blob: fca692cb153f17082e3bcb918841b8e2bbbfe2c4
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 | |
30 | #include "avcodec.h" |
31 | #include "aactab.h" |
32 | #include "psymodel.h" |
33 | |
34 | /*********************************** |
35 | * TODOs: |
36 | * try other bitrate controlling mechanism (maybe use ratecontrol.c?) |
37 | * control quality for quality-based output |
38 | **********************************/ |
39 | |
40 | /** |
41 | * constants for 3GPP AAC psychoacoustic model |
42 | * @{ |
43 | */ |
44 | #define PSY_3GPP_THR_SPREAD_HI 1.5f // spreading factor for low-to-hi threshold spreading (15 dB/Bark) |
45 | #define PSY_3GPP_THR_SPREAD_LOW 3.0f // spreading factor for hi-to-low threshold spreading (30 dB/Bark) |
46 | /* spreading factor for low-to-hi energy spreading, long block, > 22kbps/channel (20dB/Bark) */ |
47 | #define PSY_3GPP_EN_SPREAD_HI_L1 2.0f |
48 | /* spreading factor for low-to-hi energy spreading, long block, <= 22kbps/channel (15dB/Bark) */ |
49 | #define PSY_3GPP_EN_SPREAD_HI_L2 1.5f |
50 | /* spreading factor for low-to-hi energy spreading, short block (15 dB/Bark) */ |
51 | #define PSY_3GPP_EN_SPREAD_HI_S 1.5f |
52 | /* spreading factor for hi-to-low energy spreading, long block (30dB/Bark) */ |
53 | #define PSY_3GPP_EN_SPREAD_LOW_L 3.0f |
54 | /* spreading factor for hi-to-low energy spreading, short block (20dB/Bark) */ |
55 | #define PSY_3GPP_EN_SPREAD_LOW_S 2.0f |
56 | |
57 | #define PSY_3GPP_RPEMIN 0.01f |
58 | #define PSY_3GPP_RPELEV 2.0f |
59 | |
60 | #define PSY_3GPP_C1 3.0f /* log2(8) */ |
61 | #define PSY_3GPP_C2 1.3219281f /* log2(2.5) */ |
62 | #define PSY_3GPP_C3 0.55935729f /* 1 - C2 / C1 */ |
63 | |
64 | #define PSY_SNR_1DB 7.9432821e-1f /* -1dB */ |
65 | #define PSY_SNR_25DB 3.1622776e-3f /* -25dB */ |
66 | |
67 | #define PSY_3GPP_SAVE_SLOPE_L -0.46666667f |
68 | #define PSY_3GPP_SAVE_SLOPE_S -0.36363637f |
69 | #define PSY_3GPP_SAVE_ADD_L -0.84285712f |
70 | #define PSY_3GPP_SAVE_ADD_S -0.75f |
71 | #define PSY_3GPP_SPEND_SLOPE_L 0.66666669f |
72 | #define PSY_3GPP_SPEND_SLOPE_S 0.81818181f |
73 | #define PSY_3GPP_SPEND_ADD_L -0.35f |
74 | #define PSY_3GPP_SPEND_ADD_S -0.26111111f |
75 | #define PSY_3GPP_CLIP_LO_L 0.2f |
76 | #define PSY_3GPP_CLIP_LO_S 0.2f |
77 | #define PSY_3GPP_CLIP_HI_L 0.95f |
78 | #define PSY_3GPP_CLIP_HI_S 0.75f |
79 | |
80 | #define PSY_3GPP_AH_THR_LONG 0.5f |
81 | #define PSY_3GPP_AH_THR_SHORT 0.63f |
82 | |
83 | #define PSY_PE_FORGET_SLOPE 511 |
84 | |
85 | enum { |
86 | PSY_3GPP_AH_NONE, |
87 | PSY_3GPP_AH_INACTIVE, |
88 | PSY_3GPP_AH_ACTIVE |
89 | }; |
90 | |
91 | #define PSY_3GPP_BITS_TO_PE(bits) ((bits) * 1.18f) |
92 | #define PSY_3GPP_PE_TO_BITS(bits) ((bits) / 1.18f) |
93 | |
94 | /* LAME psy model constants */ |
95 | #define PSY_LAME_FIR_LEN 21 ///< LAME psy model FIR order |
96 | #define AAC_BLOCK_SIZE_LONG 1024 ///< long block size |
97 | #define AAC_BLOCK_SIZE_SHORT 128 ///< short block size |
98 | #define AAC_NUM_BLOCKS_SHORT 8 ///< number of blocks in a short sequence |
99 | #define PSY_LAME_NUM_SUBBLOCKS 3 ///< Number of sub-blocks in each short block |
100 | |
101 | /** |
102 | * @} |
103 | */ |
104 | |
105 | /** |
106 | * information for single band used by 3GPP TS26.403-inspired psychoacoustic model |
107 | */ |
108 | typedef struct AacPsyBand{ |
109 | float energy; ///< band energy |
110 | float thr; ///< energy threshold |
111 | float thr_quiet; ///< threshold in quiet |
112 | float nz_lines; ///< number of non-zero spectral lines |
