FFmpeg
aacenc_tns.c
Go to the documentation of this file.
1 /*
2  * AAC encoder TNS
3  * Copyright (C) 2015 Rostislav Pehlivanov
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg 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  * FFmpeg 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 FFmpeg; 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 temporal noise shaping
25  * @author Rostislav Pehlivanov ( atomnuker gmail com )
26  */
27 
28 #include "libavutil/libm.h"
29 #include "aacenc.h"
30 #include "aacenc_tns.h"
31 #include "aactab.h"
32 #include "aacenc_utils.h"
33 #include "aacenc_quantization.h"
34 
35 /* Could be set to 3 to save an additional bit at the cost of little quality */
36 #define TNS_Q_BITS 4
37 
38 /* Coefficient resolution in short windows */
39 #define TNS_Q_BITS_IS8 4
40 
41 /* We really need the bits we save here elsewhere */
42 #define TNS_ENABLE_COEF_COMPRESSION
43 
44 /* TNS will only be used if the LPC gain is within these margins */
45 #define TNS_GAIN_THRESHOLD_LOW 1.4f
46 #define TNS_GAIN_THRESHOLD_HIGH 1.16f*TNS_GAIN_THRESHOLD_LOW
47 
48 static inline int compress_coeffs(int *coef, int order, int c_bits)
49 {
50  int i;
51  const int low_idx = c_bits ? 4 : 2;
52  const int shift_val = c_bits ? 8 : 4;
53  const int high_idx = c_bits ? 11 : 5;
54 #ifndef TNS_ENABLE_COEF_COMPRESSION
55  return 0;
56 #endif /* TNS_ENABLE_COEF_COMPRESSION */
57  for (i = 0; i < order; i++)
58  if (coef[i] >= low_idx && coef[i] <= high_idx)
59  return 0;
60  for (i = 0; i < order; i++)
61  coef[i] -= (coef[i] > high_idx) ? shift_val : 0;
62  return 1;
63 }
64 
65 /**
66  * Encode TNS data.
67  * Coefficient compression is simply not lossless as it should be
68  * on any decoder tested and as such is not active.
69  */
71 {
72  TemporalNoiseShaping *tns = &sce->tns;
73  int i, w, filt, coef_compress = 0, coef_len;
74  const int is8 = sce->ics.window_sequence[0] == EIGHT_SHORT_SEQUENCE;
75  const int c_bits = is8 ? TNS_Q_BITS_IS8 == 4 : TNS_Q_BITS == 4;
76 
77  if (!sce->tns.present)
78  return;
79 
80  for (i = 0; i < sce->ics.num_windows; i++) {
81  put_bits(&s->pb, 2 - is8, sce->tns.n_filt[i]);
82  if (!tns->n_filt[i])
83  continue;
84  put_bits(&s->pb, 1, c_bits);
85  for (filt = 0; filt < tns->n_filt[i]; filt++) {
86  put_bits(&s->pb, 6 - 2 * is8, tns->length[i][filt]);
87  put_bits(&s->pb, 5 - 2 * is8, tns->order[i][filt]);
88  if (!tns->order[i][filt])
89  continue;
90  put_bits(&s->pb, 1, tns->direction[i][filt]);
91  coef_compress = compress_coeffs(tns->coef_idx[i][filt],
92  tns->order[i][filt], c_bits);
93  put_bits(&s->pb, 1, coef_compress);
94  coef_len = c_bits + 3 - coef_compress;
95  for (w = 0; w < tns->order[i][filt]; w++)
96  put_bits(&s->pb, coef_len, tns->coef_idx[i][filt][w]);
97  }
98  }
99 }
100 
101 /* Apply TNS filter */
103 {
104  TemporalNoiseShaping *tns = &sce->tns;
105  IndividualChannelStream *ics = &sce->ics;
106  int w, filt, m, i, top, order, bottom, start, end, size, inc;
107  const int mmm = FFMIN(ics->tns_max_bands, ics->max_sfb);
108  float lpc[TNS_MAX_ORDER];
109 
110  for (w = 0; w < ics->num_windows; w++) {
111  bottom = ics->num_swb;
112  for (filt = 0; filt < tns->n_filt[w]; filt++) {
113  top = bottom;
114  bottom = FFMAX(0, top - tns->length[w][filt]);
115  order = tns->order[w][filt];
116  if (order == 0)
117  continue;
118 
119  // tns_decode_coef
120  compute_lpc_coefs(tns->coef[w][filt], order, lpc, 0, 0, 0);
121 
122  start = ics->swb_offset[FFMIN(bottom, mmm)];
123  end = ics->swb_offset[FFMIN( top, mmm)];
124  if ((size = end - start) <= 0)
125  continue;
126  if (tns->direction[w][filt]) {
127  inc = -1;
128  start = end - 1;
129  } else {
130  inc = 1;
131  }
