FFmpeg
aacenc_utils.h
Go to the documentation of this file.
1 /*
2  * AAC encoder utilities
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 utilities
25  * @author Rostislav Pehlivanov ( atomnuker gmail com )
26  */
27 
28 #ifndef AVCODEC_AACENC_UTILS_H
29 #define AVCODEC_AACENC_UTILS_H
30 
31 #include "libavutil/ffmath.h"
32 #include "aacenc.h"
33 #include "aacenctab.h"
34 #include "aactab.h"
35 
36 #define ROUND_STANDARD 0.4054f
37 #define ROUND_TO_ZERO 0.1054f
38 #define C_QUANT 0.4054f
39 
40 static inline float pos_pow34(float a)
41 {
42  return sqrtf(a * sqrtf(a));
43 }
44 
45 /**
46  * Quantize one coefficient.
47  * @return absolute value of the quantized coefficient
48  * @see 3GPP TS26.403 5.6.2 "Scalefactor determination"
49  */
50 static inline int quant(float coef, const float Q, const float rounding)
51 {
52  float a = coef * Q;
53  return sqrtf(a * sqrtf(a)) + rounding;
54 }
55 
56 static inline float find_max_val(int group_len, int swb_size, const float *scaled)
57 {
58  float maxval = 0.0f;
59  int w2, i;
60  for (w2 = 0; w2 < group_len; w2++) {
61  for (i = 0; i < swb_size; i++) {
62  maxval = FFMAX(maxval, scaled[w2*128+i]);
63  }
64  }
65  return maxval;
66 }
67 
68 static inline int find_min_book(float maxval, int sf)
69 {
71  int qmaxval, cb;
72  qmaxval = maxval * Q34 + C_QUANT;
73  if (qmaxval >= (FF_ARRAY_ELEMS(aac_maxval_cb)))
74  cb = 11;
75  else
76  cb = aac_maxval_cb[qmaxval];
77  return cb;
78 }
79 
80 static inline float find_form_factor(int group_len, int swb_size, float thresh,
81  const float *scaled, float nzslope) {
82  const float iswb_size = 1.0f / swb_size;
83  const float iswb_sizem1 = 1.0f / (swb_size - 1);
84  const float ethresh = thresh;
85  float form = 0.0f, weight = 0.0f;
86  int w2, i;
87  for (w2 = 0; w2 < group_len; w2++) {
88  float e = 0.0f, e2 = 0.0f, var = 0.0f, maxval = 0.0f;
89  float nzl = 0;
90  for (i = 0; i < swb_size; i++) {
91  float s = fabsf(scaled[w2*128+i]);
92  maxval = FFMAX(maxval, s);
93  e += s;
94  e2 += s *= s;
95  /* We really don't want a hard non-zero-line count, since
96  * even below-threshold lines do add up towards band spectral power.
97  * So, fall steeply towards zero, but smoothly
98  */
99  if (s >= ethresh) {
100  nzl += 1.0f;
101  } else {
102  if (nzslope == 2.f)
103  nzl += (s / ethresh) * (s / ethresh);
104  else
105  nzl += ff_fast_powf(s / ethresh, nzslope);
106  }
107  }
108  if (e2 > thresh) {
109  float frm;
110  e *= iswb_size;
111 
112  /** compute variance */
113  for (i = 0; i < swb_size; i++) {
114  float d = fabsf(scaled[w2*128+i]) - e;
115  var += d*d;
116  }
117  var = sqrtf(var * iswb_sizem1);
118 
119  e2 *= iswb_size;
120  frm = e / FFMIN(e+4*var,maxval);
121  form += e2 * sqrtf(frm) / FFMAX(0.5f,nzl);
122  weight += e2;
123  }
124  }
125  if (weight > 0) {
126  return form / weight;
127  } else {
128  return 1.0f;
129  }
130 }
131 
132 /** Return the minimum scalefactor where the quantized coef does not clip. */
133 static inline uint8_t coef2minsf(float coef)
134 {
135  return av_clip_uint8(log2f(coef)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
136 }
137 
138 /** Return the maximum scalefactor where the quantized coef is not zero. */
139 static inline uint8_t coef2maxsf(float coef)
140 {
141  return av_clip_uint8(log2f(coef)*4 + 6 + SCALE_ONE_POS - SCALE_DIV_512);
142 }
143 
144 /*
145  * Returns the closest possible index to an array of float values, given a value.
146  */
147 static inline int quant_array_idx(const float val, const float *arr, const int num)
148 {
149  int i, index = 0;
150  float quant_min_err = INFINITY;
151  for (i = 0; i < num; i++) {
152  float error = (val - arr[i])*(val - arr[i]);
153  if (error < quant_min_err) {
154  quant_min_err = error;
155  index = i;
156  }
157  }
158  return index;
159 }
160 
161 /**
162  * approximates exp10f(-3.0f*(0.5f + 0.5f * cosf(FFMIN(b,15.5f) / 15.5f)))
163  */
164 static av_always_inline float bval2bmax(float b)
165 {
166  return 0.001f + 0.0035f * (b*b*b) / (15.5f*15.5f*15.5f);
167 }
168 
169 /*
170  * Compute a nextband map to be used with SF delta constraint utilities.
