/src/ffmpeg/libavcodec/aacpsy.c
Line | Count | Source |
1 | | /* |
2 | | * AAC encoder psychoacoustic model |
3 | | * Copyright (C) 2008 Konstantin Shishkov |
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 psychoacoustic model |
25 | | */ |
26 | | |
27 | | #include "libavutil/attributes.h" |
28 | | #include "libavutil/ffmath.h" |
29 | | #include "libavutil/mem.h" |
30 | | |
31 | | #include "avcodec.h" |
32 | | #include "aac.h" |
33 | | #include "psymodel.h" |
34 | | |
35 | | /*********************************** |
36 | | * TODOs: |
37 | | * try other bitrate controlling mechanism (maybe use ratecontrol.c?) |
38 | | * control quality for quality-based output |
39 | | **********************************/ |
40 | | |
41 | | /** |
42 | | * constants for 3GPP AAC psychoacoustic model |
43 | | * @{ |
44 | | */ |
45 | 0 | #define PSY_3GPP_THR_SPREAD_HI 1.5f // spreading factor for low-to-hi threshold spreading (15 dB/Bark) |
46 | 0 | #define PSY_3GPP_THR_SPREAD_LOW 3.0f // spreading factor for hi-to-low threshold spreading (30 dB/Bark) |
47 | | /* spreading factor for low-to-hi energy spreading, long block, > 22kbps/channel (20dB/Bark) */ |
48 | 0 | #define PSY_3GPP_EN_SPREAD_HI_L1 2.0f |
49 | | /* spreading factor for low-to-hi energy spreading, long block, <= 22kbps/channel (15dB/Bark) */ |
50 | | #define PSY_3GPP_EN_SPREAD_HI_L2 1.5f |
51 | | /* spreading factor for low-to-hi energy spreading, short block (15 dB/Bark) */ |
52 | 0 | #define PSY_3GPP_EN_SPREAD_HI_S 1.5f |
53 | | /* spreading factor for hi-to-low energy spreading, long block (30dB/Bark) */ |
54 | 0 | #define PSY_3GPP_EN_SPREAD_LOW_L 3.0f |
55 | | /* spreading factor for hi-to-low energy spreading, short block (20dB/Bark) */ |
56 | 0 | #define PSY_3GPP_EN_SPREAD_LOW_S 2.0f |
57 | | |
58 | | #define PSY_3GPP_RPEMIN 0.01f |
59 | | #define PSY_3GPP_RPELEV 2.0f |
60 | | |
61 | 0 | #define PSY_3GPP_C1 3.0f /* log2(8) */ |
62 | 0 | #define PSY_3GPP_C2 1.3219281f /* log2(2.5) */ |
63 | 0 | #define PSY_3GPP_C3 0.55935729f /* 1 - C2 / C1 */ |
64 | | |
65 | 0 | #define PSY_SNR_1DB 7.9432821e-1f /* -1dB */ |
66 | 0 | #define PSY_SNR_25DB 3.1622776e-3f /* -25dB */ |
67 | | |
68 | 0 | #define PSY_3GPP_SAVE_SLOPE_L -0.46666667f |
69 | 0 | #define PSY_3GPP_SAVE_SLOPE_S -0.36363637f |
70 | 0 | #define PSY_3GPP_SAVE_ADD_L -0.84285712f |
71 | 0 | #define PSY_3GPP_SAVE_ADD_S -0.75f |
72 | 0 | #define PSY_3GPP_SPEND_SLOPE_L 0.66666669f |
73 | 0 | #define PSY_3GPP_SPEND_SLOPE_S 0.81818181f |
74 | 0 | #define PSY_3GPP_SPEND_ADD_L -0.35f |
75 | 0 | #define PSY_3GPP_SPEND_ADD_S -0.26111111f |
76 | 0 | #define PSY_3GPP_CLIP_LO_L 0.2f |
77 | 0 | #define PSY_3GPP_CLIP_LO_S 0.2f |
78 | 0 | #define PSY_3GPP_CLIP_HI_L 0.95f |
79 | 0 | #define PSY_3GPP_CLIP_HI_S 0.75f |
80 | | |
81 | 0 | #define PSY_3GPP_AH_THR_LONG 0.5f |
82 | 0 | #define PSY_3GPP_AH_THR_SHORT 0.63f |
83 | | |
84 | 0 | #define PSY_PE_FORGET_SLOPE 511 |
85 | | |
86 | | enum { |
87 | | PSY_3GPP_AH_NONE, |
88 | | PSY_3GPP_AH_INACTIVE, |
89 | | PSY_3GPP_AH_ACTIVE |
90 | | }; |
91 | | |
92 | 0 | #define PSY_3GPP_BITS_TO_PE(bits) ((bits) * 1.18f) |
93 | 0 | #define PSY_3GPP_PE_TO_BITS(bits) ((bits) / 1.18f) |
94 | | |
95 | | /* LAME psy model constants */ |
96 | 0 | #define PSY_LAME_FIR_LEN 21 ///< LAME psy model FIR order |
97 | 0 | #define AAC_BLOCK_SIZE_LONG 1024 ///< long block size |
98 | 0 | #define AAC_BLOCK_SIZE_SHORT 128 ///< short block size |
99 | 0 | #define AAC_NUM_BLOCKS_SHORT 8 ///< number of blocks in a short sequence |
100 | 0 | #define PSY_LAME_NUM_SUBBLOCKS 2 ///< Number of sub-blocks in each short block |
101 | | |
102 | | /** |
103 | | * @} |
104 | | */ |
105 | | |
106 | | /** |
107 | | * information for single band used by 3GPP TS26.403-inspired psychoacoustic model |
108 | | */ |
109 | | typedef struct AacPsyBand{ |
110 | | float energy; ///< band energy |
111 | | float thr; ///< energy threshold |
112 | | float thr_quiet; ///< threshold in quiet |
113 | | float nz_lines; ///< number of non-zero spectral lines |
114 | | float active_lines; ///< number of active spectral lines |
115 | | float pe; ///< perceptual entropy |
116 | | float pe_const; ///< constant part of the PE calculation |
117 | | float norm_fac; ///< normalization factor for linearization |
118 | | int avoid_holes; ///< hole avoidance flag |
119 | | }AacPsyBand; |
120 | | |
121 | | /** |
122 | | * single/pair channel context for psychoacoustic model |
123 | | */ |
124 | | typedef struct AacPsyChannel{ |
125 | | AacPsyBand band[128]; ///< bands information |
126 | | AacPsyBand prev_band[128]; ///< bands information from the previous frame |
127 | | |
128 | | float win_energy; ///< sliding average of channel energy |
129 | | float iir_state[2]; ///< hi-pass IIR filter state |
130 | | uint8_t next_grouping; ///< stored grouping scheme for the next frame (in case of 8 short window sequence) |
131 | | enum WindowSequence next_window_seq; ///< window sequence to be used in the next frame |
132 | | /* LAME psy model specific members */ |
133 | | float attack_threshold; ///< attack threshold for this channel |
134 | | float prev_energy_subshort[AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS]; |
135 | | int prev_attack; ///< attack value for the last short block in the previous sequence |
136 | | int next_attack0_zero; ///< whether attack[0] of the next frame is zero |
137 | | }AacPsyChannel; |
138 | | |
139 | | /** |