132  start += w * 128;
133 
134  /* AR filter */
135  for (m = 0; m < size; m++, start += inc) {
136  for (i = 1; i <= FFMIN(m, order); i++) {
137  sce->coeffs[start] += lpc[i-1]*sce->pcoeffs[start - i*inc];
138  }
139  }
140  }
141  }
142 }
143 
144 /*
145  * c_bits - 1 if 4 bit coefficients, 0 if 3 bit coefficients
146  */
147 static inline void quantize_coefs(double *coef, int *idx, float *lpc, int order,
148  int c_bits)
149 {
150  int i;
151  const float *quant_arr = tns_tmp2_map[c_bits];
152  for (i = 0; i < order; i++) {
153  idx[i] = quant_array_idx(coef[i], quant_arr, c_bits ? 16 : 8);
154  lpc[i] = quant_arr[idx[i]];
155  }
156 }
157 
158 /*
159  * 3 bits per coefficient with 8 short windows
160  */
162 {
163  TemporalNoiseShaping *tns = &sce->tns;
164  int w, g, count = 0;
165  double gain, coefs[MAX_LPC_ORDER];
166  const int mmm = FFMIN(sce->ics.tns_max_bands, sce->ics.max_sfb);
167  const int is8 = sce->ics.window_sequence[0] == EIGHT_SHORT_SEQUENCE;
168  const int c_bits = is8 ? TNS_Q_BITS_IS8 == 4 : TNS_Q_BITS == 4;
169  const int sfb_start = av_clip(tns_min_sfb[is8][s->samplerate_index], 0, mmm);
170  const int sfb_end = av_clip(sce->ics.num_swb, 0, mmm);
171  const int order = is8 ? 7 : s->profile == FF_PROFILE_AAC_LOW ? 12 : TNS_MAX_ORDER;
172  const int slant = sce->ics.window_sequence[0] == LONG_STOP_SEQUENCE ? 1 :
173  sce->ics.window_sequence[0] == LONG_START_SEQUENCE ? 0 : 2;
174  const int sfb_len = sfb_end - sfb_start;
175  const int coef_len = sce->ics.swb_offset[sfb_end] - sce->ics.swb_offset[sfb_start];
176 
177  if (coef_len <= 0 || sfb_len <= 0) {
178  sce->tns.present = 0;
179  return;
180  }
181 
182  for (w = 0; w < sce->ics.num_windows; w++) {
183  float en[2] = {0.0f, 0.0f};
184  int oc_start = 0, os_start = 0;
185  int coef_start = sce->ics.swb_offset[sfb_start];
186 
187  for (g = sfb_start; g < sce->ics.num_swb && g <= sfb_end; g++) {
188  FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[w*16+g];
189  if (g > sfb_start + (sfb_len/2))
190  en[1] += band->energy;
191  else
192  en[0] += band->energy;
193  }
194 
195  /* LPC */
196  gain = ff_lpc_calc_ref_coefs_f(&s->lpc, &sce->coeffs[w*128 + coef_start],
197  coef_len, order, coefs);
198 
199  if (!order || !isfinite(gain) || gain < TNS_GAIN_THRESHOLD_LOW || gain > TNS_GAIN_THRESHOLD_HIGH)
200  continue;
201 
202  tns->n_filt[w] = is8 ? 1 : order != TNS_MAX_ORDER ? 2 : 3;
203  for (g = 0; g < tns->n_filt[w]; g++) {
204  tns->direction[w][g] = slant != 2 ? slant : en[g] < en[!g];
205  tns->order[w][g] = g < tns->n_filt[w] ? order/tns->n_filt[w] : order - oc_start;
206  tns->length[w][g] = g < tns->n_filt[w] ? sfb_len/tns->n_filt[w] : sfb_len - os_start;
207  quantize_coefs(&coefs[oc_start], tns->coef_idx[w][g], tns->coef[w][g],
208  tns->order[w][g], c_bits);
209  oc_start += tns->order[w][g];
210  os_start += tns->length[w][g];
211  }
212  count++;
213  }
214  sce->tns.present = !!count;
215 }
av_clip
#define av_clip
Definition: common.h:122
TNS_Q_BITS_IS8
#define TNS_Q_BITS_IS8
Definition: aacenc_tns.c:39
TemporalNoiseShaping::order
int order[8][4]
Definition: aac.h:204
libm.h
TemporalNoiseShaping::direction
int direction[8][4]
Definition: aac.h:203
TNS_GAIN_THRESHOLD_HIGH
#define TNS_GAIN_THRESHOLD_HIGH
Definition: aacenc_tns.c:46
TemporalNoiseShaping::coef_idx
int coef_idx[8][4][TNS_MAX_ORDER]
Definition: aac.h:205
put_bits
static void put_bits(Jpeg2000EncoderContext *s, int val, int n)
put n times val bit
Definition: j2kenc.c:218
w
uint8_t w
Definition: llviddspenc.c:39
compute_lpc_coefs
static int AAC_RENAME() compute_lpc_coefs(const LPC_TYPE *autoc, int max_order, LPC_TYPE *lpc, int lpc_stride, int fail, int normalize)
Levinson-Durbin recursion.
