Mercurial > libavcodec.hg
comparison aacpsy.c @ 9935:d09283aeeef8 libavcodec
Merge the AAC encoder from SoC svn. It is still considered experimental.
author | alexc |
---|---|
date | Wed, 08 Jul 2009 20:01:31 +0000 |
parents | e9d9d946f213 |
children | 7f42ae22c351 |
comparison
equal
deleted
inserted
replaced
9934:ff96ee73b08b | 9935:d09283aeeef8 |
---|---|
23 * @file libavcodec/aacpsy.c | 23 * @file libavcodec/aacpsy.c |
24 * AAC encoder psychoacoustic model | 24 * AAC encoder psychoacoustic model |
25 */ | 25 */ |
26 | 26 |
27 #include "avcodec.h" | 27 #include "avcodec.h" |
28 #include "aacpsy.h" | |
29 #include "aactab.h" | 28 #include "aactab.h" |
29 #include "psymodel.h" | |
30 | 30 |
31 /*********************************** | 31 /*********************************** |
32 * TODOs: | 32 * TODOs: |
33 * General: | |
34 * better audio preprocessing (add DC highpass filter?) | |
35 * more psy models | |
36 * maybe improve coefficient quantization function in some way | |
37 * | |
38 * 3GPP-based psy model: | |
39 * thresholds linearization after their modifications for attaining given bitrate | 33 * thresholds linearization after their modifications for attaining given bitrate |
40 * try other bitrate controlling mechanism (maybe use ratecontrol.c?) | 34 * try other bitrate controlling mechanism (maybe use ratecontrol.c?) |
41 * control quality for quality-based output | 35 * control quality for quality-based output |
42 **********************************/ | 36 **********************************/ |
43 | 37 |
44 /** | 38 /** |
45 * Quantize one coefficient. | |
46 * @return absolute value of the quantized coefficient | |
47 * @see 3GPP TS26.403 5.6.2 "Scalefactor determination" | |
48 */ | |
49 static av_always_inline int quant(float coef, const float Q) | |
50 { | |
51 return av_clip((int)(pow(fabsf(coef) * Q, 0.75) + 0.4054), 0, 8191); | |
52 } | |
53 | |
54 static inline float get_approximate_quant_error(float *c, int size, int scale_idx) | |
55 { | |
56 int i; | |
57 int q; | |
58 float coef, unquant, sum = 0.0f; | |
59 const float Q = ff_aac_pow2sf_tab[200 - scale_idx + SCALE_ONE_POS - SCALE_DIV_512]; | |
60 const float IQ = ff_aac_pow2sf_tab[200 + scale_idx - SCALE_ONE_POS + SCALE_DIV_512]; | |
61 for(i = 0; i < size; i++){ | |
62 coef = fabs(c[i]); | |
63 q = quant(c[i], Q); | |
64 unquant = (q * cbrt(q)) * IQ; | |
65 sum += (coef - unquant) * (coef - unquant); | |
66 } | |
67 return sum; | |
68 } | |
69 | |
70 /** | |
71 * constants for 3GPP AAC psychoacoustic model | 39 * constants for 3GPP AAC psychoacoustic model |
72 * @{ | 40 * @{ |
73 */ | 41 */ |
74 #define PSY_3GPP_SPREAD_LOW 1.5f // spreading factor for ascending threshold spreading (15 dB/Bark) | 42 #define PSY_3GPP_SPREAD_LOW 1.5f // spreading factor for ascending threshold spreading (15 dB/Bark) |
75 #define PSY_3GPP_SPREAD_HI 3.0f // spreading factor for descending threshold spreading (30 dB/Bark) | 43 #define PSY_3GPP_SPREAD_HI 3.0f // spreading factor for descending threshold spreading (30 dB/Bark) |
44 | |
45 #define PSY_3GPP_RPEMIN 0.01f | |
46 #define PSY_3GPP_RPELEV 2.0f | |
76 /** | 47 /** |
77 * @} | 48 * @} |
78 */ | 49 */ |
79 | 50 |
80 /** | 51 /** |
81 * information for single band used by 3GPP TS26.403-inspired psychoacoustic model | 52 * information for single band used by 3GPP TS26.403-inspired psychoacoustic model |