113 | float active_lines; ///< number of active spectral lines |
114 | float pe; ///< perceptual entropy |
115 | float pe_const; ///< constant part of the PE calculation |
116 | float norm_fac; ///< normalization factor for linearization |
117 | int avoid_holes; ///< hole avoidance flag |
118 | }AacPsyBand; |
119 | |
120 | /** |
121 | * single/pair channel context for psychoacoustic model |
122 | */ |
123 | typedef struct AacPsyChannel{ |
124 | AacPsyBand band[128]; ///< bands information |
125 | AacPsyBand prev_band[128]; ///< bands information from the previous frame |
126 | |
127 | float win_energy; ///< sliding average of channel energy |
128 | float iir_state[2]; ///< hi-pass IIR filter state |
129 | uint8_t next_grouping; ///< stored grouping scheme for the next frame (in case of 8 short window sequence) |
130 | enum WindowSequence next_window_seq; ///< window sequence to be used in the next frame |
131 | /* LAME psy model specific members */ |
132 | float attack_threshold; ///< attack threshold for this channel |
133 | float prev_energy_subshort[AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS]; |
134 | int prev_attack; ///< attack value for the last short block in the previous sequence |
135 | }AacPsyChannel; |
136 | |
137 | /** |
138 | * psychoacoustic model frame type-dependent coefficients |
139 | */ |
140 | typedef struct AacPsyCoeffs{ |
141 | float ath; ///< absolute threshold of hearing per bands |
142 | float barks; ///< Bark value for each spectral band in long frame |
143 | float spread_low[2]; ///< spreading factor for low-to-high threshold spreading in long frame |
144 | float spread_hi [2]; ///< spreading factor for high-to-low threshold spreading in long frame |
145 | float min_snr; ///< minimal SNR |
146 | }AacPsyCoeffs; |
147 | |
148 | /** |
149 | * 3GPP TS26.403-inspired psychoacoustic model specific data |
150 | */ |
151 | typedef struct AacPsyContext{ |
152 | int chan_bitrate; ///< bitrate per channel |
153 | int frame_bits; ///< average bits per frame |
154 | int fill_level; ///< bit reservoir fill level |
155 | struct { |
156 | float min; ///< minimum allowed PE for bit factor calculation |
157 | float max; ///< maximum allowed PE for bit factor calculation |
158 | float previous; ///< allowed PE of the previous frame |
159 | float correction; ///< PE correction factor |
160 | } pe; |
161 | AacPsyCoeffs psy_coef[2][64]; |
162 | AacPsyChannel *ch; |
163 | float global_quality; ///< normalized global quality taken from avctx |
164 | }AacPsyContext; |
165 | |
166 | /** |
167 | * LAME psy model preset struct |
168 | */ |
169 | typedef struct PsyLamePreset { |
170 | int quality; ///< Quality to map the rest of the vaules to. |
171 | /* This is overloaded to be both kbps per channel in ABR mode, and |
172 | * requested quality in constant quality mode. |
173 | */ |
174 | float st_lrm; ///< short threshold for L, R, and M channels |
175 | } PsyLamePreset; |
176 | |
177 | /** |
178 | * LAME psy model preset table for ABR |
179 | */ |
180 | static const PsyLamePreset psy_abr_map[] = { |
181 | /* TODO: Tuning. These were taken from LAME. */ |
182 | /* kbps/ch st_lrm */ |
183 | { 8, 6.60}, |
184 | { 16, 6.60}, |
185 | { 24, 6.60}, |
186 | { 32, 6.60}, |
187 | { 40, 6.60}, |
188 | { 48, 6.60}, |
189 | { 56, 6.60}, |
190 | { 64, 6.40}, |
191 | { 80, 6.00}, |
192 | { 96, 5.60}, |
193 | {112, 5.20}, |
194 | {128, 5.20}, |
195 | {160, 5.20} |
196 | }; |
197 | |
198 | /** |
199 | * LAME psy model preset table for constant quality |