140 | | * psychoacoustic model frame type-dependent coefficients |
141 | | */ |
142 | | typedef struct AacPsyCoeffs{ |
143 | | float ath; ///< absolute threshold of hearing per bands |
144 | | float barks; ///< Bark value for each spectral band in long frame |
145 | | float spread_low[2]; ///< spreading factor for low-to-high threshold spreading in long frame |
146 | | float spread_hi [2]; ///< spreading factor for high-to-low threshold spreading in long frame |
147 | | float min_snr; ///< minimal SNR |
148 | | }AacPsyCoeffs; |
149 | | |
150 | | /** |
151 | | * 3GPP TS26.403-inspired psychoacoustic model specific data |
152 | | */ |
153 | | typedef struct AacPsyContext{ |
154 | | int chan_bitrate; ///< bitrate per channel |
155 | | int frame_bits; ///< average bits per frame |
156 | | int fill_level; ///< bit reservoir fill level |
157 | | struct { |
158 | | float min; ///< minimum allowed PE for bit factor calculation |
159 | | float max; ///< maximum allowed PE for bit factor calculation |
160 | | float previous; ///< allowed PE of the previous frame |
161 | | float correction; ///< PE correction factor |
162 | | } pe; |
163 | | AacPsyCoeffs psy_coef[2][64]; |
164 | | AacPsyChannel *ch; |
165 | | float global_quality; ///< normalized global quality taken from avctx |
166 | | }AacPsyContext; |
167 | | |
168 | | /** |
169 | | * LAME psy model preset struct |
170 | | */ |
171 | | typedef struct PsyLamePreset { |
172 | | int quality; ///< Quality to map the rest of the values to. |
173 | | /* This is overloaded to be both kbps per channel in ABR mode, and |
174 | | * requested quality in constant quality mode. |
175 | | */ |
176 | | float st_lrm; ///< short threshold for L, R, and M channels |
177 | | } PsyLamePreset; |
178 | | |
179 | | /** |
180 | | * LAME psy model preset table for ABR |
181 | | */ |
182 | | static const PsyLamePreset psy_abr_map[] = { |
183 | | /* TODO: Tuning. These were taken from LAME. */ |
184 | | /* kbps/ch st_lrm */ |
185 | | { 8, 7.60}, |
186 | | { 16, 7.60}, |
187 | | { 24, 7.60}, |
188 | | { 32, 7.60}, |
189 | | { 40, 7.60}, |
190 | | { 48, 7.60}, |
191 | | { 56, 7.60}, |
192 | | { 64, 7.40}, |
193 | | { 80, 7.00}, |
194 | | { 96, 6.60}, |
195 | | {112, 6.20}, |
196 | | {128, 6.20}, |
197 | | {160, 6.20} |
198 | | }; |
199 | | |
200 | | /** |
201 | | * LAME psy model preset table for constant quality |
202 | | */ |
203 | | static const PsyLamePreset psy_vbr_map[] = { |
204 | | /* vbr_q st_lrm */ |
205 | | { 0, 4.20}, |
206 | | { 1, 4.20}, |
207 | | { 2, 4.20}, |
208 | | { 3, 4.20}, |
209 | | { 4, 4.20}, |
210 | | { 5, 4.20}, |
211 | | { 6, 4.20}, |
212 | | { 7, 4.20}, |
213 | | { 8, 4.20}, |
214 | | { 9, 4.20}, |
215 | | {10, 4.20} |
216 | | }; |
217 | | |
218 | | /** |
219 | | * LAME psy model FIR coefficient table |
220 | | */ |
221 | | static const float psy_fir_coeffs[] = { |
222 | | -8.65163e-18 * 2, -0.00851586 * 2, -6.74764e-18 * 2, 0.0209036 * 2, |
223 | | -3.36639e-17 * 2, -0.0438162 * 2, -1.54175e-17 * 2, 0.0931738 * 2, |
224 | | -5.52212e-17 * 2, -0.313819 * 2 |
225 | | }; |
226 | | |
227 | | /** |
228 | | * Calculate the ABR attack threshold from the above LAME psymodel table. |
229 | | */ |
230 | | static float lame_calc_attack_threshold(int bitrate) |
231 | 0 | { |
232 | | /* Assume max bitrate to start with */ |
233 | 0 | int lower_range = 12, upper_range = 12; |
234 | 0 | int lower_range_kbps = psy_abr_map[12].quality; |
235 | 0 | int upper_range_kbps = psy_abr_map[12].quality; |
236 | 0 | int i; |
237 | | |
238 | | /* Determine which bitrates the value specified falls between. |
239 | | * If the loop ends without breaking our above assumption of 320kbps was correct. |
240 | | */ |
241 | 0 | for (i = 1; i < 13; i++) { |
242 | 0 | if (FFMAX(bitrate, psy_abr_map[i].quality) != bitrate) { |
243 | 0 | upper_range = i; |
244 | 0 | upper_range_kbps = psy_abr_map[i ].quality; |
245 | 0 | lower_range = i - 1; |
246 | 0 | lower_range_kbps = psy_abr_map[i - 1].quality; |
247 | 0 | break; /* Upper range found */ |
248 | 0 | } |
249 | 0 | } |
250 | | |
251 | | /* Determine which range the value specified is closer to */ |
252 | 0 | if ((upper_range_kbps - bitrate) > (bitrate - lower_range_kbps)) |
253 | 0 | return psy_abr_map[lower_range].st_lrm; |
254 | 0 | return psy_abr_map[upper_range].st_lrm; |
255 | 0 | } |
256 | | |
257 | | /** |
258 | | * LAME psy model specific initialization |
259 | | */ |
260 | | static av_cold void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx) |
261 | 0 | { |
262 | 0 | int i, j; |
263 | |
|
264 | 0 | for (i = 0; i < avctx->ch_layout.nb_channels; i++) { |
265 | 0 | AacPsyChannel *pch = &ctx->ch[i]; |
266 | |
|
267 | 0 | if (avctx->flags & AV_CODEC_FLAG_QSCALE) |
268 | 0 | pch->attack_threshold = psy_vbr_map[av_clip(avctx->global_quality / FF_QP2LAMBDA, 0, 10)].st_lrm; |
269 | 0 | else |
270 | 0 | pch->attack_threshold = lame_calc_attack_threshold(avctx->bit_rate / avctx->ch_layout.nb_channels / 1000); |
271 | |
|
272 | 0 | for (j = 0; j < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; j++) |
273 | 0 | pch->prev_energy_subshort[j] = 10.0f; |
274 | 0 | } |
275 | 0 | } |
276 | | |
277 | | /** |
278 | | * Calculate Bark value for given line. |
279 | | */ |
280 | | static av_cold float calc_bark(float f) |
281 | 0 | { |
282 | 0 | return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f)); |
283 | 0 | } |
284 | | |
285 | 0 | #define ATH_ADD 4 |