82 */ | 53 */ |
83 typedef struct Psy3gppBand{ | 54 typedef struct Psy3gppBand{ |
84 float energy; ///< band energy | 55 float energy; ///< band energy |
85 float ffac; ///< form factor | 56 float ffac; ///< form factor |
57 float thr; ///< energy threshold | |
58 float min_snr; ///< minimal SNR | |
59 float thr_quiet; ///< threshold in quiet | |
86 }Psy3gppBand; | 60 }Psy3gppBand; |
61 | |
62 /** | |
63 * single/pair channel context for psychoacoustic model | |
64 */ | |
65 typedef struct Psy3gppChannel{ | |
66 Psy3gppBand band[128]; ///< bands information | |
67 Psy3gppBand prev_band[128]; ///< bands information from the previous frame | |
68 | |
69 float win_energy; ///< sliding average of channel energy | |
70 float iir_state[2]; ///< hi-pass IIR filter state | |
71 uint8_t next_grouping; ///< stored grouping scheme for the next frame (in case of 8 short window sequence) | |
72 enum WindowSequence next_window_seq; ///< window sequence to be used in the next frame | |
73 }Psy3gppChannel; | |
87 | 74 |
88 /** | 75 /** |
89 * psychoacoustic model frame type-dependent coefficients | 76 * psychoacoustic model frame type-dependent coefficients |
90 */ | 77 */ |
91 typedef struct Psy3gppCoeffs{ | 78 typedef struct Psy3gppCoeffs{ |
94 float spread_low[64]; ///< spreading factor for low-to-high threshold spreading in long frame | 81 float spread_low[64]; ///< spreading factor for low-to-high threshold spreading in long frame |
95 float spread_hi [64]; ///< spreading factor for high-to-low threshold spreading in long frame | 82 float spread_hi [64]; ///< spreading factor for high-to-low threshold spreading in long frame |
96 }Psy3gppCoeffs; | 83 }Psy3gppCoeffs; |
97 | 84 |
98 /** | 85 /** |
86 * 3GPP TS26.403-inspired psychoacoustic model specific data | |
87 */ | |
88 typedef struct Psy3gppContext{ | |
89 Psy3gppCoeffs psy_coef[2]; | |
90 Psy3gppChannel *ch; | |
91 }Psy3gppContext; | |
92 | |
93 /** | |
99 * Calculate Bark value for given line. | 94 * Calculate Bark value for given line. |
100 */ | 95 */ |
101 static inline float calc_bark(float f) | 96 static av_cold float calc_bark(float f) |
102 { | 97 { |
103 return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f)); | 98 return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f)); |
104 } | 99 } |
100 | |
101 #define ATH_ADD 4 | |
102 /** | |
103 * Calculate ATH value for given frequency. | |
104 * Borrowed from Lame. | |
105 */ | |
106 static av_cold float ath(float f, float add) | |
107 { | |
108 f /= 1000.0f; | |
109 return 3.64 * pow(f, -0.8) | |
110 - 6.8 * exp(-0.6 * (f - 3.4) * (f - 3.4)) | |
111 + 6.0 * exp(-0.15 * (f - 8.7) * (f - 8.7)) | |
112 + (0.6 + 0.04 * add) * 0.001 * f * f * f * f; | |
113 } | |
114 | |
115 static av_cold int psy_3gpp_init(FFPsyContext *ctx){ | |
116 Psy3gppContext *pctx; | |
117 float barks[1024]; | |
118 int i, j, g, start; | |
119 float prev, minscale, minath; | |
120 | |
121 ctx->model_priv_data = av_mallocz(sizeof(Psy3gppContext)); | |
122 pctx = (Psy3gppContext*) ctx->model_priv_data; | |
123 | |
124 for(i = 0; i < 1024; i++) | |
125 barks[i] = calc_bark(i * ctx->avctx->sample_rate / 2048.0); | |
126 minath = ath(3410, ATH_ADD); | |
127 for(j = 0; j < 2; j++){ | |
128 Psy3gppCoeffs *coeffs = &pctx->psy_coef[j]; | |
129 i = 0; | |
130 prev = 0.0; | |
131 for(g = 0; g < ctx->num_bands[j]; g++){ | |