200 | */ |
201 | static const PsyLamePreset psy_vbr_map[] = { |
202 | /* vbr_q st_lrm */ |
203 | { 0, 4.20}, |
204 | { 1, 4.20}, |
205 | { 2, 4.20}, |
206 | { 3, 4.20}, |
207 | { 4, 4.20}, |
208 | { 5, 4.20}, |
209 | { 6, 4.20}, |
210 | { 7, 4.20}, |
211 | { 8, 4.20}, |
212 | { 9, 4.20}, |
213 | {10, 4.20} |
214 | }; |
215 | |
216 | /** |
217 | * LAME psy model FIR coefficient table |
218 | */ |
219 | static const float psy_fir_coeffs[] = { |
220 | -8.65163e-18 * 2, -0.00851586 * 2, -6.74764e-18 * 2, 0.0209036 * 2, |
221 | -3.36639e-17 * 2, -0.0438162 * 2, -1.54175e-17 * 2, 0.0931738 * 2, |
222 | -5.52212e-17 * 2, -0.313819 * 2 |
223 | }; |
224 | |
225 | #if ARCH_MIPS |
226 | # include "mips/aacpsy_mips.h" |
227 | #endif /* ARCH_MIPS */ |
228 | |
229 | /** |
230 | * Calculate the ABR attack threshold from the above LAME psymodel table. |
231 | */ |
232 | static float lame_calc_attack_threshold(int bitrate) |
233 | { |
234 | /* Assume max bitrate to start with */ |
235 | int lower_range = 12, upper_range = 12; |
236 | int lower_range_kbps = psy_abr_map[12].quality; |
237 | int upper_range_kbps = psy_abr_map[12].quality; |
238 | int i; |
239 | |
240 | /* Determine which bitrates the value specified falls between. |
241 | * If the loop ends without breaking our above assumption of 320kbps was correct. |
242 | */ |
243 | for (i = 1; i < 13; i++) { |
244 | if (FFMAX(bitrate, psy_abr_map[i].quality) != bitrate) { |
245 | upper_range = i; |
246 | upper_range_kbps = psy_abr_map[i ].quality; |
247 | lower_range = i - 1; |
248 | lower_range_kbps = psy_abr_map[i - 1].quality; |
249 | break; /* Upper range found */ |
250 | } |
251 | } |
252 | |
253 | /* Determine which range the value specified is closer to */ |
254 | if ((upper_range_kbps - bitrate) > (bitrate - lower_range_kbps)) |
255 | return psy_abr_map[lower_range].st_lrm; |
256 | return psy_abr_map[upper_range].st_lrm; |
257 | } |
258 | |
259 | /** |
260 | * LAME psy model specific initialization |
261 | */ |
262 | static av_cold void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx) |
263 | { |
264 | int i, j; |
265 | |
266 | for (i = 0; i < avctx->channels; i++) { |
267 | AacPsyChannel *pch = &ctx->ch[i]; |
268 | |
269 | if (avctx->flags & AV_CODEC_FLAG_QSCALE) |
270 | pch->attack_threshold = psy_vbr_map[avctx->global_quality / FF_QP2LAMBDA].st_lrm; |
271 | else |
272 | pch->attack_threshold = lame_calc_attack_threshold(avctx->bit_rate / avctx->channels / 1000); |
273 | |
274 | for (j = 0; j < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; j++) |
275 | pch->prev_energy_subshort[j] = 10.0f; |
276 | } |
277 | } |
278 | |
279 | /** |
280 | * Calculate Bark value for given line. |
281 | */ |
282 | static av_cold float calc_bark(float f) |
283 | { |
284 | return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f)); |
285 | } |
286 | |
287 | #define ATH_ADD 4 |
288 | /** |
289 | * Calculate ATH value for given frequency. |
290 | * Borrowed from Lame. |
291 | */ |
292 | static av_cold float ath(float f, float add) |
293 | { |
294 | f /= 1000.0f; |
295 | return 3.64 * pow(f, -0.8) |
296 | - 6.8 * exp(-0.6 * (f - 3.4) * (f - 3.4)) |
297 | + 6.0 * exp(-0.15 * (f - 8.7) * (f - 8.7)) |
298 | + (0.6 + 0.04 * add) * 0.001 * f * f * f * f; |
299 | } |
300 | |
301 | static av_cold int psy_3gpp_init(FFPsyContext *ctx) { |
302 | AacPsyContext *pctx; |
303 | float bark; |
304 | int i, j, g, start; |
305 | float prev, minscale, minath, minsnr, pe_min; |