171  * The nextband array should contain 128 elements, and positions that don't
172  * map to valid, nonzero bands of the form w*16+g (with w being the initial
173  * window of the window group, only) are left indetermined.
174  */
175 static inline void ff_init_nextband_map(const SingleChannelElement *sce, uint8_t *nextband)
176 {
177  unsigned char prevband = 0;
178  int w, g;
179  /** Just a safe default */
180  for (g = 0; g < 128; g++)
181  nextband[g] = g;
182 
183  /** Now really navigate the nonzero band chain */
184  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
185  for (g = 0; g < sce->ics.num_swb; g++) {
186  if (!sce->zeroes[w*16+g] && sce->band_type[w*16+g] < RESERVED_BT)
187  prevband = nextband[prevband] = w*16+g;
188  }
189  }
190  nextband[prevband] = prevband; /* terminate */
191 }
192 
193 /*
194  * Updates nextband to reflect a removed band (equivalent to
195  * calling ff_init_nextband_map after marking a band as zero)
196  */
197 static inline void ff_nextband_remove(uint8_t *nextband, int prevband, int band)
198 {
199  nextband[prevband] = nextband[band];
200 }
201 
202 /*
203  * Checks whether the specified band could be removed without inducing
204  * scalefactor delta that violates SF delta encoding constraints.
205  * prev_sf has to be the scalefactor of the previous nonzero, nonspecial
206  * band, in encoding order, or negative if there was no such band.
207  */
209  const uint8_t *nextband, int prev_sf, int band)
210 {
211  return prev_sf >= 0
212  && sce->sf_idx[nextband[band]] >= (prev_sf - SCALE_MAX_DIFF)
213  && sce->sf_idx[nextband[band]] <= (prev_sf + SCALE_MAX_DIFF);
214 }
215 
216 /*
217  * Checks whether the specified band's scalefactor could be replaced
218  * with another one without violating SF delta encoding constraints.
219  * prev_sf has to be the scalefactor of the previous nonzero, nonsepcial
220  * band, in encoding order, or negative if there was no such band.
221  */
222 static inline int ff_sfdelta_can_replace(const SingleChannelElement *sce,
223  const uint8_t *nextband, int prev_sf, int new_sf, int band)
224 {
225  return new_sf >= (prev_sf - SCALE_MAX_DIFF)
226  && new_sf <= (prev_sf + SCALE_MAX_DIFF)
227  && sce->sf_idx[nextband[band]] >= (new_sf - SCALE_MAX_DIFF)
228  && sce->sf_idx[nextband[band]] <= (new_sf + SCALE_MAX_DIFF);
229 }
230 
231 /**
232  * linear congruential pseudorandom number generator
233  *
234  * @param previous_val pointer to the current state of the generator
235  *
236  * @return Returns a 32-bit pseudorandom integer
237  */
238 static av_always_inline int lcg_random(unsigned previous_val)
239 {
240  union { unsigned u; int s; } v = { previous_val * 1664525u + 1013904223 };
241  return v.s;
242 }
243 
244 #define ERROR_IF(cond, ...) \
245  if (cond) { \
246  av_log(avctx, AV_LOG_ERROR, __VA_ARGS__); \
247  return AVERROR(EINVAL); \
248  }
249 
250 #define WARN_IF(cond, ...) \
251  if (cond) { \
252  av_log(avctx, AV_LOG_WARNING, __VA_ARGS__); \
253  }
254 
255 #endif /* AVCODEC_AACENC_UTILS_H */
error
static void error(const char *err)
Definition: target_bsf_fuzzer.c:31
INFINITY
#define INFINITY
Definition: mathematics.h:118
cb
static double cb(void *priv, double x, double y)
Definition: vf_geq.c:241
u
#define u(width, name, range_min, range_max)
Definition: cbs_h2645.c:250
aacenctab.h
log2f
#define log2f(x)
Definition: libm.h:409
SingleChannelElement::zeroes
uint8_t zeroes[128]
band is not coded
Definition: aacenc.h:120
w
uint8_t w
Definition: llviddspenc.c:38
bval2bmax
static av_always_inline float bval2bmax(float b)
approximates exp10f(-3.0f*(0.5f + 0.5f * cosf(FFMIN(b,15.5f) / 15.5f)))
Definition: aacenc_utils.h:164
b
#define b
Definition: input.c:41
ff_fast_powf
static av_always_inline float ff_fast_powf(float x, float y)
Compute x^y for floating point x, y.
Definition: ffmath.h:62
ff_sfdelta_can_remove_band
static int ff_sfdelta_can_remove_band(const SingleChannelElement *sce, const uint8_t *nextband, int prev_sf, int band)
Definition: aacenc_utils.h:208
FFMAX
#define FFMAX(a, b)
Definition: macros.h:47
coef2maxsf
static uint8_t coef2maxsf(float coef)
Return the maximum scalefactor where the quantized coef is not zero.