132 i += ctx->bands[j][g]; | |
133 coeffs->barks[g] = (barks[i - 1] + prev) / 2.0; | |
134 prev = barks[i - 1]; | |
135 } | |
136 for(g = 0; g < ctx->num_bands[j] - 1; g++){ | |
137 coeffs->spread_low[g] = pow(10.0, -(coeffs->barks[g+1] - coeffs->barks[g]) * PSY_3GPP_SPREAD_LOW); | |
138 coeffs->spread_hi [g] = pow(10.0, -(coeffs->barks[g+1] - coeffs->barks[g]) * PSY_3GPP_SPREAD_HI); | |
139 } | |
140 start = 0; | |
141 for(g = 0; g < ctx->num_bands[j]; g++){ | |
142 minscale = ath(ctx->avctx->sample_rate * start / 1024.0, ATH_ADD); | |
143 for(i = 1; i < ctx->bands[j][g]; i++){ | |
144 minscale = fminf(minscale, ath(ctx->avctx->sample_rate * (start + i) / 1024.0 / 2.0, ATH_ADD)); | |
145 } | |
146 coeffs->ath[g] = minscale - minath; | |
147 start += ctx->bands[j][g]; | |
148 } | |
149 } | |
150 | |
151 pctx->ch = av_mallocz(sizeof(Psy3gppChannel) * ctx->avctx->channels); | |
152 return 0; | |
153 } | |
154 | |
155 /** | |
156 * IIR filter used in block switching decision | |
157 */ | |
158 static float iir_filter(int in, float state[2]) | |
159 { | |
160 float ret; | |
161 | |
162 ret = 0.7548f * (in - state[0]) + 0.5095f * state[1]; | |
163 state[0] = in; | |
164 state[1] = ret; | |
165 return ret; | |
166 } | |
167 | |
168 /** | |
169 * window grouping information stored as bits (0 - new group, 1 - group continues) | |
170 */ | |
171 static const uint8_t window_grouping[9] = { | |
172 0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36 | |
173 }; | |
174 | |
175 /** | |
176 * Tell encoder which window types to use. | |
177 * @see 3GPP TS26.403 5.4.1 "Blockswitching" | |
178 */ | |
179 static FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx, | |
180 const int16_t *audio, const int16_t *la, | |
181 int channel, int prev_type) | |
182 { | |
183 int i, j; | |
184 int br = ctx->avctx->bit_rate / ctx->avctx->channels; | |
185 int attack_ratio = br <= 16000 ? 18 : 10; | |
186 Psy3gppContext *pctx = (Psy3gppContext*) ctx->model_priv_data; | |
187 Psy3gppChannel *pch = &pctx->ch[channel]; | |
188 uint8_t grouping = 0; | |
189 FFPsyWindowInfo wi; | |
190 | |
191 memset(&wi, 0, sizeof(wi)); | |
192 if(la){ | |
193 float s[8], v; | |
194 int switch_to_eight = 0; | |
195 float sum = 0.0, sum2 = 0.0; | |
196 int attack_n = 0; | |
197 for(i = 0; i < 8; i++){ | |
198 for(j = 0; j < 128; j++){ | |
199 v = iir_filter(audio[(i*128+j)*ctx->avctx->channels], pch->iir_state); | |
200 sum += v*v; | |
201 } | |
202 s[i] = sum; | |
203 sum2 += sum; | |
204 } | |
205 for(i = 0; i < 8; i++){ | |
206 if(s[i] > pch->win_energy * attack_ratio){ | |
207 attack_n = i + 1; | |
208 switch_to_eight = 1; | |
209 break; | |
210 } | |
211 } | |
212 pch->win_energy = pch->win_energy*7/8 + sum2/64; | |
213 | |
214 wi.window_type[1] = prev_type; | |
215 switch(prev_type){ | |
216 case ONLY_LONG_SEQUENCE: | |
217 wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE; | |
218 break; | |
219 case LONG_START_SEQUENCE: | |
220 wi.window_type[0] = EIGHT_SHORT_SEQUENCE; | |
221 grouping = pch->next_grouping; | |
222 break; | |
223 case LONG_STOP_SEQUENCE: | |
224 wi.window_type[0] = ONLY_LONG_SEQUENCE; | |
225 break; | |
226 case EIGHT_SHORT_SEQUENCE: | |
227 wi.window_type[0] = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE; | |
228 grouping = switch_to_eight ? pch->next_grouping : 0; | |
229 break; | |
230 } | |