306 | int chan_bitrate = ctx->avctx->bit_rate / ((ctx->avctx->flags & AV_CODEC_FLAG_QSCALE) ? 2.0f : ctx->avctx->channels); |
307 | |
308 | const int bandwidth = ctx->cutoff ? ctx->cutoff : AAC_CUTOFF(ctx->avctx); |
309 | const float num_bark = calc_bark((float)bandwidth); |
310 | |
311 | ctx->model_priv_data = av_mallocz(sizeof(AacPsyContext)); |
312 | if (!ctx->model_priv_data) |
313 | return AVERROR(ENOMEM); |
314 | pctx = ctx->model_priv_data; |
315 | pctx->global_quality = (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120) * 0.01f; |
316 | |
317 | if (ctx->avctx->flags & AV_CODEC_FLAG_QSCALE) { |
318 | /* Use the target average bitrate to compute spread parameters */ |
319 | chan_bitrate = (int)(chan_bitrate / 120.0 * (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120)); |
320 | } |
321 | |
322 | pctx->chan_bitrate = chan_bitrate; |
323 | pctx->frame_bits = FFMIN(2560, chan_bitrate * AAC_BLOCK_SIZE_LONG / ctx->avctx->sample_rate); |
324 | pctx->pe.min = 8.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f); |
325 | pctx->pe.max = 12.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f); |
326 | ctx->bitres.size = 6144 - pctx->frame_bits; |
327 | ctx->bitres.size -= ctx->bitres.size % 8; |
328 | pctx->fill_level = ctx->bitres.size; |
329 | minath = ath(3410 - 0.733 * ATH_ADD, ATH_ADD); |
330 | for (j = 0; j < 2; j++) { |
331 | AacPsyCoeffs *coeffs = pctx->psy_coef[j]; |
332 | const uint8_t *band_sizes = ctx->bands[j]; |
333 | float line_to_frequency = ctx->avctx->sample_rate / (j ? 256.f : 2048.0f); |
334 | float avg_chan_bits = chan_bitrate * (j ? 128.0f : 1024.0f) / ctx->avctx->sample_rate; |
335 | /* reference encoder uses 2.4% here instead of 60% like the spec says */ |
336 | float bark_pe = 0.024f * PSY_3GPP_BITS_TO_PE(avg_chan_bits) / num_bark; |
337 | float en_spread_low = j ? PSY_3GPP_EN_SPREAD_LOW_S : PSY_3GPP_EN_SPREAD_LOW_L; |
338 | /* High energy spreading for long blocks <= 22kbps/channel and short blocks are the same. */ |
339 | float en_spread_hi = (j || (chan_bitrate <= 22.0f)) ? PSY_3GPP_EN_SPREAD_HI_S : PSY_3GPP_EN_SPREAD_HI_L1; |
340 | |
341 | i = 0; |
342 | prev = 0.0; |
343 | for (g = 0; g < ctx->num_bands[j]; g++) { |
344 | i += band_sizes[g]; |
345 | bark = calc_bark((i-1) * line_to_frequency); |
346 | coeffs[g].barks = (bark + prev) / 2.0; |
347 | prev = bark; |
348 | } |
349 | for (g = 0; g < ctx->num_bands[j] - 1; g++) { |
350 | AacPsyCoeffs *coeff = &coeffs[g]; |
351 | float bark_width = coeffs[g+1].barks - coeffs->barks; |
352 | coeff->spread_low[0] = ff_exp10(-bark_width * PSY_3GPP_THR_SPREAD_LOW); |
353 | coeff->spread_hi [0] = ff_exp10(-bark_width * PSY_3GPP_THR_SPREAD_HI); |
354 | coeff->spread_low[1] = ff_exp10(-bark_width * en_spread_low); |
355 | coeff->spread_hi [1] = ff_exp10(-bark_width * en_spread_hi); |
356 | pe_min = bark_pe * bark_width; |
357 | minsnr = exp2(pe_min / band_sizes[g]) - 1.5f; |
358 | coeff->min_snr = av_clipf(1.0f / minsnr, PSY_SNR_25DB, PSY_SNR_1DB); |
359 | } |
360 | start = 0; |
361 | for (g = 0; g < ctx->num_bands[j]; g++) { |
362 | minscale = ath(start * line_to_frequency, ATH_ADD); |
363 | for (i = 1; i < band_sizes[g]; i++) |
364 | minscale = FFMIN(minscale, ath((start + i) * line_to_frequency, ATH_ADD)); |
365 | coeffs[g].ath = minscale - minath; |
366 | start += band_sizes[g]; |
367 | } |
368 | } |
369 | |
370 | pctx->ch = av_mallocz_array(ctx->avctx->channels, sizeof(AacPsyChannel)); |
371 | if (!pctx->ch) { |
372 | av_freep(&ctx->model_priv_data); |
373 | return AVERROR(ENOMEM); |