286 | | /** |
287 | | * Calculate ATH value for given frequency. |
288 | | * Borrowed from Lame. |
289 | | */ |
290 | | static av_cold float ath(float f, float add) |
291 | 0 | { |
292 | 0 | f /= 1000.0f; |
293 | 0 | return 3.64 * pow(f, -0.8) |
294 | 0 | - 6.8 * exp(-0.6 * (f - 3.4) * (f - 3.4)) |
295 | 0 | + 6.0 * exp(-0.15 * (f - 8.7) * (f - 8.7)) |
296 | 0 | + (0.6 + 0.04 * add) * 0.001 * f * f * f * f; |
297 | 0 | } |
298 | | |
299 | 0 | static av_cold int psy_3gpp_init(FFPsyContext *ctx) { |
300 | 0 | AacPsyContext *pctx; |
301 | 0 | float bark; |
302 | 0 | int i, j, g, start; |
303 | 0 | float prev, minscale, minath, minsnr, pe_min; |
304 | 0 | int chan_bitrate = ctx->avctx->bit_rate / ((ctx->avctx->flags & AV_CODEC_FLAG_QSCALE) ? 2.0f : ctx->avctx->ch_layout.nb_channels); |
305 | |
|
306 | 0 | const int bandwidth = ctx->cutoff ? ctx->cutoff : AAC_CUTOFF(ctx->avctx); |
307 | 0 | const float num_bark = calc_bark((float)bandwidth); |
308 | |
|
309 | 0 | if (bandwidth <= 0) |
310 | 0 | return AVERROR(EINVAL); |
311 | | |
312 | 0 | ctx->model_priv_data = av_mallocz(sizeof(AacPsyContext)); |
313 | 0 | if (!ctx->model_priv_data) |
314 | 0 | return AVERROR(ENOMEM); |
315 | 0 | pctx = ctx->model_priv_data; |
316 | 0 | pctx->global_quality = (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120) * 0.01f; |
317 | |
|
318 | 0 | if (ctx->avctx->flags & AV_CODEC_FLAG_QSCALE) { |
319 | | /* Use the target average bitrate to compute spread parameters */ |
320 | 0 | chan_bitrate = (int)(chan_bitrate / 120.0 * (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120)); |
321 | 0 | } |
322 | |
|
323 | 0 | pctx->chan_bitrate = chan_bitrate; |
324 | 0 | pctx->frame_bits = FFMIN(2560, chan_bitrate * AAC_BLOCK_SIZE_LONG / ctx->avctx->sample_rate); |
325 | 0 | pctx->pe.min = 8.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f); |
326 | 0 | pctx->pe.max = 12.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f); |
327 | 0 | ctx->bitres.size = 6144 - pctx->frame_bits; |
328 | 0 | ctx->bitres.size -= ctx->bitres.size % 8; |
329 | 0 | pctx->fill_level = ctx->bitres.size; |
330 | 0 | minath = ath(3410 - 0.733 * ATH_ADD, ATH_ADD); |
331 | 0 | for (j = 0; j < 2; j++) { |
332 | 0 | AacPsyCoeffs *coeffs = pctx->psy_coef[j]; |
333 | 0 | const uint8_t *band_sizes = ctx->bands[j]; |
334 | 0 | float line_to_frequency = ctx->avctx->sample_rate / (j ? 256.f : 2048.0f); |
335 | 0 | float avg_chan_bits = chan_bitrate * (j ? 128.0f : 1024.0f) / ctx->avctx->sample_rate; |
336 | | /* reference encoder uses 2.4% here instead of 60% like the spec says */ |
337 | 0 | float bark_pe = 0.024f * PSY_3GPP_BITS_TO_PE(avg_chan_bits) / num_bark; |
338 | 0 | float en_spread_low = j ? PSY_3GPP_EN_SPREAD_LOW_S : PSY_3GPP_EN_SPREAD_LOW_L; |
339 | | /* High energy spreading for long blocks <= 22kbps/channel and short blocks are the same. */ |
340 | 0 | float en_spread_hi = (j || (chan_bitrate <= 22.0f)) ? PSY_3GPP_EN_SPREAD_HI_S : PSY_3GPP_EN_SPREAD_HI_L1; |
341 | |
|
342 | 0 | i = 0; |
343 | 0 | prev = 0.0; |
344 | 0 | for (g = 0; g < ctx->num_bands[j]; g++) { |
345 | 0 | i += band_sizes[g]; |
346 | 0 | bark = calc_bark((i-1) * line_to_frequency); |
347 | 0 | coeffs[g].barks = (bark + prev) / 2.0; |
348 | 0 | prev = bark; |
349 | 0 | } |
350 | 0 | for (g = 0; g < ctx->num_bands[j] - 1; g++) { |
351 | 0 | AacPsyCoeffs *coeff = &coeffs[g]; |
352 | 0 | float bark_width = coeffs[g+1].barks - coeffs->barks; |
353 | 0 | coeff->spread_low[0] = ff_exp10(-bark_width * PSY_3GPP_THR_SPREAD_LOW); |
354 | 0 | coeff->spread_hi [0] = ff_exp10(-bark_width * PSY_3GPP_THR_SPREAD_HI); |
355 | 0 | coeff->spread_low[1] = ff_exp10(-bark_width * en_spread_low); |
356 | 0 | coeff->spread_hi [1] = ff_exp10(-bark_width * en_spread_hi); |
357 | 0 | pe_min = bark_pe * bark_width; |
358 | 0 | minsnr = exp2(pe_min / band_sizes[g]) - 1.5f; |
359 | 0 | coeff->min_snr = av_clipf(1.0f / minsnr, PSY_SNR_25DB, PSY_SNR_1DB); |
360 | 0 | } |
361 | 0 | start = 0; |
362 | 0 | for (g = 0; g < ctx->num_bands[j]; g++) { |
363 | 0 | minscale = ath(start * line_to_frequency, ATH_ADD); |
364 | 0 | for (i = 1; i < band_sizes[g]; i++) |
365 | 0 | minscale = FFMIN(minscale, ath((start + i) * line_to_frequency, ATH_ADD)); |
366 | 0 | coeffs[g].ath = minscale - minath; |
367 | 0 | start += band_sizes[g]; |
368 | 0 | } |
369 | 0 | } |
370 | |
|
371 | 0 | pctx->ch = av_calloc(ctx->avctx->ch_layout.nb_channels, sizeof(*pctx->ch)); |
372 | 0 | if (!pctx->ch) { |
373 | 0 | av_freep(&ctx->model_priv_data); |
374 | 0 | return AVERROR(ENOMEM); |
375 | 0 | } |
376 | | |
377 | 0 | lame_window_init(pctx, ctx->avctx); |
378 | |
|
379 | 0 | return 0; |
380 | 0 | } |
381 | | |
382 | | /** |
383 | | * IIR filter used in block switching decision |
384 | | */ |
385 | | static float iir_filter(int in, float state[2]) |
386 | 0 | { |
387 | 0 | float ret; |
388 | 0 |
|
389 | 0 | ret = 0.7548f * (in - state[0]) + 0.5095f * state[1]; |
390 | 0 | state[0] = in; |
391 | 0 | state[1] = ret; |
392 | 0 | return ret; |
393 | 0 | } |
394 | | |
395 | | /** |
396 | | * window grouping information stored as bits (0 - new group, 1 - group continues) |
397 | | */ |
398 | | static const uint8_t window_grouping[9] = { |
399 | | 0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36 |
400 | | }; |
401 | | |
402 | | /** |
403 | | * Tell encoder which window types to use. |
404 | | * @see 3GPP TS26.403 5.4.1 "Blockswitching" |
405 | | */ |
