Statistics
| Branch: | Revision:

ffmpeg / libavcodec / aacpsy.c @ f50d9377

History | View | Annotate | Download (35.2 KB)

1
/*
2
 * AAC encoder psychoacoustic model
3
 * Copyright (C) 2008 Konstantin Shishkov
4
 *
5
 * This file is part of Libav.
6
 *
7
 * Libav is free software; you can redistribute it and/or
8
 * modify it under the terms of the GNU Lesser General Public
9
 * License as published by the Free Software Foundation; either
10
 * version 2.1 of the License, or (at your option) any later version.
11
 *
12
 * Libav is distributed in the hope that it will be useful,
13
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15
 * Lesser General Public License for more details.
16
 *
17
 * You should have received a copy of the GNU Lesser General Public
18
 * License along with Libav; if not, write to the Free Software
19
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20
 */
21

    
22
/**
23
 * @file
24
 * AAC encoder psychoacoustic model
25
 */
26

    
27
#include "avcodec.h"
28
#include "aactab.h"
29
#include "psymodel.h"
30

    
31
/***********************************
32
 *              TODOs:
33
 * try other bitrate controlling mechanism (maybe use ratecontrol.c?)
34
 * control quality for quality-based output
35
 **********************************/
36

    
37
/**
38
 * constants for 3GPP AAC psychoacoustic model
39
 * @{
40
 */
41
#define PSY_3GPP_THR_SPREAD_HI   1.5f // spreading factor for low-to-hi threshold spreading  (15 dB/Bark)
42
#define PSY_3GPP_THR_SPREAD_LOW  3.0f // spreading factor for hi-to-low threshold spreading  (30 dB/Bark)
43
/* spreading factor for low-to-hi energy spreading, long block, > 22kbps/channel (20dB/Bark) */
44
#define PSY_3GPP_EN_SPREAD_HI_L1 2.0f
45
/* spreading factor for low-to-hi energy spreading, long block, <= 22kbps/channel (15dB/Bark) */
46
#define PSY_3GPP_EN_SPREAD_HI_L2 1.5f
47
/* spreading factor for low-to-hi energy spreading, short block (15 dB/Bark) */
48
#define PSY_3GPP_EN_SPREAD_HI_S  1.5f
49
/* spreading factor for hi-to-low energy spreading, long block (30dB/Bark) */
50
#define PSY_3GPP_EN_SPREAD_LOW_L 3.0f
51
/* spreading factor for hi-to-low energy spreading, short block (20dB/Bark) */
52
#define PSY_3GPP_EN_SPREAD_LOW_S 2.0f
53

    
54
#define PSY_3GPP_RPEMIN      0.01f
55
#define PSY_3GPP_RPELEV      2.0f
56

    
57
#define PSY_3GPP_C1          3.0f           /* log2(8) */
58
#define PSY_3GPP_C2          1.3219281f     /* log2(2.5) */
59
#define PSY_3GPP_C3          0.55935729f    /* 1 - C2 / C1 */
60

    
61
#define PSY_SNR_1DB          7.9432821e-1f  /* -1dB */
62
#define PSY_SNR_25DB         3.1622776e-3f  /* -25dB */
63

    
64
#define PSY_3GPP_SAVE_SLOPE_L  -0.46666667f
65
#define PSY_3GPP_SAVE_SLOPE_S  -0.36363637f
66
#define PSY_3GPP_SAVE_ADD_L    -0.84285712f
67
#define PSY_3GPP_SAVE_ADD_S    -0.75f
68
#define PSY_3GPP_SPEND_SLOPE_L  0.66666669f
69
#define PSY_3GPP_SPEND_SLOPE_S  0.81818181f
70
#define PSY_3GPP_SPEND_ADD_L   -0.35f
71
#define PSY_3GPP_SPEND_ADD_S   -0.26111111f
72
#define PSY_3GPP_CLIP_LO_L      0.2f
73
#define PSY_3GPP_CLIP_LO_S      0.2f
74
#define PSY_3GPP_CLIP_HI_L      0.95f
75
#define PSY_3GPP_CLIP_HI_S      0.75f
76

    
77
#define PSY_3GPP_AH_THR_LONG    0.5f
78
#define PSY_3GPP_AH_THR_SHORT   0.63f
79

    
80
enum {
81
    PSY_3GPP_AH_NONE,
82
    PSY_3GPP_AH_INACTIVE,
83
    PSY_3GPP_AH_ACTIVE
84
};
85

    
86
#define PSY_3GPP_BITS_TO_PE(bits) ((bits) * 1.18f)
87

    
88
/* LAME psy model constants */
89
#define PSY_LAME_FIR_LEN 21         ///< LAME psy model FIR order
90
#define AAC_BLOCK_SIZE_LONG 1024    ///< long block size
91
#define AAC_BLOCK_SIZE_SHORT 128    ///< short block size
92
#define AAC_NUM_BLOCKS_SHORT 8      ///< number of blocks in a short sequence
93
#define PSY_LAME_NUM_SUBBLOCKS 3    ///< Number of sub-blocks in each short block
94

    
95
/**
96
 * @}
97
 */
98

    
99
/**
100
 * information for single band used by 3GPP TS26.403-inspired psychoacoustic model
101
 */
102
typedef struct AacPsyBand{
103
    float energy;       ///< band energy
104
    float thr;          ///< energy threshold
105
    float thr_quiet;    ///< threshold in quiet
106
    float nz_lines;     ///< number of non-zero spectral lines
107
    float active_lines; ///< number of active spectral lines
108
    float pe;           ///< perceptual entropy
109
    float pe_const;     ///< constant part of the PE calculation
110
    float norm_fac;     ///< normalization factor for linearization
111
    int   avoid_holes;  ///< hole avoidance flag
112
}AacPsyBand;
113