374 | } |
375 | |
376 | lame_window_init(pctx, ctx->avctx); |
377 | |
378 | return 0; |
379 | } |
380 | |
381 | /** |
382 | * IIR filter used in block switching decision |
383 | */ |
384 | static float iir_filter(int in, float state[2]) |
385 | { |
386 | float ret; |
387 | |
388 | ret = 0.7548f * (in - state[0]) + 0.5095f * state[1]; |
389 | state[0] = in; |
390 | state[1] = ret; |
391 | return ret; |
392 | } |
393 | |
394 | /** |
395 | * window grouping information stored as bits (0 - new group, 1 - group continues) |
396 | */ |
397 | static const uint8_t window_grouping[9] = { |
398 | 0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36 |
399 | }; |
400 | |
401 | /** |
402 | * Tell encoder which window types to use. |
403 | * @see 3GPP TS26.403 5.4.1 "Blockswitching" |
404 | */ |
405 | static av_unused FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx, |
406 | const int16_t *audio, |
407 | const int16_t *la, |
408 | int channel, int prev_type) |
409 | { |
410 | int i, j; |
411 | int br = ((AacPsyContext*)ctx->model_priv_data)->chan_bitrate; |
412 | int attack_ratio = br <= 16000 ? 18 : 10; |
413 | AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data; |
414 | AacPsyChannel *pch = &pctx->ch[channel]; |
415 | uint8_t grouping = 0; |
416 | int next_type = pch->next_window_seq; |
417 | FFPsyWindowInfo wi = { { 0 } }; |
418 | |
419 | if (la) { |
420 | float s[8], v; |
421 | int switch_to_eight = 0; |
422 | float sum = 0.0, sum2 = 0.0; |
423 | int attack_n = 0; |
424 | int stay_short = 0; |
425 | for (i = 0; i < 8; i++) { |
426 | for (j = 0; j < 128; j++) { |
427 | v = iir_filter(la[i*128+j], pch->iir_state); |
428 | sum += v*v; |
429 | } |
430 | s[i] = sum; |
431 | sum2 += sum; |
432 | } |
433 | for (i = 0; i < 8; i++) { |
434 | if (s[i] > pch->win_energy * attack_ratio) { |
435 | attack_n = i + 1; |
436 | switch_to_eight = 1; |
437 | break; |
438 | } |
439 | } |
440 | pch->win_energy = pch->win_energy*7/8 + sum2/64; |
441 | |
442 | wi.window_type[1] = prev_type; |
443 | switch (prev_type) { |
444 | case ONLY_LONG_SEQUENCE: |
445 | wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE; |
446 | next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE; |
447 | break; |
448 | case LONG_START_SEQUENCE: |
449 | wi.window_type[0] = EIGHT_SHORT_SEQUENCE; |
450 | grouping = pch->next_grouping; |
451 | next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE; |
452 | break; |
453 | case LONG_STOP_SEQUENCE: |
454 | wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE; |
455 | next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE; |
456 | break; |
457 | case EIGHT_SHORT_SEQUENCE: |
458 | stay_short = next_type == EIGHT_SHORT_SEQUENCE || switch_to_eight; |
459 | wi.window_type[0] = stay_short ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE; |
460 | grouping = next_type == EIGHT_SHORT_SEQUENCE ? pch->next_grouping : 0; |
461 | next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE; |
462 | break; |
463 | } |
464 | |
465 | pch->next_grouping = window_grouping[attack_n]; |
466 | pch->next_window_seq = next_type; |
467 | } else { |
468 | for (i = 0; i < 3; i++) |
469 | wi.window_type[i] = prev_type; |
470 | grouping = (prev_type == EIGHT_SHORT_SEQUENCE) ? window_grouping[0] : 0; |
471 | } |
472 | |
473 | wi.window_shape = 1; |
474 | if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) { |
475 | wi.num_windows = 1; |
476 | wi.grouping[0] = 1; |
477 | } else { |
478 | int lastgrp = 0; |
479 | wi.num_windows = 8; |