406 | | av_unused static FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx, |
407 | | const int16_t *audio, |
408 | | const int16_t *la, |
409 | | int channel, int prev_type) |
410 | 0 | { |
411 | 0 | int i, j; |
412 | 0 | int br = ((AacPsyContext*)ctx->model_priv_data)->chan_bitrate; |
413 | 0 | int attack_ratio = br <= 16000 ? 18 : 10; |
414 | 0 | AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data; |
415 | 0 | AacPsyChannel *pch = &pctx->ch[channel]; |
416 | 0 | uint8_t grouping = 0; |
417 | 0 | int next_type = pch->next_window_seq; |
418 | 0 | FFPsyWindowInfo wi = { { 0 } }; |
419 | 0 |
|
420 | 0 | if (la) { |
421 | 0 | float s[8], v; |
422 | 0 | int switch_to_eight = 0; |
423 | 0 | float sum = 0.0, sum2 = 0.0; |
424 | 0 | int attack_n = 0; |
425 | 0 | int stay_short = 0; |
426 | 0 | for (i = 0; i < 8; i++) { |
427 | 0 | for (j = 0; j < 128; j++) { |
428 | 0 | v = iir_filter(la[i*128+j], pch->iir_state); |
429 | 0 | sum += v*v; |
430 | 0 | } |
431 | 0 | s[i] = sum; |
432 | 0 | sum2 += sum; |
433 | 0 | } |
434 | 0 | for (i = 0; i < 8; i++) { |
435 | 0 | if (s[i] > pch->win_energy * attack_ratio) { |
436 | 0 | attack_n = i + 1; |
437 | 0 | switch_to_eight = 1; |
438 | 0 | break; |
439 | 0 | } |
440 | 0 | } |
441 | 0 | pch->win_energy = pch->win_energy*7/8 + sum2/64; |
442 | 0 |
|
443 | 0 | wi.window_type[1] = prev_type; |
444 | 0 | switch (prev_type) { |
445 | 0 | case ONLY_LONG_SEQUENCE: |
446 | 0 | wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE; |
447 | 0 | next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE; |
448 | 0 | break; |
449 | 0 | case LONG_START_SEQUENCE: |
450 | 0 | wi.window_type[0] = EIGHT_SHORT_SEQUENCE; |
451 | 0 | grouping = pch->next_grouping; |
452 | 0 | next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE; |
453 | 0 | break; |
454 | 0 | case LONG_STOP_SEQUENCE: |
455 | 0 | wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE; |
456 | 0 | next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE; |
457 | 0 | break; |
458 | 0 | case EIGHT_SHORT_SEQUENCE: |
459 | 0 | stay_short = next_type == EIGHT_SHORT_SEQUENCE || switch_to_eight; |
460 | 0 | wi.window_type[0] = stay_short ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE; |
461 | 0 | grouping = next_type == EIGHT_SHORT_SEQUENCE ? pch->next_grouping : 0; |
462 | 0 | next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE; |
463 | 0 | break; |
464 | 0 | } |
465 | 0 |
|
466 | 0 | pch->next_grouping = window_grouping[attack_n]; |
467 | 0 | pch->next_window_seq = next_type; |
468 | 0 | } else { |
469 | 0 | for (i = 0; i < 3; i++) |
470 | 0 | wi.window_type[i] = prev_type; |
471 | 0 | grouping = (prev_type == EIGHT_SHORT_SEQUENCE) ? window_grouping[0] : 0; |
472 | 0 | } |
473 | 0 |
|
474 | 0 | wi.window_shape = 1; |
475 | 0 | if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) { |
476 | 0 | wi.num_windows = 1; |
477 | 0 | wi.grouping[0] = 1; |
478 | 0 | } else { |
479 | 0 | int lastgrp = 0; |
480 | 0 | wi.num_windows = 8; |
481 | 0 | for (i = 0; i < 8; i++) { |
482 | 0 | if (!((grouping >> i) & 1)) |
483 | 0 | lastgrp = i; |
484 | 0 | wi.grouping[lastgrp]++; |
485 | 0 | } |
486 | 0 | } |
487 | 0 |
|
488 | 0 | return wi; |
489 | 0 | } |
490 | | |
491 | | /* 5.6.1.2 "Calculation of Bit Demand" */ |
492 | | static int calc_bit_demand(AacPsyContext *ctx, float pe, int bits, int size, |
493 | | int short_window) |
494 | 0 | { |
495 | 0 | const float bitsave_slope = short_window ? PSY_3GPP_SAVE_SLOPE_S : PSY_3GPP_SAVE_SLOPE_L; |
496 | 0 | const float bitsave_add = short_window ? PSY_3GPP_SAVE_ADD_S : PSY_3GPP_SAVE_ADD_L; |
497 | 0 | const float bitspend_slope = short_window ? PSY_3GPP_SPEND_SLOPE_S : PSY_3GPP_SPEND_SLOPE_L; |
498 | 0 | const float bitspend_add = short_window ? PSY_3GPP_SPEND_ADD_S : PSY_3GPP_SPEND_ADD_L; |
499 | 0 | const float clip_low = short_window ? PSY_3GPP_CLIP_LO_S : PSY_3GPP_CLIP_LO_L; |
500 | 0 | const float clip_high = short_window ? PSY_3GPP_CLIP_HI_S : PSY_3GPP_CLIP_HI_L; |
501 | 0 | float clipped_pe, bit_save, bit_spend, bit_factor, fill_level, forgetful_min_pe; |
502 | |
|
503 | 0 | ctx->fill_level += ctx->frame_bits - bits; |
504 | 0 | ctx->fill_level = av_clip(ctx->fill_level, 0, size); |
505 | 0 | fill_level = av_clipf((float)ctx->fill_level / size, clip_low, clip_high); |
506 | 0 | clipped_pe = av_clipf(pe, ctx->pe.min, ctx->pe.max); |
507 | 0 | bit_save = (fill_level + bitsave_add) * bitsave_slope; |
508 | 0 | assert(bit_save <= 0.3f && bit_save >= -0.05000001f); |
509 | 0 | bit_spend = (fill_level + bitspend_add) * bitspend_slope; |
510 | 0 | assert(bit_spend <= 0.5f && bit_spend >= -0.1f); |
511 | | /* The bit factor graph in the spec is obviously incorrect. |
512 | | * bit_spend + ((bit_spend - bit_spend))... |
513 | | * The reference encoder subtracts everything from 1, but also seems incorrect. |
514 | | * 1 - bit_save + ((bit_spend + bit_save))... |
515 | | * Hopefully below is correct. |
516 | | */ |
517 | 0 | bit_factor = 1.0f - bit_save + ((bit_spend - bit_save) / (ctx->pe.max - ctx->pe.min)) * (clipped_pe - ctx->pe.min); |
518 | | /* NOTE: The reference encoder attempts to center pe max/min around the current pe. |
519 | | * Here we do that by slowly forgetting pe.min when pe stays in a range that makes |
520 | | * it unlikely (ie: above the mean) |
521 | | */ |
522 | 0 | ctx->pe.max = FFMAX(pe, ctx->pe.max); |
523 | 0 | forgetful_min_pe = ((ctx->pe.min * PSY_PE_FORGET_SLOPE) |