    
114
/**
115
 * single/pair channel context for psychoacoustic model
116
 */
117
typedef struct AacPsyChannel{
118
    AacPsyBand band[128];               ///< bands information
119
    AacPsyBand prev_band[128];          ///< bands information from the previous frame
120

    
121
    float       win_energy;              ///< sliding average of channel energy
122
    float       iir_state[2];            ///< hi-pass IIR filter state
123
    uint8_t     next_grouping;           ///< stored grouping scheme for the next frame (in case of 8 short window sequence)
124
    enum WindowSequence next_window_seq; ///< window sequence to be used in the next frame
125
    /* LAME psy model specific members */
126
    float attack_threshold;              ///< attack threshold for this channel
127
    float prev_energy_subshort[AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS];
128
    int   prev_attack;                   ///< attack value for the last short block in the previous sequence
129
}AacPsyChannel;
130

    
131
/**
132
 * psychoacoustic model frame type-dependent coefficients
133
 */
134
typedef struct AacPsyCoeffs{
135
    float ath;           ///< absolute threshold of hearing per bands
136
    float barks;         ///< Bark value for each spectral band in long frame
137
    float spread_low[2]; ///< spreading factor for low-to-high threshold spreading in long frame
138
    float spread_hi [2]; ///< spreading factor for high-to-low threshold spreading in long frame
139
    float min_snr;       ///< minimal SNR
140
}AacPsyCoeffs;
141

    
142
/**
143
 * 3GPP TS26.403-inspired psychoacoustic model specific data
144
 */
145
typedef struct AacPsyContext{
146
    int chan_bitrate;     ///< bitrate per channel
147
    int frame_bits;       ///< average bits per frame
148
    int fill_level;       ///< bit reservoir fill level
149
    struct {
150
        float min;        ///< minimum allowed PE for bit factor calculation
151
        float max;        ///< maximum allowed PE for bit factor calculation
152
        float previous;   ///< allowed PE of the previous frame
153
        float correction; ///< PE correction factor
154
    } pe;
155
    AacPsyCoeffs psy_coef[2][64];
156
    AacPsyChannel *ch;
157
}AacPsyContext;
158

    
159
/**
160
 * LAME psy model preset struct
161
 */
162
typedef struct {
163
    int   quality;  ///< Quality to map the rest of the vaules to.
164
     /* This is overloaded to be both kbps per channel in ABR mode, and
165
      * requested quality in constant quality mode.
166
      */
167
    float st_lrm;   ///< short threshold for L, R, and M channels
168
} PsyLamePreset;
169

    
170
/**
171
 * LAME psy model preset table for ABR
172
 */
173
static const PsyLamePreset psy_abr_map[] = {
174
/* TODO: Tuning. These were taken from LAME. */
175
/* kbps/ch st_lrm   */
176
    {  8,  6.60},
177
    { 16,  6.60},
178
    { 24,  6.60},
179
    { 32,  6.60},
180
    { 40,  6.60},
181
    { 48,  6.60},
182
    { 56,  6.60},
183
    { 64,  6.40},
184
    { 80,  6.00},
185
    { 96,  5.60},
186
    {112,  5.20},
187
    {128,  5.20},
188
    {160,  5.20}
189
};
190

    
191
/**
192
* LAME psy model preset table for constant quality
193
*/
194
static const PsyLamePreset psy_vbr_map[] = {
195
/* vbr_q  st_lrm    */
196
    { 0,  4.20},
197
    { 1,  4.20},
198
    { 2,  4.20},
199
    { 3,  4.20},
200
    { 4,  4.20},
201
    { 5,  4.20},
202
    { 6,  4.20},
203
    { 7,  4.20},
204
    { 8,  4.20},
205
    { 9,  4.20},
206
    {10,  4.20}
207
};
208

    
209
/**
210
 * LAME psy model FIR coefficient table
211
 */
212
static const float psy_fir_coeffs[] = {
213
    -8.65163e-18 * 2, -0.00851586 * 2, -6.74764e-18 * 2, 0.0209036 * 2,
214
    -3.36639e-17 * 2, -0.0438162 * 2,  -1.54175e-17 * 2, 0.0931738 * 2,
215
    -5.52212e-17 * 2, -0.313819 * 2
216
};
217

    
218
/**
219
 * calculates the attack threshold for ABR from the above table for the LAME psy model
220
 */
221
static float lame_calc_attack_threshold(int bitrate)
222
{
223
    /* Assume max bitrate to start with */
224
    int lower_range = 12, upper_range = 12;
225
    int lower_range_kbps = psy_abr_map[12].quality;
226
    int upper_range_kbps = psy_abr_map[12].quality;
227
    int i;
228

    
229
    /* Determine which bitrates the value specified falls between.
230
     * If the loop ends without breaking our above assumption of 320kbps was correct.
231
     */
232
    for (i = 1; i < 13; i++) {
233
        if (FFMAX(bitrate, psy_abr_map[i].quality) != bitrate) {
234
            upper_range = i;
235
            upper_range_kbps = psy_abr_map[i    ].quality;
236
            lower_range = i - 1;
237
            lower_range_kbps = psy_abr_map[i - 1].quality;
238
            break; /* Upper range found */
239
        }
240
    }
241