Definition: aacenc_utils.h:139
IndividualChannelStream::num_swb
int num_swb
number of scalefactor window bands
Definition: aacdec.h:92
Q
#define Q(x)
Definition: vvc_filter_template.c:433
SCALE_DIV_512
#define SCALE_DIV_512
scalefactor difference that corresponds to scale difference in 512 times
Definition: aac.h:104
ff_sfdelta_can_replace
static int ff_sfdelta_can_replace(const SingleChannelElement *sce, const uint8_t *nextband, int prev_sf, int new_sf, int band)
Definition: aacenc_utils.h:222
find_form_factor
static float find_form_factor(int group_len, int swb_size, float thresh, const float *scaled, float nzslope)
Definition: aacenc_utils.h:80
POW_SF2_ZERO
#define POW_SF2_ZERO
ff_aac_pow2sf_tab index corresponding to pow(2, 0);
Definition: aac.h:110
val
static double val(void *priv, double ch)
Definition: aeval.c:78
fabsf
static __device__ float fabsf(float a)
Definition: cuda_runtime.h:181
quant
static int quant(float coef, const float Q, const float rounding)
Quantize one coefficient.
Definition: aacenc_utils.h:50
SingleChannelElement::ics
IndividualChannelStream ics
Definition: aacdec.h:132
FF_ARRAY_ELEMS
#define FF_ARRAY_ELEMS(a)
Definition: sinewin_tablegen.c:29
s
#define s(width, name)
Definition: cbs_vp9.c:198
g
const char * g
Definition: vf_curves.c:127
form
This is the more generic form
Definition: tablegen.txt:34
aactab.h
sqrtf
static __device__ float sqrtf(float a)
Definition: cuda_runtime.h:184
ff_init_nextband_map
static void ff_init_nextband_map(const SingleChannelElement *sce, uint8_t *nextband)
Definition: aacenc_utils.h:175
index
int index
Definition: gxfenc.c:89
SingleChannelElement::sf_idx
int sf_idx[128]
scalefactor indices
Definition: aacenc.h:119
weight
static int weight(int i, int blen, int offset)
Definition: diracdec.c:1562
ff_aac_pow34sf_tab
float ff_aac_pow34sf_tab[428]
coef2minsf
static uint8_t coef2minsf(float coef)
Return the minimum scalefactor where the quantized coef does not clip.
Definition: aacenc_utils.h:133
f
f
Definition: af_crystalizer.c:121
C_QUANT
#define C_QUANT
Definition: aacenc_utils.h:38
quant_array_idx
static int quant_array_idx(const float val, const float *arr, const int num)
Definition: aacenc_utils.h:147
a
The reader does not expect b to be semantically here and if the code is changed by maybe adding a a division or other the signedness will almost certainly be mistaken To avoid this confusion a new type was SUINT is the C unsigned type but it holds a signed int to use the same example SUINT a
Definition: undefined.txt:41
SingleChannelElement::band_type
enum BandType band_type[128]
band types
Definition: aacdec.h:134
SCALE_MAX_DIFF
#define SCALE_MAX_DIFF
maximum scalefactor difference allowed by standard
Definition: aac.h:107
pos_pow34
static float pos_pow34(float a)
Definition: aacenc_utils.h:40
SingleChannelElement
Single Channel Element - used for both SCE and LFE elements.
Definition: aacdec.h:131
i
#define i(width, name, range_min, range_max)
Definition: cbs_h2645.c:255
IndividualChannelStream::num_windows
int num_windows
Definition: aacdec.h:93
SCALE_ONE_POS
#define SCALE_ONE_POS
scalefactor index that corresponds to scale=1.0
Definition: aac.h:105
find_min_book
static int find_min_book(float maxval, int sf)
Definition: aacenc_utils.h:68
av_always_inline
#define av_always_inline
Definition: attributes.h:49
FFMIN
#define FFMIN(a, b)
Definition: macros.h:49
lcg_random
static av_always_inline int lcg_random(unsigned previous_val)
linear congruential pseudorandom number generator
Definition: aacenc_utils.h:238
RESERVED_BT
@ RESERVED_BT
Band types following are encoded differently from others.
Definition: aac.h:73
aac_maxval_cb
static const unsigned char aac_maxval_cb[]
Definition: aacenctab.h:123
ffmath.h
av_clip_uint8
#define av_clip_uint8
Definition: common.h:104
find_max_val
static float find_max_val(int group_len, int swb_size, const float *scaled)
Definition: aacenc_utils.h:56
d
d
Definition: ffmpeg_filter.c:425
IndividualChannelStream::group_len
uint8_t group_len[8]
Definition: aacdec.h:89
ff_nextband_remove
static void ff_nextband_remove(uint8_t *nextband, int prevband, int band)
Definition: aacenc_utils.h:197
aacenc.h