524 | 0 | + FFMAX(ctx->pe.min, pe * (pe / ctx->pe.max))) / (PSY_PE_FORGET_SLOPE + 1); |
525 | 0 | ctx->pe.min = FFMIN(pe, forgetful_min_pe); |
526 | | |
527 | | /* NOTE: allocate a minimum of 1/8th average frame bits, to avoid |
528 | | * reservoir starvation from producing zero-bit frames |
529 | | */ |
530 | 0 | return FFMIN( |
531 | 0 | ctx->frame_bits * bit_factor, |
532 | 0 | FFMAX(ctx->frame_bits + size - bits, ctx->frame_bits / 8)); |
533 | 0 | } |
534 | | |
535 | | static float calc_pe_3gpp(AacPsyBand *band) |
536 | 0 | { |
537 | 0 | float pe, a; |
538 | |
|
539 | 0 | band->pe = 0.0f; |
540 | 0 | band->pe_const = 0.0f; |
541 | 0 | band->active_lines = 0.0f; |
542 | 0 | if (band->energy > band->thr) { |
543 | 0 | a = log2f(band->energy); |
544 | 0 | pe = a - log2f(band->thr); |
545 | 0 | band->active_lines = band->nz_lines; |
546 | 0 | if (pe < PSY_3GPP_C1) { |
547 | 0 | pe = pe * PSY_3GPP_C3 + PSY_3GPP_C2; |
548 | 0 | a = a * PSY_3GPP_C3 + PSY_3GPP_C2; |
549 | 0 | band->active_lines *= PSY_3GPP_C3; |
550 | 0 | } |
551 | 0 | band->pe = pe * band->nz_lines; |
552 | 0 | band->pe_const = a * band->nz_lines; |
553 | 0 | } |
554 | |
|
555 | 0 | return band->pe; |
556 | 0 | } |
557 | | |
558 | | static float calc_reduction_3gpp(float a, float desired_pe, float pe, |
559 | | float active_lines) |
560 | 0 | { |
561 | 0 | float thr_avg, reduction; |
562 | |
|
563 | 0 | if(active_lines == 0.0) |
564 | 0 | return 0; |
565 | | |
566 | 0 | thr_avg = exp2f((a - pe) / (4.0f * active_lines)); |
567 | 0 | reduction = exp2f((a - desired_pe) / (4.0f * active_lines)) - thr_avg; |
568 | |
|
569 | 0 | return FFMAX(reduction, 0.0f); |
570 | 0 | } |
571 | | |
572 | | static float calc_reduced_thr_3gpp(AacPsyBand *band, float min_snr, |
573 | | float reduction) |
574 | 0 | { |
575 | 0 | float thr = band->thr; |
576 | |
|
577 | 0 | if (band->energy > thr) { |
578 | 0 | thr = sqrtf(thr); |
579 | 0 | thr = sqrtf(thr) + reduction; |
580 | 0 | thr *= thr; |
581 | 0 | thr *= thr; |
582 | | |
583 | | /* This deviates from the 3GPP spec to match the reference encoder. |
584 | | * It performs min(thr_reduced, max(thr, energy/min_snr)) only for bands |
585 | | * that have hole avoidance on (active or inactive). It always reduces the |
586 | | * threshold of bands with hole avoidance off. |
587 | | */ |
588 | 0 | if (thr > band->energy * min_snr && band->avoid_holes != PSY_3GPP_AH_NONE) { |
589 | 0 | thr = FFMAX(band->thr, band->energy * min_snr); |
590 | 0 | band->avoid_holes = PSY_3GPP_AH_ACTIVE; |
591 | 0 | } |
592 | 0 | } |
593 | |
|
594 | 0 | return thr; |
595 | 0 | } |
596 | | |
597 | | static void calc_thr_3gpp(const FFPsyWindowInfo *wi, const int num_bands, AacPsyChannel *pch, |
598 | | const uint8_t *band_sizes, const float *coefs, const int cutoff) |
599 | 0 | { |
600 | 0 | int i, w, g; |
601 | 0 | int start = 0, wstart = 0; |
602 | 0 | for (w = 0; w < wi->num_windows*16; w += 16) { |
603 | 0 | wstart = 0; |
604 | 0 | for (g = 0; g < num_bands; g++) { |
605 | 0 | AacPsyBand *band = &pch->band[w+g]; |
606 | |
|
607 | 0 | float form_factor = 0.0f; |
608 | 0 | float Temp; |
609 | 0 | band->energy = 0.0f; |
610 | 0 | if (wstart < cutoff) { |
611 | 0 | for (i = 0; i < band_sizes[g]; i++) { |
612 | 0 | band->energy += coefs[start+i] * coefs[start+i]; |
613 | 0 | form_factor += sqrtf(fabs(coefs[start+i])); |
614 | 0 | } |
615 | 0 | } |
616 | 0 | Temp = band->energy > 0 ? sqrtf((float)band_sizes[g] / band->energy) : 0; |
617 | 0 | band->thr = band->energy * 0.001258925f; |
618 | 0 | band->nz_lines = form_factor * sqrtf(Temp); |
619 | |
|
620 | 0 | start += band_sizes[g]; |
621 | 0 | wstart += band_sizes[g]; |
622 | 0 | } |
623 | 0 | } |
624 | 0 | } |
625 | | |
626 | | static void psy_hp_filter(const float *firbuf, float *hpfsmpl, const float *psy_fir_coeffs) |
627 | 0 | { |
628 | 0 | int i, j; |
629 | 0 | for (i = 0; i < AAC_BLOCK_SIZE_LONG; i++) { |
630 | 0 | float sum1, sum2; |
631 | 0 | sum1 = firbuf[i + (PSY_LAME_FIR_LEN - 1) / 2]; |
632 | 0 | sum2 = 0.0; |
633 | 0 | for (j = 0; j < ((PSY_LAME_FIR_LEN - 1) / 2) - 1; j += 2) { |
634 | 0 | sum1 += psy_fir_coeffs[j] * (firbuf[i + j] + firbuf[i + PSY_LAME_FIR_LEN - j]); |
635 | 0 | sum2 += psy_fir_coeffs[j + 1] * (firbuf[i + j + 1] + firbuf[i + PSY_LAME_FIR_LEN - j - 1]); |
636 | 0 | } |
637 | | /* NOTE: The LAME psymodel expects it's input in the range -32768 to 32768. |
638 | | * Tuning this for normalized floats would be difficult. */ |
639 | 0 | hpfsmpl[i] = (sum1 + sum2) * 32768.0f; |
640 | 0 | } |
641 | 0 | } |
642 | | |
643 | | /** |
644 | | * Calculate band thresholds as suggested in 3GPP TS26.403 |
645 | | */ |
646 | | static void psy_3gpp_analyze_channel(FFPsyContext *ctx, int channel, |
647 | | const float *coefs, const FFPsyWindowInfo *wi) |
648 | 0 | { |
649 | 0 | AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data; |
650 | 0 | AacPsyChannel *pch = &pctx->ch[channel]; |
651 | 0 | int i, w, g; |
652 | 0 | float desired_bits, desired_pe, delta_pe, reduction= NAN, spread_en[128] = {0}; |
653 | 0 | float a = 0.0f, active_lines = 0.0f, norm_fac = 0.0f; |
654 | 0 | float pe = pctx->chan_bitrate > 32000 ? 0.0f : FFMAX(50.0f, 100.0f - pctx->chan_bitrate * 100.0f / 32000.0f); |
655 | 0 | const int num_bands = ctx->num_bands[wi->num_windows == 8]; |
656 | 0 | const uint8_t *band_sizes = ctx->bands[wi->num_windows == 8]; |
657 | 0 | AacPsyCoeffs *coeffs = pctx->psy_coef[wi->num_windows == 8]; |