    
242
    /* Determine which range the value specified is closer to */
243
    if ((upper_range_kbps - bitrate) > (bitrate - lower_range_kbps))
244
        return psy_abr_map[lower_range].st_lrm;
245
    return psy_abr_map[upper_range].st_lrm;
246
}
247

    
248
/**
249
 * LAME psy model specific initialization
250
 */
251
static void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx) {
252
    int i, j;
253

    
254
    for (i = 0; i < avctx->channels; i++) {
255
        AacPsyChannel *pch = &ctx->ch[i];
256

    
257
        if (avctx->flags & CODEC_FLAG_QSCALE)
258
            pch->attack_threshold = psy_vbr_map[avctx->global_quality / FF_QP2LAMBDA].st_lrm;
259
        else
260
            pch->attack_threshold = lame_calc_attack_threshold(avctx->bit_rate / avctx->channels / 1000);
261

    
262
        for (j = 0; j < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; j++)
263
            pch->prev_energy_subshort[j] = 10.0f;
264
    }
265
}
266

    
267
/**
268
 * Calculate Bark value for given line.
269
 */
270
static av_cold float calc_bark(float f)
271
{
272
    return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f));
273
}
274

    
275
#define ATH_ADD 4
276
/**
277
 * Calculate ATH value for given frequency.
278
 * Borrowed from Lame.
279
 */
280
static av_cold float ath(float f, float add)
281
{
282
    f /= 1000.0f;
283
    return    3.64 * pow(f, -0.8)
284
            - 6.8  * exp(-0.6  * (f - 3.4) * (f - 3.4))
285
            + 6.0  * exp(-0.15 * (f - 8.7) * (f - 8.7))
286
            + (0.6 + 0.04 * add) * 0.001 * f * f * f * f;
287
}
288

    
289
static av_cold int psy_3gpp_init(FFPsyContext *ctx) {
290
    AacPsyContext *pctx;
291
    float bark;
292
    int i, j, g, start;
293
    float prev, minscale, minath, minsnr, pe_min;
294
    const int chan_bitrate = ctx->avctx->bit_rate / ctx->avctx->channels;
295
    const int bandwidth    = ctx->avctx->cutoff ? ctx->avctx->cutoff : ctx->avctx->sample_rate / 2;
296
    const float num_bark   = calc_bark((float)bandwidth);
297

    
298
    ctx->model_priv_data = av_mallocz(sizeof(AacPsyContext));
299
    pctx = (AacPsyContext*) ctx->model_priv_data;
300

    
301
    pctx->chan_bitrate = chan_bitrate;
302
    pctx->frame_bits   = chan_bitrate * AAC_BLOCK_SIZE_LONG / ctx->avctx->sample_rate;
303
    pctx->pe.min       =  8.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
304
    pctx->pe.max       = 12.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
305
    ctx->bitres.size   = 6144 - pctx->frame_bits;
306
    ctx->bitres.size  -= ctx->bitres.size % 8;
307
    pctx->fill_level   = ctx->bitres.size;
308
    minath = ath(3410, ATH_ADD);
309
    for (j = 0; j < 2; j++) {
310
        AacPsyCoeffs *coeffs = pctx->psy_coef[j];
311
        const uint8_t *band_sizes = ctx->bands[j];
312
        float line_to_frequency = ctx->avctx->sample_rate / (j ? 256.f : 2048.0f);
313
        float avg_chan_bits = chan_bitrate / ctx->avctx->sample_rate * (j ? 128.0f : 1024.0f);
314
        /* reference encoder uses 2.4% here instead of 60% like the spec says */
315
        float bark_pe = 0.024f * PSY_3GPP_BITS_TO_PE(avg_chan_bits) / num_bark;
316
        float en_spread_low = j ? PSY_3GPP_EN_SPREAD_LOW_S : PSY_3GPP_EN_SPREAD_LOW_L;
317
        /* High energy spreading for long blocks <= 22kbps/channel and short blocks are the same. */
318
        float en_spread_hi  = (j || (chan_bitrate <= 22.0f)) ? PSY_3GPP_EN_SPREAD_HI_S : PSY_3GPP_EN_SPREAD_HI_L1;
319

    
320
        i = 0;
321
        prev = 0.0;
322
        for (g = 0; g < ctx->num_bands[j]; g++) {
323
            i += band_sizes[g];
324
            bark = calc_bark((i-1) * line_to_frequency);
325
            coeffs[g].barks = (bark + prev) / 2.0;
326
            prev = bark;
327
        }
328
        for (g = 0; g < ctx->num_bands[j] - 1; g++) {
329
            AacPsyCoeffs *coeff = &coeffs[g];
330
            float bark_width = coeffs[g+1].barks - coeffs->barks;
331
            coeff->spread_low[0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_LOW);
332
            coeff->spread_hi [0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_HI);
333
            coeff->spread_low[1] = pow(10.0, -bark_width * en_spread_low);
334
            coeff->spread_hi [1] = pow(10.0, -bark_width * en_spread_hi);
335
            pe_min = bark_pe * bark_width;
336
            minsnr = pow(2.0f, pe_min / band_sizes[g]) - 1.5f;
337
            coeff->min_snr = av_clipf(1.0f / minsnr, PSY_SNR_25DB, PSY_SNR_1DB);
338
        }
339
        start = 0;
340
        for (g = 0; g < ctx->num_bands[j]; g++) {
341
            minscale = ath(start * line_to_frequency, ATH_ADD);
342
            for (i = 1; i < band_sizes[g]; i++)
343
                minscale = FFMIN(minscale, ath((start + i) * line_to_frequency, ATH_ADD));
344
            coeffs[g].ath = minscale - minath;
345
            start += band_sizes[g];
346
        }
347
    }
348