480 | for (i = 0; i < 8; i++) { |
481 | if (!((grouping >> i) & 1)) |
482 | lastgrp = i; |
483 | wi.grouping[lastgrp]++; |
484 | } |
485 | } |
486 | |
487 | return wi; |
488 | } |
489 | |
490 | /* 5.6.1.2 "Calculation of Bit Demand" */ |
491 | static int calc_bit_demand(AacPsyContext *ctx, float pe, int bits, int size, |
492 | int short_window) |
493 | { |
494 | const float bitsave_slope = short_window ? PSY_3GPP_SAVE_SLOPE_S : PSY_3GPP_SAVE_SLOPE_L; |
495 | const float bitsave_add = short_window ? PSY_3GPP_SAVE_ADD_S : PSY_3GPP_SAVE_ADD_L; |
496 | const float bitspend_slope = short_window ? PSY_3GPP_SPEND_SLOPE_S : PSY_3GPP_SPEND_SLOPE_L; |
497 | const float bitspend_add = short_window ? PSY_3GPP_SPEND_ADD_S : PSY_3GPP_SPEND_ADD_L; |
498 | const float clip_low = short_window ? PSY_3GPP_CLIP_LO_S : PSY_3GPP_CLIP_LO_L; |
499 | const float clip_high = short_window ? PSY_3GPP_CLIP_HI_S : PSY_3GPP_CLIP_HI_L; |
500 | float clipped_pe, bit_save, bit_spend, bit_factor, fill_level, forgetful_min_pe; |
501 | |
502 | ctx->fill_level += ctx->frame_bits - bits; |
503 | ctx->fill_level = av_clip(ctx->fill_level, 0, size); |
504 | fill_level = av_clipf((float)ctx->fill_level / size, clip_low, clip_high); |
505 | clipped_pe = av_clipf(pe, ctx->pe.min, ctx->pe.max); |
506 | bit_save = (fill_level + bitsave_add) * bitsave_slope; |
507 | assert(bit_save <= 0.3f && bit_save >= -0.05000001f); |
508 | bit_spend = (fill_level + bitspend_add) * bitspend_slope; |
509 | assert(bit_spend <= 0.5f && bit_spend >= -0.1f); |
510 | /* The bit factor graph in the spec is obviously incorrect. |
511 | * bit_spend + ((bit_spend - bit_spend))... |
512 | * The reference encoder subtracts everything from 1, but also seems incorrect. |
513 | * 1 - bit_save + ((bit_spend + bit_save))... |
514 | * Hopefully below is correct. |
515 | */ |
516 | bit_factor = 1.0f - bit_save + ((bit_spend - bit_save) / (ctx->pe.max - ctx->pe.min)) * (clipped_pe - ctx->pe.min); |
517 | /* NOTE: The reference encoder attempts to center pe max/min around the current pe. |
518 | * Here we do that by slowly forgetting pe.min when pe stays in a range that makes |
519 | * it unlikely (ie: above the mean) |
520 | */ |
521 | ctx->pe.max = FFMAX(pe, ctx->pe.max); |
522 | forgetful_min_pe = ((ctx->pe.min * PSY_PE_FORGET_SLOPE) |
523 | + FFMAX(ctx->pe.min, pe * (pe / ctx->pe.max))) / (PSY_PE_FORGET_SLOPE + 1); |
524 | ctx->pe.min = FFMIN(pe, forgetful_min_pe); |
525 | |
526 | /* NOTE: allocate a minimum of 1/8th average frame bits, to avoid |
527 | * reservoir starvation from producing zero-bit frames |
528 | */ |
529 | return FFMIN( |
530 | ctx->frame_bits * bit_factor, |
531 | FFMAX(ctx->frame_bits + size - bits, ctx->frame_bits / 8)); |
532 | } |
533 | |
534 | static float calc_pe_3gpp(AacPsyBand *band) |
535 | { |
536 | float pe, a; |
537 | |
538 | band->pe = 0.0f; |
539 | band->pe_const = 0.0f; |
540 | band->active_lines = 0.0f; |
541 | if (band->energy > band->thr) { |
542 | a = log2f(band->energy); |
543 | pe = a - log2f(band->thr); |
544 | band->active_lines = band->nz_lines; |
545 | if (pe < PSY_3GPP_C1) { |
546 | pe = pe * PSY_3GPP_C3 + PSY_3GPP_C2; |
547 | a = a * PSY_3GPP_C3 + PSY_3GPP_C2; |
548 | band->active_lines *= PSY_3GPP_C3; |
549 | } |
550 | band->pe = pe * band->nz_lines; |
551 | band->pe_const = a * band->nz_lines; |
552 | } |
553 | |
554 | return band->pe; |
555 | } |
556 | |
557 | static float calc_reduction_3gpp(float a, float desired_pe, float pe, |
558 | float active_lines) |
559 | { |