658 | 0 | const float avoid_hole_thr = wi->num_windows == 8 ? PSY_3GPP_AH_THR_SHORT : PSY_3GPP_AH_THR_LONG; |
659 | 0 | const int bandwidth = ctx->cutoff ? ctx->cutoff : AAC_CUTOFF(ctx->avctx); |
660 | 0 | const int cutoff = bandwidth * 2048 / wi->num_windows / ctx->avctx->sample_rate; |
661 | | |
662 | | //calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation" |
663 | 0 | calc_thr_3gpp(wi, num_bands, pch, band_sizes, coefs, cutoff); |
664 | | |
665 | | //modify thresholds and energies - spread, threshold in quiet, pre-echo control |
666 | 0 | for (w = 0; w < wi->num_windows*16; w += 16) { |
667 | 0 | AacPsyBand *bands = &pch->band[w]; |
668 | | |
669 | | /* 5.4.2.3 "Spreading" & 5.4.3 "Spread Energy Calculation" */ |
670 | 0 | spread_en[0] = bands[0].energy; |
671 | 0 | for (g = 1; g < num_bands; g++) { |
672 | 0 | bands[g].thr = FFMAX(bands[g].thr, bands[g-1].thr * coeffs[g].spread_hi[0]); |
673 | 0 | spread_en[w+g] = FFMAX(bands[g].energy, spread_en[w+g-1] * coeffs[g].spread_hi[1]); |
674 | 0 | } |
675 | 0 | for (g = num_bands - 2; g >= 0; g--) { |
676 | 0 | bands[g].thr = FFMAX(bands[g].thr, bands[g+1].thr * coeffs[g].spread_low[0]); |
677 | 0 | spread_en[w+g] = FFMAX(spread_en[w+g], spread_en[w+g+1] * coeffs[g].spread_low[1]); |
678 | 0 | } |
679 | | //5.4.2.4 "Threshold in quiet" |
680 | 0 | for (g = 0; g < num_bands; g++) { |
681 | 0 | AacPsyBand *band = &bands[g]; |
682 | |
|
683 | 0 | band->thr_quiet = band->thr = FFMAX(band->thr, coeffs[g].ath); |
684 | | //5.4.2.5 "Pre-echo control" |
685 | 0 | if (!(wi->window_type[0] == LONG_STOP_SEQUENCE || (!w && wi->window_type[1] == LONG_START_SEQUENCE))) |
686 | 0 | band->thr = FFMAX(PSY_3GPP_RPEMIN*band->thr, FFMIN(band->thr, |
687 | 0 | PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet)); |
688 | | |
689 | | /* 5.6.1.3.1 "Preparatory steps of the perceptual entropy calculation" */ |
690 | 0 | pe += calc_pe_3gpp(band); |
691 | 0 | a += band->pe_const; |
692 | 0 | active_lines += band->active_lines; |
693 | | |
694 | | /* 5.6.1.3.3 "Selection of the bands for avoidance of holes" */ |
695 | 0 | if (spread_en[w+g] * avoid_hole_thr > band->energy || coeffs[g].min_snr > 1.0f) |
696 | 0 | band->avoid_holes = PSY_3GPP_AH_NONE; |
697 | 0 | else |
698 | 0 | band->avoid_holes = PSY_3GPP_AH_INACTIVE; |
699 | 0 | } |
700 | 0 | } |
701 | | |
702 | | /* 5.6.1.3.2 "Calculation of the desired perceptual entropy" */ |
703 | 0 | ctx->ch[channel].entropy = pe; |
704 | 0 | if (ctx->avctx->flags & AV_CODEC_FLAG_QSCALE) { |
705 | | /* (2.5 * 120) achieves almost transparent rate, and we want to give |
706 | | * ample room downwards, so we make that equivalent to QSCALE=2.4 |
707 | | */ |
708 | 0 | desired_pe = pe * (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120) / (2 * 2.5f * 120.0f); |
709 | 0 | desired_bits = FFMIN(2560, PSY_3GPP_PE_TO_BITS(desired_pe)); |
710 | 0 | desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits); // reflect clipping |
711 | | |
712 | | /* PE slope smoothing */ |
713 | 0 | if (ctx->bitres.bits > 0) { |
714 | 0 | desired_bits = FFMIN(2560, PSY_3GPP_PE_TO_BITS(desired_pe)); |
715 | 0 | desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits); // reflect clipping |
716 | 0 | } |
717 | |
|
718 | 0 | pctx->pe.max = FFMAX(pe, pctx->pe.max); |
719 | 0 | pctx->pe.min = FFMIN(pe, pctx->pe.min); |
720 | 0 | } else { |
721 | 0 | desired_bits = calc_bit_demand(pctx, pe, ctx->bitres.bits, ctx->bitres.size, wi->num_windows == 8); |
722 | 0 | desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits); |
723 | | |
724 | | /* NOTE: PE correction is kept simple. During initial testing it had very |
725 | | * little effect on the final bitrate. Probably a good idea to come |
726 | | * back and do more testing later. |
727 | | */ |
728 | 0 | if (ctx->bitres.bits > 0) |
729 | 0 | desired_pe *= av_clipf(pctx->pe.previous / PSY_3GPP_BITS_TO_PE(ctx->bitres.bits), |
730 | 0 | 0.85f, 1.15f); |
731 | 0 | } |
732 | 0 | pctx->pe.previous = PSY_3GPP_BITS_TO_PE(desired_bits); |
733 | 0 | ctx->bitres.alloc = desired_bits; |
734 | |
|
735 | 0 | if (desired_pe < pe) { |
736 | | /* 5.6.1.3.4 "First Estimation of the reduction value" */ |
737 | 0 | for (w = 0; w < wi->num_windows*16; w += 16) { |
738 | 0 | reduction = calc_reduction_3gpp(a, desired_pe, pe, active_lines); |
739 | 0 | pe = 0.0f; |
740 | 0 | a = 0.0f; |
741 | 0 | active_lines = 0.0f; |
742 | 0 | for (g = 0; g < num_bands; g++) { |
743 | 0 | AacPsyBand *band = &pch->band[w+g]; |
744 | |
|
745 | 0 | band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction); |
746 | | /* recalculate PE */ |
747 | 0 | pe += calc_pe_3gpp(band); |
748 | 0 | a += band->pe_const; |
749 | 0 | active_lines += band->active_lines; |
750 | 0 | } |
751 | 0 | } |
752 | | |
753 | | /* 5.6.1.3.5 "Second Estimation of the reduction value" */ |
754 | 0 | for (i = 0; i < 2; i++) { |
755 | 0 | float pe_no_ah = 0.0f, desired_pe_no_ah; |
756 | 0 | active_lines = a = 0.0f; |
757 | 0 | for (w = 0; w < wi->num_windows*16; w += 16) { |
758 | 0 | for (g = 0; g < num_bands; g++) { |
759 | 0 | AacPsyBand *band = &pch->band[w+g]; |
760 | |
|
761 | 0 | if (band->avoid_holes != PSY_3GPP_AH_ACTIVE) { |
762 | 0 | pe_no_ah += band->pe; |
763 | 0 | a += band->pe_const; |
764 | 0 | active_lines += band->active_lines; |
765 | 0 | } |
766 | 0 | } |
767 | 0 | } |
768 | 0 | desired_pe_no_ah = FFMAX(desired_pe - (pe - pe_no_ah), 0.0f); |