    
349
    pctx->ch = av_mallocz(sizeof(AacPsyChannel) * ctx->avctx->channels);
350

    
351
    lame_window_init(pctx, ctx->avctx);
352

    
353
    return 0;
354
}
355

    
356
/**
357
 * IIR filter used in block switching decision
358
 */
359
static float iir_filter(int in, float state[2])
360
{
361
    float ret;
362

    
363
    ret = 0.7548f * (in - state[0]) + 0.5095f * state[1];
364
    state[0] = in;
365
    state[1] = ret;
366
    return ret;
367
}
368

    
369
/**
370
 * window grouping information stored as bits (0 - new group, 1 - group continues)
371
 */
372
static const uint8_t window_grouping[9] = {
373
    0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36
374
};
375

    
376
/**
377
 * Tell encoder which window types to use.
378
 * @see 3GPP TS26.403 5.4.1 "Blockswitching"
379
 */
380
static FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx,
381
                                       const int16_t *audio, const int16_t *la,
382
                                       int channel, int prev_type)
383
{
384
    int i, j;
385
    int br               = ctx->avctx->bit_rate / ctx->avctx->channels;
386
    int attack_ratio     = br <= 16000 ? 18 : 10;
387
    AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
388
    AacPsyChannel *pch  = &pctx->ch[channel];
389
    uint8_t grouping     = 0;
390
    int next_type        = pch->next_window_seq;
391
    FFPsyWindowInfo wi;
392

    
393
    memset(&wi, 0, sizeof(wi));
394
    if (la) {
395
        float s[8], v;
396
        int switch_to_eight = 0;
397
        float sum = 0.0, sum2 = 0.0;
398
        int attack_n = 0;
399
        int stay_short = 0;
400
        for (i = 0; i < 8; i++) {
401
            for (j = 0; j < 128; j++) {
402
                v = iir_filter(la[(i*128+j)*ctx->avctx->channels], pch->iir_state);
403
                sum += v*v;
404
            }
405
            s[i]  = sum;
406
            sum2 += sum;
407
        }
408
        for (i = 0; i < 8; i++) {
409
            if (s[i] > pch->win_energy * attack_ratio) {
410
                attack_n        = i + 1;
411
                switch_to_eight = 1;
412
                break;
413
            }
414
        }
415
        pch->win_energy = pch->win_energy*7/8 + sum2/64;
416

    
417
        wi.window_type[1] = prev_type;
418
        switch (prev_type) {
419
        case ONLY_LONG_SEQUENCE:
420
            wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
421
            next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
422
            break;
423
        case LONG_START_SEQUENCE:
424
            wi.window_type[0] = EIGHT_SHORT_SEQUENCE;
425
            grouping = pch->next_grouping;
426
            next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
427
            break;
428
        case LONG_STOP_SEQUENCE:
429
            wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
430
            next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
431
            break;
432
        case EIGHT_SHORT_SEQUENCE:
433
            stay_short = next_type == EIGHT_SHORT_SEQUENCE || switch_to_eight;
434
            wi.window_type[0] = stay_short ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
435
            grouping = next_type == EIGHT_SHORT_SEQUENCE ? pch->next_grouping : 0;
436
            next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
437
            break;
438
        }
439

    
440
        pch->next_grouping = window_grouping[attack_n];
441
        pch->next_window_seq = next_type;
442
    } else {
443
        for (i = 0; i < 3; i++)
444
            wi.window_type[i] = prev_type;
445
        grouping = (prev_type == EIGHT_SHORT_SEQUENCE) ? window_grouping[0] : 0;
446
    }
447

    
448
    wi.window_shape   = 1;
449
    if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
450
        wi.num_windows = 1;
451
        wi.grouping[0] = 1;
452
    } else {
453
        int lastgrp = 0;
454
        wi.num_windows = 8;
455
        for (i = 0; i < 8; i++) {
456
            if (!((grouping >> i) & 1))
457
                lastgrp = i;
458
            wi.grouping[lastgrp]++;
459
        }
460
    }
461

    
462
    return wi;
463
}
464

    
465
/* 5.6.1.2 "Calculation of Bit Demand" */
466
static int calc_bit_demand(AacPsyContext *ctx, float pe, int bits, int size,
467
                           int short_window)
468
{
469
    const float bitsave_slope  = short_window ? PSY_3GPP_SAVE_SLOPE_S  : PSY_3GPP_SAVE_SLOPE_L;
470
    const float bitsave_add    = short_window ? PSY_3GPP_SAVE_ADD_S    : PSY_3GPP_SAVE_ADD_L;
471
    const float bitspend_slope = short_window ? PSY_3GPP_SPEND_SLOPE_S : PSY_3GPP_SPEND_SLOPE_L;
472
    const float bitspend_add   = short_window ? PSY_3GPP_SPEND_ADD_S   : PSY_3GPP_SPEND_ADD_L;
473
    const float clip_low       = short_window ? PSY_3GPP_CLIP_LO_S     : PSY_3GPP_CLIP_LO_L;
474
    const float clip_high      = short_window ? PSY_3GPP_CLIP_HI_S     : PSY_3GPP_CLIP_HI_L;
475
    float clipped_pe, bit_save, bit_spend, bit_factor, fill_level;
476