560 | float thr_avg, reduction; |
561 | |
562 | if(active_lines == 0.0) |
563 | return 0; |
564 | |
565 | thr_avg = exp2f((a - pe) / (4.0f * active_lines)); |
566 | reduction = exp2f((a - desired_pe) / (4.0f * active_lines)) - thr_avg; |
567 | |
568 | return FFMAX(reduction, 0.0f); |
569 | } |
570 | |
571 | static float calc_reduced_thr_3gpp(AacPsyBand *band, float min_snr, |
572 | float reduction) |
573 | { |
574 | float thr = band->thr; |
575 | |
576 | if (band->energy > thr) { |
577 | thr = sqrtf(thr); |
578 | thr = sqrtf(thr) + reduction; |
579 | thr *= thr; |
580 | thr *= thr; |
581 | |
582 | /* This deviates from the 3GPP spec to match the reference encoder. |
583 | * It performs min(thr_reduced, max(thr, energy/min_snr)) only for bands |
584 | * that have hole avoidance on (active or inactive). It always reduces the |
585 | * threshold of bands with hole avoidance off. |
586 | */ |
587 | if (thr > band->energy * min_snr && band->avoid_holes != PSY_3GPP_AH_NONE) { |
588 | thr = FFMAX(band->thr, band->energy * min_snr); |
589 | band->avoid_holes = PSY_3GPP_AH_ACTIVE; |
590 | } |
591 | } |
592 | |
593 | return thr; |
594 | } |
595 | |
596 | #ifndef calc_thr_3gpp |
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 | { |
600 | int i, w, g; |
601 | int start = 0, wstart = 0; |
602 | for (w = 0; w < wi->num_windows*16; w += 16) { |
603 | wstart = 0; |
604 | for (g = 0; g < num_bands; g++) { |
605 | AacPsyBand *band = &pch->band[w+g]; |
606 | |
607 | float form_factor = 0.0f; |
608 | float Temp; |
609 | band->energy = 0.0f; |
610 | if (wstart < cutoff) { |
611 | for (i = 0; i < band_sizes[g]; i++) { |
612 | band->energy += coefs[start+i] * coefs[start+i]; |
613 | form_factor += sqrtf(fabs(coefs[start+i])); |
614 | } |
615 | } |
616 | Temp = band->energy > 0 ? sqrtf((float)band_sizes[g] / band->energy) : 0; |
617 | band->thr = band->energy * 0.001258925f; |
618 | band->nz_lines = form_factor * sqrtf(Temp); |
619 | |
620 | start += band_sizes[g]; |
621 | wstart += band_sizes[g]; |
622 | } |
623 | } |
624 | } |
625 | #endif /* calc_thr_3gpp */ |
626 | |
627 | #ifndef psy_hp_filter |
628 | static void psy_hp_filter(const float *firbuf, float *hpfsmpl, const float *psy_fir_coeffs) |
629 | { |
630 | int i, j; |
631 | for (i = 0; i < AAC_BLOCK_SIZE_LONG; i++) { |
632 | float sum1, sum2; |
633 | sum1 = firbuf[i + (PSY_LAME_FIR_LEN - 1) / 2]; |
634 | sum2 = 0.0; |
635 | for (j = 0; j < ((PSY_LAME_FIR_LEN - 1) / 2) - 1; j += 2) { |
636 | sum1 += psy_fir_coeffs[j] * (firbuf[i + j] + firbuf[i + PSY_LAME_FIR_LEN - j]); |
637 | sum2 += psy_fir_coeffs[j + 1] * (firbuf[i + j + 1] + firbuf[i + PSY_LAME_FIR_LEN - j - 1]); |
638 | } |
639 | /* NOTE: The LAME psymodel expects it's input in the range -32768 to 32768. |
640 | * Tuning this for normalized floats would be difficult. */ |
641 | hpfsmpl[i] = (sum1 + sum2) * 32768.0f; |
642 | } |
643 | } |
644 | #endif /* psy_hp_filter */ |
645 | |
646 | /** |
647 | * Calculate band thresholds as suggested in 3GPP TS26.403 |
648 | */ |
649 | static void psy_3gpp_analyze_channel(FFPsyContext *ctx, int channel, |
650 | const float *coefs, const FFPsyWindowInfo *wi) |
651 | { |
652 | AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data; |
653 | AacPsyChannel *pch = &pctx->ch[channel]; |
654 | int i, w, g; |
655 | float desired_bits, desired_pe, delta_pe, reduction= NAN, spread_en[128] = {0}; |
656 | float a = 0.0f, active_lines = 0.0f, norm_fac = 0.0f; |
