                                                 sce0->sf_idx[(w+w2)*16+g],
yading@10:                                                 sce0->band_type[(w+w2)*16+g],
yading@10:                                                 lambda / band0->threshold, INFINITY, NULL);
yading@10:                     dist1 += quantize_band_cost(s, sce1->coeffs + start + w2*128,
yading@10:                                                 R34,
yading@10:                                                 sce1->ics.swb_sizes[g],
yading@10:                                                 sce1->sf_idx[(w+w2)*16+g],
yading@10:                                                 sce1->band_type[(w+w2)*16+g],
yading@10:                                                 lambda / band1->threshold, INFINITY, NULL);
yading@10:                     dist2 += quantize_band_cost(s, M,
yading@10:                                                 M34,
yading@10:                                                 sce0->ics.swb_sizes[g],
yading@10:                                                 sce0->sf_idx[(w+w2)*16+g],
yading@10:                                                 sce0->band_type[(w+w2)*16+g],
yading@10:                                                 lambda / maxthr, INFINITY, NULL);
yading@10:                     dist2 += quantize_band_cost(s, S,
yading@10:                                                 S34,
yading@10:                                                 sce1->ics.swb_sizes[g],
yading@10:                                                 sce1->sf_idx[(w+w2)*16+g],
yading@10:                                                 sce1->band_type[(w+w2)*16+g],
yading@10:                                                 lambda / minthr, INFINITY, NULL);
yading@10:                 }
yading@10:                 cpe->ms_mask[w*16+g] = dist2 < dist1;
yading@10:             }
yading@10:             start += sce0->ics.swb_sizes[g];
yading@10:         }
yading@10:     }
yading@10: }
yading@10: 
yading@10: AACCoefficientsEncoder ff_aac_coders[AAC_CODER_NB] = {
yading@10:     {
yading@10:         search_for_quantizers_faac,
yading@10:         encode_window_bands_info,
yading@10:         quantize_and_encode_band,
yading@10:         search_for_ms,
yading@10:     },
yading@10:     {
yading@10:         search_for_quantizers_anmr,
yading@10:         encode_window_bands_info,
yading@10:         quantize_and_encode_band,
yading@10:         search_for_ms,
yading@10:     },
yading@10:     {
yading@10:         search_for_quantizers_twoloop,
yading@10:         codebook_trellis_rate,
yading@10:         quantize_and_encode_band,
yading@10:         search_for_ms,
yading@10:     },
yading@10:     {
yading@10:         search_for_quantizers_fast,
yading@10:         encode_window_bands_info,
yading@10:         quantize_and_encode_band,
yading@10:         search_for_ms,
yading@10:     },
yading@10: };