    
477
    ctx->fill_level += ctx->frame_bits - bits;
478
    ctx->fill_level  = av_clip(ctx->fill_level, 0, size);
479
    fill_level = av_clipf((float)ctx->fill_level / size, clip_low, clip_high);
480
    clipped_pe = av_clipf(pe, ctx->pe.min, ctx->pe.max);
481
    bit_save   = (fill_level + bitsave_add) * bitsave_slope;
482
    assert(bit_save <= 0.3f && bit_save >= -0.05000001f);
483
    bit_spend  = (fill_level + bitspend_add) * bitspend_slope;
484
    assert(bit_spend <= 0.5f && bit_spend >= -0.1f);
485
    /* The bit factor graph in the spec is obviously incorrect.
486
     *      bit_spend + ((bit_spend - bit_spend))...
487
     * The reference encoder subtracts everything from 1, but also seems incorrect.
488
     *      1 - bit_save + ((bit_spend + bit_save))...
489
     * Hopefully below is correct.
490
     */
491
    bit_factor = 1.0f - bit_save + ((bit_spend - bit_save) / (ctx->pe.max - ctx->pe.min)) * (clipped_pe - ctx->pe.min);
492
    /* NOTE: The reference encoder attempts to center pe max/min around the current pe. */
493
    ctx->pe.max = FFMAX(pe, ctx->pe.max);
494
    ctx->pe.min = FFMIN(pe, ctx->pe.min);
495

    
496
    return FFMIN(ctx->frame_bits * bit_factor, ctx->frame_bits + size - bits);
497
}
498

    
499
static float calc_pe_3gpp(AacPsyBand *band)
500
{
501
    float pe, a;
502

    
503
    band->pe           = 0.0f;
504
    band->pe_const     = 0.0f;
505
    band->active_lines = 0.0f;
506
    if (band->energy > band->thr) {
507
        a  = log2f(band->energy);
508
        pe = a - log2f(band->thr);
509
        band->active_lines = band->nz_lines;
510
        if (pe < PSY_3GPP_C1) {
511
            pe = pe * PSY_3GPP_C3 + PSY_3GPP_C2;
512
            a  = a  * PSY_3GPP_C3 + PSY_3GPP_C2;
513
            band->active_lines *= PSY_3GPP_C3;
514
        }
515
        band->pe       = pe * band->nz_lines;
516
        band->pe_const = a  * band->nz_lines;
517
    }
518

    
519
    return band->pe;
520
}
521

    
522
static float calc_reduction_3gpp(float a, float desired_pe, float pe,
523
                                 float active_lines)
524
{
525
    float thr_avg, reduction;
526

    
527
    thr_avg   = powf(2.0f, (a - pe) / (4.0f * active_lines));
528
    reduction = powf(2.0f, (a - desired_pe) / (4.0f * active_lines)) - thr_avg;
529

    
530
    return FFMAX(reduction, 0.0f);
531
}
532

    
533
static float calc_reduced_thr_3gpp(AacPsyBand *band, float min_snr,
534
                                   float reduction)
535
{
536
    float thr = band->thr;
537

    
538
    if (band->energy > thr) {
539
        thr = powf(thr, 0.25f) + reduction;
540
        thr = powf(thr, 4.0f);
541

    
542
        /* This deviates from the 3GPP spec to match the reference encoder.
543
         * It performs min(thr_reduced, max(thr, energy/min_snr)) only for bands
544
         * that have hole avoidance on (active or inactive). It always reduces the
545
         * threshold of bands with hole avoidance off.
546
         */
547
        if (thr > band->energy * min_snr && band->avoid_holes != PSY_3GPP_AH_NONE) {
548
            thr = FFMAX(band->thr, band->energy * min_snr);
549
            band->avoid_holes = PSY_3GPP_AH_ACTIVE;
550
        }
551
    }
552

    
553
    return thr;
554
}
555

    
556
/**
557
 * Calculate band thresholds as suggested in 3GPP TS26.403
558
 */
559
static void psy_3gpp_analyze(FFPsyContext *ctx, int channel,
560
                             const float *coefs, const FFPsyWindowInfo *wi)
561
{
562
    AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
563
    AacPsyChannel *pch  = &pctx->ch[channel];
564
    int start = 0;
565
    int i, w, g;
566
    float desired_bits, desired_pe, delta_pe, reduction, spread_en[128] = {0};
567
    float a = 0.0f, active_lines = 0.0f, norm_fac = 0.0f;
568
    float pe = pctx->chan_bitrate > 32000 ? 0.0f : FFMAX(50.0f, 100.0f - pctx->chan_bitrate * 100.0f / 32000.0f);
569
    const int      num_bands   = ctx->num_bands[wi->num_windows == 8];
570
    const uint8_t *band_sizes  = ctx->bands[wi->num_windows == 8];
571
    AacPsyCoeffs  *coeffs      = pctx->psy_coef[wi->num_windows == 8];
572
    const float avoid_hole_thr = wi->num_windows == 8 ? PSY_3GPP_AH_THR_SHORT : PSY_3GPP_AH_THR_LONG;
573

    
574
    //calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation"
575
    for (w = 0; w < wi->num_windows*16; w += 16) {
576
        for (g = 0; g < num_bands; g++) {
577
            AacPsyBand *band = &pch->band[w+g];
578