Definition: lpc.h:166
compress_coeffs
static int compress_coeffs(int *coef, int order, int c_bits)
Definition: aacenc_tns.c:48
TemporalNoiseShaping::present
int present
Definition: aac.h:200
SingleChannelElement::pcoeffs
INTFLOAT pcoeffs[1024]
coefficients for IMDCT, pristine
Definition: aac.h:262
IndividualChannelStream::num_swb
int num_swb
number of scalefactor window bands
Definition: aac.h:184
TNS_Q_BITS
#define TNS_Q_BITS
Definition: aacenc_tns.c:36
TemporalNoiseShaping::length
int length[8][4]
Definition: aac.h:202
SingleChannelElement::ics
IndividualChannelStream ics
Definition: aac.h:250
s
#define s(width, name)
Definition: cbs_vp9.c:257
SingleChannelElement::coeffs
INTFLOAT coeffs[1024]
coefficients for IMDCT, maybe processed
Definition: aac.h:263
g
const char * g
Definition: vf_curves.c:117
EIGHT_SHORT_SEQUENCE
@ EIGHT_SHORT_SEQUENCE
Definition: aac.h:79
isfinite
#define isfinite(x)
Definition: libm.h:359
IndividualChannelStream
Individual Channel Stream.
Definition: aac.h:175
TemporalNoiseShaping::coef
INTFLOAT coef[8][4][TNS_MAX_ORDER]
Definition: aac.h:206
IndividualChannelStream::swb_offset
const uint16_t * swb_offset
table of offsets to the lowest spectral coefficient of a scalefactor band, sfb, for a particular wind...
Definition: aac.h:182
aacenc_quantization.h
ff_aac_apply_tns
void ff_aac_apply_tns(AACEncContext *s, SingleChannelElement *sce)
Definition: aacenc_tns.c:102
FFPsyBand
single band psychoacoustic information
Definition: psymodel.h:50
aactab.h
ff_aac_encode_tns_info
void ff_aac_encode_tns_info(AACEncContext *s, SingleChannelElement *sce)
Encode TNS data.
Definition: aacenc_tns.c:70
TNS_MAX_ORDER
#define TNS_MAX_ORDER
Definition: aac.h:51
ff_lpc_calc_ref_coefs_f
double ff_lpc_calc_ref_coefs_f(LPCContext *s, const float *samples, int len, int order, double *ref)
Definition: lpc.c:172
TemporalNoiseShaping::n_filt
int n_filt[8]
Definition: aac.h:201
FF_PROFILE_AAC_LOW
#define FF_PROFILE_AAC_LOW
Definition: avcodec.h:1863
MAX_LPC_ORDER
#define MAX_LPC_ORDER
Definition: lpc.h:38
FFMAX
#define FFMAX(a, b)
Definition: common.h:103
size
int size
Definition: twinvq_data.h:10344
quant_array_idx
static int quant_array_idx(const float val, const float *arr, const int num)
Definition: aacenc_utils.h:171
FFMIN
#define FFMIN(a, b)
Definition: common.h:105
tns_min_sfb
static const uint8_t *const tns_min_sfb[2]
Definition: aacenctab.h:112
i
int i
Definition: input.c:407
SingleChannelElement
Single Channel Element - used for both SCE and LFE elements.
Definition: aac.h:249
IndividualChannelStream::num_windows
int num_windows
Definition: aac.h:185
ff_aac_search_for_tns
void ff_aac_search_for_tns(AACEncContext *s, SingleChannelElement *sce)
Definition: aacenc_tns.c:161
LONG_STOP_SEQUENCE
@ LONG_STOP_SEQUENCE
Definition: aac.h:80
filt
static const int8_t filt[NUMTAPS *2]
Definition: af_earwax.c:39
IndividualChannelStream::tns_max_bands
int tns_max_bands
Definition: aac.h:186
AACEncContext::lpc
LPCContext lpc
used by TNS
Definition: aacenc.h:390
LONG_START_SEQUENCE
@ LONG_START_SEQUENCE
Definition: aac.h:78
SingleChannelElement::tns
TemporalNoiseShaping tns
Definition: aac.h:251
AACEncContext
AAC encoder context.
Definition: aacenc.h:378
FFPsyBand::energy
float energy
Definition: psymodel.h:52
IndividualChannelStream::window_sequence
enum WindowSequence window_sequence[2]
Definition: aac.h:177
TemporalNoiseShaping
Temporal Noise Shaping.
Definition: aac.h:199
quantize_coefs
static void quantize_coefs(double *coef, int *idx, float *lpc, int order, int c_bits)
Definition: aacenc_tns.c:147
aacenc_tns.h
IndividualChannelStream::max_sfb
uint8_t max_sfb
number of scalefactor bands per group
Definition: aac.h:176
tns_tmp2_map
static const INTFLOAT *const tns_tmp2_map[4]
Definition: aactab.h:82
aacenc_utils.h
aacenc.h