769 | 0 | if (active_lines > 0.0f) |
770 | 0 | reduction = calc_reduction_3gpp(a, desired_pe_no_ah, pe_no_ah, active_lines); |
771 | |
|
772 | 0 | pe = 0.0f; |
773 | 0 | for (w = 0; w < wi->num_windows*16; w += 16) { |
774 | 0 | for (g = 0; g < num_bands; g++) { |
775 | 0 | AacPsyBand *band = &pch->band[w+g]; |
776 | |
|
777 | 0 | if (active_lines > 0.0f) |
778 | 0 | band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction); |
779 | 0 | pe += calc_pe_3gpp(band); |
780 | 0 | if (band->thr > 0.0f) |
781 | 0 | band->norm_fac = band->active_lines / band->thr; |
782 | 0 | else |
783 | 0 | band->norm_fac = 0.0f; |
784 | 0 | norm_fac += band->norm_fac; |
785 | 0 | } |
786 | 0 | } |
787 | 0 | delta_pe = desired_pe - pe; |
788 | 0 | if (fabs(delta_pe) > 0.05f * desired_pe) |
789 | 0 | break; |
790 | 0 | } |
791 | |
|
792 | 0 | if (pe < 1.15f * desired_pe) { |
793 | | /* 6.6.1.3.6 "Final threshold modification by linearization" */ |
794 | 0 | norm_fac = norm_fac ? 1.0f / norm_fac : 0; |
795 | 0 | for (w = 0; w < wi->num_windows*16; w += 16) { |
796 | 0 | for (g = 0; g < num_bands; g++) { |
797 | 0 | AacPsyBand *band = &pch->band[w+g]; |
798 | |
|
799 | 0 | if (band->active_lines > 0.5f) { |
800 | 0 | float delta_sfb_pe = band->norm_fac * norm_fac * delta_pe; |
801 | 0 | float thr = band->thr; |
802 | |
|
803 | 0 | thr *= exp2f(delta_sfb_pe / band->active_lines); |
804 | 0 | if (thr > coeffs[g].min_snr * band->energy && band->avoid_holes == PSY_3GPP_AH_INACTIVE) |
805 | 0 | thr = FFMAX(band->thr, coeffs[g].min_snr * band->energy); |
806 | 0 | band->thr = thr; |
807 | 0 | } |
808 | 0 | } |
809 | 0 | } |
810 | 0 | } else { |
811 | | /* 5.6.1.3.7 "Further perceptual entropy reduction" */ |
812 | 0 | g = num_bands; |
813 | 0 | while (pe > desired_pe && g--) { |
814 | 0 | for (w = 0; w < wi->num_windows*16; w+= 16) { |
815 | 0 | AacPsyBand *band = &pch->band[w+g]; |
816 | 0 | if (band->avoid_holes != PSY_3GPP_AH_NONE && coeffs[g].min_snr < PSY_SNR_1DB) { |
817 | 0 | coeffs[g].min_snr = PSY_SNR_1DB; |
818 | 0 | band->thr = band->energy * PSY_SNR_1DB; |
819 | 0 | pe += band->active_lines * 1.5f - band->pe; |
820 | 0 | } |
821 | 0 | } |
822 | 0 | } |
823 | | /* TODO: allow more holes (unused without mid/side) */ |
824 | 0 | } |
825 | 0 | } |
826 | |
|
827 | 0 | for (w = 0; w < wi->num_windows*16; w += 16) { |
828 | 0 | for (g = 0; g < num_bands; g++) { |
829 | 0 | AacPsyBand *band = &pch->band[w+g]; |
830 | 0 | FFPsyBand *psy_band = &ctx->ch[channel].psy_bands[w+g]; |
831 | |
|
832 | 0 | psy_band->threshold = band->thr; |
833 | 0 | psy_band->energy = band->energy; |
834 | 0 | psy_band->spread = band->active_lines * 2.0f / band_sizes[g]; |
835 | 0 | psy_band->bits = PSY_3GPP_PE_TO_BITS(band->pe); |
836 | 0 | } |
837 | 0 | } |
838 | |
|
839 | 0 | memcpy(pch->prev_band, pch->band, sizeof(pch->band)); |
840 | 0 | } |
841 | | |
842 | | static void psy_3gpp_analyze(FFPsyContext *ctx, int channel, |
843 | | const float **coeffs, const FFPsyWindowInfo *wi) |
844 | 0 | { |
845 | 0 | int ch; |
846 | 0 | FFPsyChannelGroup *group = ff_psy_find_group(ctx, channel); |
847 | |
|
848 | 0 | for (ch = 0; ch < group->num_ch; ch++) |
849 | 0 | psy_3gpp_analyze_channel(ctx, channel + ch, coeffs[ch], &wi[ch]); |
850 | 0 | } |
851 | | |
852 | | static av_cold void psy_3gpp_end(FFPsyContext *apc) |
853 | 0 | { |
854 | 0 | AacPsyContext *pctx = (AacPsyContext*) apc->model_priv_data; |
855 | 0 | if (pctx) |
856 | 0 | av_freep(&pctx->ch); |
857 | 0 | av_freep(&apc->model_priv_data); |
858 | 0 | } |
859 | | |
860 | | static void lame_apply_block_type(AacPsyChannel *ctx, FFPsyWindowInfo *wi, int uselongblock) |
861 | 0 | { |
862 | 0 | int blocktype = ONLY_LONG_SEQUENCE; |
863 | 0 | if (uselongblock) { |
864 | 0 | if (ctx->next_window_seq == EIGHT_SHORT_SEQUENCE) |
865 | 0 | blocktype = LONG_STOP_SEQUENCE; |
866 | 0 | } else { |
867 | 0 | blocktype = EIGHT_SHORT_SEQUENCE; |
868 | 0 | if (ctx->next_window_seq == ONLY_LONG_SEQUENCE) |
869 | 0 | ctx->next_window_seq = LONG_START_SEQUENCE; |
870 | 0 | if (ctx->next_window_seq == LONG_STOP_SEQUENCE) |
871 | 0 | ctx->next_window_seq = EIGHT_SHORT_SEQUENCE; |
872 | 0 | } |
873 | |
|
874 | 0 | wi->window_type[0] = ctx->next_window_seq; |
875 | 0 | ctx->next_window_seq = blocktype; |
876 | 0 | } |
877 | | |
878 | | static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx, const float *audio, |
879 | | const float *la, int channel, int prev_type) |
880 | 0 | { |
881 | 0 | AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data; |
882 | 0 | AacPsyChannel *pch = &pctx->ch[channel]; |
883 | 0 | int grouping = 0; |
884 | 0 | int uselongblock = 1; |
885 | 0 | int attacks[AAC_NUM_BLOCKS_SHORT + 1] = { 0 }; |
886 | 0 | int i; |
887 | 0 | FFPsyWindowInfo wi = { { 0 } }; |
888 | |
|
889 | 0 | if (la) { |
890 | 0 | float hpfsmpl[AAC_BLOCK_SIZE_LONG]; |
891 | 0 | const float *pf = hpfsmpl; |
892 | 0 | float attack_intensity[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS]; |
893 | 0 | float energy_subshort[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS]; |
894 | 0 | float energy_short[AAC_NUM_BLOCKS_SHORT + 1] = { 0 }; |
895 | 0 | const float *firbuf = la + (AAC_BLOCK_SIZE_SHORT/4 - PSY_LAME_FIR_LEN); |
896 | 0 | int att_sum = 0; |
897 | | |
898 | | /* LAME comment: apply high pass filter of fs/4 */ |
899 | 0 | psy_hp_filter(firbuf, hpfsmpl, psy_fir_coeffs); |
900 | | |