    
579
            float form_factor = 0.0f;
580
            band->energy = 0.0f;
581
            for (i = 0; i < band_sizes[g]; i++) {
582
                band->energy += coefs[start+i] * coefs[start+i];
583
                form_factor  += sqrtf(fabs(coefs[start+i]));
584
            }
585
            band->thr      = band->energy * 0.001258925f;
586
            band->nz_lines = form_factor / powf(band->energy / band_sizes[g], 0.25f);
587

    
588
            start += band_sizes[g];
589
        }
590
    }
591
    //modify thresholds and energies - spread, threshold in quiet, pre-echo control
592
    for (w = 0; w < wi->num_windows*16; w += 16) {
593
        AacPsyBand *bands = &pch->band[w];
594

    
595
        //5.4.2.3 "Spreading" & 5.4.3 "Spreaded Energy Calculation"
596
        spread_en[0] = bands[0].energy;
597
        for (g = 1; g < num_bands; g++) {
598
            bands[g].thr   = FFMAX(bands[g].thr,    bands[g-1].thr * coeffs[g].spread_hi[0]);
599
            spread_en[w+g] = FFMAX(bands[g].energy, spread_en[w+g-1] * coeffs[g].spread_hi[1]);
600
        }
601
        for (g = num_bands - 2; g >= 0; g--) {
602
            bands[g].thr   = FFMAX(bands[g].thr,   bands[g+1].thr * coeffs[g].spread_low[0]);
603
            spread_en[w+g] = FFMAX(spread_en[w+g], spread_en[w+g+1] * coeffs[g].spread_low[1]);
604
        }
605
        //5.4.2.4 "Threshold in quiet"
606
        for (g = 0; g < num_bands; g++) {
607
            AacPsyBand *band = &bands[g];
608

    
609
            band->thr_quiet = band->thr = FFMAX(band->thr, coeffs[g].ath);
610
            //5.4.2.5 "Pre-echo control"
611
            if (!(wi->window_type[0] == LONG_STOP_SEQUENCE || (wi->window_type[1] == LONG_START_SEQUENCE && !w)))
612
                band->thr = FFMAX(PSY_3GPP_RPEMIN*band->thr, FFMIN(band->thr,
613
                                  PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet));
614

    
615
            /* 5.6.1.3.1 "Prepatory steps of the perceptual entropy calculation" */
616
            pe += calc_pe_3gpp(band);
617
            a  += band->pe_const;
618
            active_lines += band->active_lines;
619

    
620
            /* 5.6.1.3.3 "Selection of the bands for avoidance of holes" */
621
            if (spread_en[w+g] * avoid_hole_thr > band->energy || coeffs[g].min_snr > 1.0f)
622
                band->avoid_holes = PSY_3GPP_AH_NONE;
623
            else
624
                band->avoid_holes = PSY_3GPP_AH_INACTIVE;
625
        }
626
    }
627

    
628
    /* 5.6.1.3.2 "Calculation of the desired perceptual entropy" */
629
    ctx->pe[channel] = pe;
630
    desired_bits = calc_bit_demand(pctx, pe, ctx->bitres.bits, ctx->bitres.size, wi->num_windows == 8);
631
    desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits);
632
    /* NOTE: PE correction is kept simple. During initial testing it had very
633
     *       little effect on the final bitrate. Probably a good idea to come
634
     *       back and do more testing later.
635
     */
636
    if (ctx->bitres.bits > 0)
637
        desired_pe *= av_clipf(pctx->pe.previous / PSY_3GPP_BITS_TO_PE(ctx->bitres.bits),
638
                               0.85f, 1.15f);
639
    pctx->pe.previous = PSY_3GPP_BITS_TO_PE(desired_bits);
640

    
641
    if (desired_pe < pe) {
642
        /* 5.6.1.3.4 "First Estimation of the reduction value" */
643
        for (w = 0; w < wi->num_windows*16; w += 16) {
644
            reduction = calc_reduction_3gpp(a, desired_pe, pe, active_lines);
645
            pe = 0.0f;
646
            a  = 0.0f;
647
            active_lines = 0.0f;
648
            for (g = 0; g < num_bands; g++) {
649
                AacPsyBand *band = &pch->band[w+g];
650

    
651
                band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
652
                /* recalculate PE */
653
                pe += calc_pe_3gpp(band);
654
                a  += band->pe_const;
655
                active_lines += band->active_lines;
656
            }
657
        }
658

    
659
        /* 5.6.1.3.5 "Second Estimation of the reduction value" */
660
        for (i = 0; i < 2; i++) {
661
            float pe_no_ah = 0.0f, desired_pe_no_ah;
662
            active_lines = a = 0.0f;
663
            for (w = 0; w < wi->num_windows*16; w += 16) {
664
                for (g = 0; g < num_bands; g++) {
665
                    AacPsyBand *band = &pch->band[w+g];
666

    
667
                    if (band->avoid_holes != PSY_3GPP_AH_ACTIVE) {
668
                        pe_no_ah += band->pe;
669
                        a        += band->pe_const;
670
                        active_lines += band->active_lines;
671
                    }
672
                }
673
            }
674
            desired_pe_no_ah = FFMAX(desired_pe - (pe - pe_no_ah), 0.0f);
675
            if (active_lines > 0.0f)
676
                reduction += calc_reduction_3gpp(a, desired_pe_no_ah, pe_no_ah, active_lines);
677

    
678
            pe = 0.0f;
679
            for (w = 0; w < wi->num_windows*16; w += 16) {
680
                for (g = 0; g < num_bands; g++) {
681
                    AacPsyBand *band = &pch->band[w+g];
682