231 pch->next_grouping = window_grouping[attack_n]; | |
232 }else{ | |
233 for(i = 0; i < 3; i++) | |
234 wi.window_type[i] = prev_type; | |
235 grouping = (prev_type == EIGHT_SHORT_SEQUENCE) ? window_grouping[0] : 0; | |
236 } | |
237 | |
238 wi.window_shape = 1; | |
239 if(wi.window_type[0] != EIGHT_SHORT_SEQUENCE){ | |
240 wi.num_windows = 1; | |
241 wi.grouping[0] = 1; | |
242 }else{ | |
243 int lastgrp = 0; | |
244 wi.num_windows = 8; | |
245 for(i = 0; i < 8; i++){ | |
246 if(!((grouping >> i) & 1)) | |
247 lastgrp = i; | |
248 wi.grouping[lastgrp]++; | |
249 } | |
250 } | |
251 | |
252 return wi; | |
253 } | |
254 | |
255 /** | |
256 * Calculate band thresholds as suggested in 3GPP TS26.403 | |
257 */ | |
258 static void psy_3gpp_analyze(FFPsyContext *ctx, int channel, const float *coefs, | |
259 FFPsyWindowInfo *wi) | |
260 { | |
261 Psy3gppContext *pctx = (Psy3gppContext*) ctx->model_priv_data; | |
262 Psy3gppChannel *pch = &pctx->ch[channel]; | |
263 int start = 0; | |
264 int i, w, g; | |
265 const int num_bands = ctx->num_bands[wi->num_windows == 8]; | |
266 const uint8_t* band_sizes = ctx->bands[wi->num_windows == 8]; | |
267 Psy3gppCoeffs *coeffs = &pctx->psy_coef[wi->num_windows == 8]; | |
268 | |
269 //calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation" | |
270 for(w = 0; w < wi->num_windows*16; w += 16){ | |
271 for(g = 0; g < num_bands; g++){ | |
272 Psy3gppBand *band = &pch->band[w+g]; | |
273 band->energy = 0.0f; | |
274 for(i = 0; i < band_sizes[g]; i++) | |
275 band->energy += coefs[start+i] * coefs[start+i]; | |
276 band->energy *= 1.0f / (512*512); | |
277 band->thr = band->energy * 0.001258925f; | |
278 start += band_sizes[g]; | |
279 | |
280 ctx->psy_bands[channel*PSY_MAX_BANDS+w+g].energy = band->energy; | |
281 } | |
282 } | |
283 //modify thresholds - spread, threshold in quiet - 5.4.3 "Spreaded Energy Calculation" | |
284 for(w = 0; w < wi->num_windows*16; w += 16){ | |
285 Psy3gppBand *band = &pch->band[w]; | |
286 for(g = 1; g < num_bands; g++){ | |
287 band[g].thr = FFMAX(band[g].thr, band[g-1].thr * coeffs->spread_low[g-1]); | |
288 } | |
289 for(g = num_bands - 2; g >= 0; g--){ | |
290 band[g].thr = FFMAX(band[g].thr, band[g+1].thr * coeffs->spread_hi [g]); | |
291 } | |
292 for(g = 0; g < num_bands; g++){ | |
293 band[g].thr_quiet = FFMAX(band[g].thr, coeffs->ath[g]); | |
294 if(wi->num_windows != 8 && wi->window_type[1] != EIGHT_SHORT_SEQUENCE){ | |
295 band[g].thr_quiet = fmaxf(PSY_3GPP_RPEMIN*band[g].thr_quiet, | |
296 fminf(band[g].thr_quiet, | |
297 PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet)); | |
298 } | |
299 band[g].thr = FFMAX(band[g].thr, band[g].thr_quiet * 0.25); | |
300 | |
301 ctx->psy_bands[channel*PSY_MAX_BANDS+w+g].threshold = band[g].thr; | |
302 } | |
303 } | |
304 memcpy(pch->prev_band, pch->band, sizeof(pch->band)); | |
305 } | |
306 | |
307 static av_cold void psy_3gpp_end(FFPsyContext *apc) | |
308 { | |
309 Psy3gppContext *pctx = (Psy3gppContext*) apc->model_priv_data; | |
310 av_freep(&pctx->ch); | |
311 av_freep(&apc->model_priv_data); | |
312 } | |
313 | |
314 | |
315 const FFPsyModel ff_aac_psy_model = | |
316 { | |
317 .name = "3GPP TS 26.403-inspired model", | |
318 .init = psy_3gpp_init, | |
319 .window = psy_3gpp_window, | |
320 .analyze = psy_3gpp_analyze, | |
321 .end = psy_3gpp_end, | |
322 }; |