901 | | /* Calculate the energies of each sub-shortblock */ |
902 | 0 | for (i = 0; i < PSY_LAME_NUM_SUBBLOCKS; i++) { |
903 | 0 | energy_subshort[i] = pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) * PSY_LAME_NUM_SUBBLOCKS)]; |
904 | 0 | assert(pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) * PSY_LAME_NUM_SUBBLOCKS - 2)] > 0); |
905 | 0 | attack_intensity[i] = energy_subshort[i] / pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) * PSY_LAME_NUM_SUBBLOCKS - 2)]; |
906 | 0 | energy_short[0] += energy_subshort[i]; |
907 | 0 | } |
908 | |
|
909 | 0 | for (i = 0; i < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; i++) { |
910 | 0 | const float *const pfe = pf + AAC_BLOCK_SIZE_LONG / (AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS); |
911 | 0 | float p = 1.0f; |
912 | 0 | for (; pf < pfe; pf++) |
913 | 0 | p = FFMAX(p, fabsf(*pf)); |
914 | 0 | pch->prev_energy_subshort[i] = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS] = p; |
915 | 0 | energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p; |
916 | | |
917 | | /* NOTE: The indexes below are [i + 3 - 2] in the LAME source. Compare each sub-block to sub-block - 2 */ |
918 | 0 | if (p > energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS - 2]) |
919 | 0 | p = p / energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS - 2]; |
920 | 0 | else if (energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS - 2] > p * 10.0f) |
921 | 0 | p = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS - 2] / (p * 10.0f); |
922 | 0 | else |
923 | 0 | p = 0.0; |
924 | |
|
925 | 0 | attack_intensity[i + PSY_LAME_NUM_SUBBLOCKS] = p; |
926 | 0 | } |
927 | | |
928 | | /* compare energy between sub-short blocks */ |
929 | 0 | for (i = 0; i < (AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS; i++) |
930 | 0 | if (!attacks[i / PSY_LAME_NUM_SUBBLOCKS]) |
931 | 0 | if (attack_intensity[i] > pch->attack_threshold) |
932 | 0 | attacks[i / PSY_LAME_NUM_SUBBLOCKS] = (i % PSY_LAME_NUM_SUBBLOCKS) + 1; |
933 | | |
934 | | /* should have energy change between short blocks, in order to avoid periodic signals */ |
935 | | /* Good samples to show the effect are Trumpet test songs */ |
936 | | /* GB: tuned (1) to avoid too many short blocks for test sample TRUMPET */ |
937 | | /* RH: tuned (2) to let enough short blocks through for test sample FSOL and SNAPS */ |
938 | 0 | for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) { |
939 | 0 | const float u = energy_short[i - 1]; |
940 | 0 | const float v = energy_short[i]; |
941 | 0 | const float m = FFMAX(u, v); |
942 | 0 | if (m < 40000) { /* (2) */ |
943 | 0 | if (u < 2.3f * v && v < 2.3f * u) { /* (1) */ |
944 | 0 | if (i == 1 && attacks[0] < attacks[i]) |
945 | 0 | attacks[0] = 0; |
946 | 0 | attacks[i] = 0; |
947 | 0 | } |
948 | 0 | } |
949 | 0 | att_sum += attacks[i]; |
950 | 0 | } |
951 | |
|
952 | 0 | if (pch->next_attack0_zero) |
953 | 0 | attacks[0] = 0; |
954 | 0 | pch->next_attack0_zero = !attacks[AAC_NUM_BLOCKS_SHORT]; |
955 | |
|
956 | 0 | if (attacks[0] <= pch->prev_attack) |
957 | 0 | attacks[0] = 0; |
958 | |
|
959 | 0 | att_sum += attacks[0]; |
960 | | |
961 | | /* If the previous attack happened in the last sub-block of the previous sequence, |
962 | | * or if there's a new attack, use short window */ |
963 | 0 | if (pch->prev_attack == PSY_LAME_NUM_SUBBLOCKS || att_sum) { |
964 | 0 | uselongblock = 0; |
965 | |
|
966 | 0 | for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) |
967 | 0 | if (attacks[i] && attacks[i-1]) |
968 | 0 | attacks[i] = 0; |
969 | 0 | } |
970 | 0 | } else { |
971 | | /* We have no lookahead info, so just use same type as the previous sequence. */ |
972 | 0 | uselongblock = !(prev_type == EIGHT_SHORT_SEQUENCE); |
973 | 0 | } |
974 | |
|
975 | 0 | lame_apply_block_type(pch, &wi, uselongblock); |
976 | |
|
977 | 0 | wi.window_type[1] = prev_type; |
978 | 0 | if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) { |
979 | |
|
980 | 0 | wi.num_windows = 1; |
981 | 0 | wi.grouping[0] = 1; |
982 | 0 | if (wi.window_type[0] == LONG_START_SEQUENCE) |
983 | 0 | wi.window_shape = 0; |
984 | 0 | else |
985 | 0 | wi.window_shape = 1; |
986 | |
|
987 | 0 | } else { |
988 | 0 | int lastgrp = 0; |
989 | |
|
990 | 0 | wi.num_windows = 8; |
991 | 0 | wi.window_shape = 0; |
992 | 0 | for (i = 0; i < 8; i++) { |
993 | 0 | if (!((pch->next_grouping >> i) & 1)) |
994 | 0 | lastgrp = i; |
995 | 0 | wi.grouping[lastgrp]++; |
996 | 0 | } |
997 | 0 | } |
998 | | |
999 | | /* Determine grouping, based on the location of the first attack, and save for |
1000 | | * the next frame. |
1001 | | * FIXME: Move this to analysis. |
1002 | | * TODO: Tune groupings depending on attack location |
1003 | | * TODO: Handle more than one attack in a group |
1004 | | */ |
1005 | 0 | for (i = 0; i < 9; i++) { |
1006 | 0 | if (attacks[i]) { |
1007 | 0 | grouping = i; |
1008 | 0 | break; |
1009 | 0 | } |
1010 | 0 | } |
1011 | 0 | pch->next_grouping = window_grouping[grouping]; |
1012 | |
|
1013 | 0 | pch->prev_attack = attacks[AAC_NUM_BLOCKS_SHORT - 1]; |
1014 | |
|
1015 | 0 | return wi; |
1016 | 0 | } |
1017 | | |
1018 | | const FFPsyModel ff_aac_psy_model = |
1019 | | { |
1020 | | .name = "3GPP TS 26.403-inspired model", |
1021 | | .init = psy_3gpp_init, |
1022 | | .window = psy_lame_window, |
1023 | | .analyze = psy_3gpp_analyze, |
1024 | | .end = psy_3gpp_end, |
1025 | | }; |