    
683
                    if (active_lines > 0.0f)
684
                        band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
685
                    pe += calc_pe_3gpp(band);
686
                    band->norm_fac = band->active_lines / band->thr;
687
                    norm_fac += band->norm_fac;
688
                }
689
            }
690
            delta_pe = desired_pe - pe;
691
            if (fabs(delta_pe) > 0.05f * desired_pe)
692
                break;
693
        }
694

    
695
        if (pe < 1.15f * desired_pe) {
696
            /* 6.6.1.3.6 "Final threshold modification by linearization" */
697
            norm_fac = 1.0f / norm_fac;
698
            for (w = 0; w < wi->num_windows*16; w += 16) {
699
                for (g = 0; g < num_bands; g++) {
700
                    AacPsyBand *band = &pch->band[w+g];
701

    
702
                    if (band->active_lines > 0.5f) {
703
                        float delta_sfb_pe = band->norm_fac * norm_fac * delta_pe;
704
                        float thr = band->thr;
705

    
706
                        thr *= powf(2.0f, delta_sfb_pe / band->active_lines);
707
                        if (thr > coeffs[g].min_snr * band->energy && band->avoid_holes == PSY_3GPP_AH_INACTIVE)
708
                            thr = FFMAX(band->thr, coeffs[g].min_snr * band->energy);
709
                        band->thr = thr;
710
                    }
711
                }
712
            }
713
        } else {
714
            /* 5.6.1.3.7 "Further perceptual entropy reduction" */
715
            g = num_bands;
716
            while (pe > desired_pe && g--) {
717
                for (w = 0; w < wi->num_windows*16; w+= 16) {
718
                    AacPsyBand *band = &pch->band[w+g];
719
                    if (band->avoid_holes != PSY_3GPP_AH_NONE && coeffs[g].min_snr < PSY_SNR_1DB) {
720
                        coeffs[g].min_snr = PSY_SNR_1DB;
721
                        band->thr = band->energy * PSY_SNR_1DB;
722
                        pe += band->active_lines * 1.5f - band->pe;
723
                    }
724
                }
725
            }
726
            /* TODO: allow more holes (unused without mid/side) */
727
        }
728
    }
729

    
730
    for (w = 0; w < wi->num_windows*16; w += 16) {
731
        for (g = 0; g < num_bands; g++) {
732
            AacPsyBand *band     = &pch->band[w+g];
733
            FFPsyBand  *psy_band = &ctx->psy_bands[channel*PSY_MAX_BANDS+w+g];
734

    
735
            psy_band->threshold = band->thr;
736
            psy_band->energy    = band->energy;
737
        }
738
    }
739

    
740
    memcpy(pch->prev_band, pch->band, sizeof(pch->band));
741
}
742

    
743
static av_cold void psy_3gpp_end(FFPsyContext *apc)
744
{
745
    AacPsyContext *pctx = (AacPsyContext*) apc->model_priv_data;
746
    av_freep(&pctx->ch);
747
    av_freep(&apc->model_priv_data);
748
}
749

    
750
static void lame_apply_block_type(AacPsyChannel *ctx, FFPsyWindowInfo *wi, int uselongblock)
751
{
752
    int blocktype = ONLY_LONG_SEQUENCE;
753
    if (uselongblock) {
754
        if (ctx->next_window_seq == EIGHT_SHORT_SEQUENCE)
755
            blocktype = LONG_STOP_SEQUENCE;
756
    } else {
757
        blocktype = EIGHT_SHORT_SEQUENCE;
758
        if (ctx->next_window_seq == ONLY_LONG_SEQUENCE)
759
            ctx->next_window_seq = LONG_START_SEQUENCE;
760
        if (ctx->next_window_seq == LONG_STOP_SEQUENCE)
761
            ctx->next_window_seq = EIGHT_SHORT_SEQUENCE;
762
    }
763

    
764
    wi->window_type[0] = ctx->next_window_seq;
765
    ctx->next_window_seq = blocktype;
766
}
767

    
768
static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx,
769
                                       const int16_t *audio, const int16_t *la,
770
                                       int channel, int prev_type)
771
{
772
    AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
773
    AacPsyChannel *pch  = &pctx->ch[channel];
774
    int grouping     = 0;
775
    int uselongblock = 1;
776
    int attacks[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
777
    int i;
778
    FFPsyWindowInfo wi;
779

    
780
    memset(&wi, 0, sizeof(wi));
781
    if (la) {
782
        float hpfsmpl[AAC_BLOCK_SIZE_LONG];
783
        float const *pf = hpfsmpl;
784
        float attack_intensity[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
785
        float energy_subshort[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
786
        float energy_short[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
787
        int chans = ctx->avctx->channels;
788
        const int16_t *firbuf = la + (AAC_BLOCK_SIZE_SHORT/4 - PSY_LAME_FIR_LEN) * chans;
789
        int j, att_sum = 0;
790

    
791
        /* LAME comment: apply high pass filter of fs/4 */
792
        for (i = 0; i < AAC_BLOCK_SIZE_LONG; i++) {
793
            float sum1, sum2;
794
            sum1 = firbuf[(i + ((PSY_LAME_FIR_LEN - 1) / 2)) * chans];
795
            sum2 = 0.0;
796
            for (j = 0; j < ((PSY_LAME_FIR_LEN - 1) / 2) - 1; j += 2) {
797
                sum1 += psy_fir_coeffs[j] * (firbuf[(i + j) * chans] + firbuf[(i + PSY_LAME_FIR_LEN - j) * chans]);
798
                sum2 += psy_fir_coeffs[j + 1] * (firbuf[(i + j + 1) * chans] + firbuf[(i + PSY_LAME_FIR_LEN - j - 1) * chans]);
799
            }
800
            hpfsmpl[i] = sum1 + sum2;
801
        }
802

    
803
        /* Calculate the energies of each sub-shortblock */
804
        for (i = 0; i < PSY_LAME_NUM_SUBBLOCKS; i++) {
805
            energy_subshort[i] = pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) * PSY_LAME_NUM_SUBBLOCKS)];
806
            assert(pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)] > 0);
807
            attack_intensity[i] = energy_subshort[i] / pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)];
808
            energy_short[0] += energy_subshort[i];
809
        }
810

    
811
        for (i = 0; i < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; i++) {
812
            float const *const pfe = pf + AAC_BLOCK_SIZE_LONG / (AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS);
813
            float p = 1.0f;
814
            for (; pf < pfe; pf++)
815
                if (p < fabsf(*pf))
816
                    p = fabsf(*pf);
817
            pch->prev_energy_subshort[i] = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS] = p;
818
            energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p;
819
            /* FIXME: The indexes below are [i + 3 - 2] in the LAME source.
820
             *          Obviously the 3 and 2 have some significance, or this would be just [i + 1]
821
             *          (which is what we use here). What the 3 stands for is ambigious, as it is both
822
             *          number of short blocks, and the number of sub-short blocks.
823
             *          It seems that LAME is comparing each sub-block to sub-block + 1 in the
824
             *          previous block.
825
             */
826
            if (p > energy_subshort[i + 1])
827
                p = p / energy_subshort[i + 1];
828
            else if (energy_subshort[i + 1] > p * 10.0f)
829
                p = energy_subshort[i + 1] / (p * 10.0f);
830
            else
831
                p = 0.0;
832
            attack_intensity[i + PSY_LAME_NUM_SUBBLOCKS] = p;
833
        }
834

    
835
        /* compare energy between sub-short blocks */
836
        for (i = 0; i < (AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS; i++)
837
            if (!attacks[i / PSY_LAME_NUM_SUBBLOCKS])
838
                if (attack_intensity[i] > pch->attack_threshold)
839
                    attacks[i / PSY_LAME_NUM_SUBBLOCKS] = (i % PSY_LAME_NUM_SUBBLOCKS) + 1;
840

    
841
        /* should have energy change between short blocks, in order to avoid periodic signals */
842
        /* Good samples to show the effect are Trumpet test songs */
843
        /* GB: tuned (1) to avoid too many short blocks for test sample TRUMPET */
844
        /* RH: tuned (2) to let enough short blocks through for test sample FSOL and SNAPS */
845
        for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) {
846
            float const u = energy_short[i - 1];
847
            float const v = energy_short[i];
848
            float const m = FFMAX(u, v);
849
            if (m < 40000) {                          /* (2) */
850
                if (u < 1.7f * v && v < 1.7f * u) {   /* (1) */
851
                    if (i == 1 && attacks[0] < attacks[i])
852
                        attacks[0] = 0;
853
                    attacks[i] = 0;
854
                }
855
            }
856
            att_sum += attacks[i];
857
        }
858

    
859
        if (attacks[0] <= pch->prev_attack)
860
            attacks[0] = 0;
861

    
862
        att_sum += attacks[0];
863
        /* 3 below indicates the previous attack happened in the last sub-block of the previous sequence */
864
        if (pch->prev_attack == 3 || att_sum) {
865
            uselongblock = 0;
866

    
867
            for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++)
868
                if (attacks[i] && attacks[i-1])
869
                    attacks[i] = 0;
870
        }
871
    } else {
872
        /* We have no lookahead info, so just use same type as the previous sequence. */
873
        uselongblock = !(prev_type == EIGHT_SHORT_SEQUENCE);
874
    }
875

    
876
    lame_apply_block_type(pch, &wi, uselongblock);
877

    
878
    wi.window_type[1] = prev_type;
879
    if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
880
        wi.num_windows  = 1;
881
        wi.grouping[0]  = 1;
882
        if (wi.window_type[0] == LONG_START_SEQUENCE)
883
            wi.window_shape = 0;
884
        else
885
            wi.window_shape = 1;
886
    } else {
887
        int lastgrp = 0;
888

    
889
        wi.num_windows = 8;
890
        wi.window_shape = 0;
891
        for (i = 0; i < 8; i++) {
892
            if (!((pch->next_grouping >> i) & 1))
893
                lastgrp = i;
894
            wi.grouping[lastgrp]++;
895
        }
896
    }
897

    
898
    /* Determine grouping, based on the location of the first attack, and save for
899
     * the next frame.
900
     * FIXME: Move this to analysis.
901
     * TODO: Tune groupings depending on attack location
902
     * TODO: Handle more than one attack in a group
903
     */
904
    for (i = 0; i < 9; i++) {
905
        if (attacks[i]) {
906
            grouping = i;
907
            break;
908
        }
909
    }
910
    pch->next_grouping = window_grouping[grouping];
911

    
912
    pch->prev_attack = attacks[8];
913

    
914
    return wi;
915
}
916

    
917
const FFPsyModel ff_aac_psy_model =
918
{
919
    .name    = "3GPP TS 26.403-inspired model",
920
    .init    = psy_3gpp_init,
921
    .window  = psy_lame_window,
922
    .analyze = psy_3gpp_analyze,
923
    .end     = psy_3gpp_end,
924
};