657 | float pe = pctx->chan_bitrate > 32000 ? 0.0f : FFMAX(50.0f, 100.0f - pctx->chan_bitrate * 100.0f / 32000.0f); |
658 | const int num_bands = ctx->num_bands[wi->num_windows == 8]; |
659 | const uint8_t *band_sizes = ctx->bands[wi->num_windows == 8]; |
660 | AacPsyCoeffs *coeffs = pctx->psy_coef[wi->num_windows == 8]; |
661 | const float avoid_hole_thr = wi->num_windows == 8 ? PSY_3GPP_AH_THR_SHORT : PSY_3GPP_AH_THR_LONG; |
662 | const int bandwidth = ctx->cutoff ? ctx->cutoff : AAC_CUTOFF(ctx->avctx); |
663 | const int cutoff = bandwidth * 2048 / wi->num_windows / ctx->avctx->sample_rate; |
664 | |
665 | //calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation" |
666 | calc_thr_3gpp(wi, num_bands, pch, band_sizes, coefs, cutoff); |
667 | |
668 | //modify thresholds and energies - spread, threshold in quiet, pre-echo control |
669 | for (w = 0; w < wi->num_windows*16; w += 16) { |
670 | AacPsyBand *bands = &pch->band[w]; |
671 | |
672 | /* 5.4.2.3 "Spreading" & 5.4.3 "Spread Energy Calculation" */ |
673 | spread_en[0] = bands[0].energy; |
674 | for (g = 1; g < num_bands; g++) { |
675 | bands[g].thr = FFMAX(bands[g].thr, bands[g-1].thr * coeffs[g].spread_hi[0]); |
676 | spread_en[w+g] = FFMAX(bands[g].energy, spread_en[w+g-1] * coeffs[g].spread_hi[1]); |
677 | } |
678 | for (g = num_bands - 2; g >= 0; g--) { |
679 | bands[g].thr = FFMAX(bands[g].thr, bands[g+1].thr * coeffs[g].spread_low[0]); |
680 | spread_en[w+g] = FFMAX(spread_en[w+g], spread_en[w+g+1] * coeffs[g].spread_low[1]); |
681 | } |
682 | //5.4.2.4 "Threshold in quiet" |
683 | for (g = 0; g < num_bands; g++) { |
684 | AacPsyBand *band = &bands[g]; |
685 | |
686 | band->thr_quiet = band->thr = FFMAX(band->thr, coeffs[g].ath); |
687 | //5.4.2.5 "Pre-echo control" |
688 | if (!(wi->window_type[0] == LONG_STOP_SEQUENCE || (!w && wi->window_type[1] == LONG_START_SEQUENCE))) |
689 | band->thr = FFMAX(PSY_3GPP_RPEMIN*band->thr, FFMIN(band->thr, |
690 | PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet)); |
691 | |
692 | /* 5.6.1.3.1 "Preparatory steps of the perceptual entropy calculation" */ |
693 | pe += calc_pe_3gpp(band); |
694 | a += band->pe_const; |
695 | active_lines += band->active_lines; |
696 | |
697 | /* 5.6.1.3.3 "Selection of the bands for avoidance of holes" */ |
698 | if (spread_en[w+g] * avoid_hole_thr > band->energy || coeffs[g].min_snr > 1.0f) |
699 | band->avoid_holes = PSY_3GPP_AH_NONE; |
700 | else |
701 | band->avoid_holes = PSY_3GPP_AH_INACTIVE; |
702 | } |
703 | } |
704 | |
705 | /* 5.6.1.3.2 "Calculation of the desired perceptual entropy" */ |
706 | ctx->ch[channel].entropy = pe; |
707 | if (ctx->avctx->flags & AV_CODEC_FLAG_QSCALE) { |
708 | /* (2.5 * 120) achieves almost transparent rate, and we want to give |
709 | * ample room downwards, so we make that equivalent to QSCALE=2.4 |
710 | */ |
711 | desired_pe = pe * (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120) / (2 * 2.5f * 120.0f); |
712 | desired_bits = FFMIN(2560, PSY_3GPP_PE_TO_BITS(desired_pe)); |
713 | desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits); // reflect clipping |
714 | |
715 | /* PE slope smoothing */ |
716 | if (ctx->bitres.bits > 0) { |
717 | desired_bits = FFMIN(2560, PSY_3GPP_PE_TO_BITS(desired_pe)); |
718 | desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits); // reflect clipping |
719 | } |
720 | |
721 | pctx->pe.max = FFMAX(pe, pctx->pe.max); |
722 | pctx->pe.min = FFMIN(pe, pctx->pe.min); |
723 | } else { |
724 | desired_bits = calc_bit_demand(pctx, pe, ctx->bitres.bits, ctx->bitres.size, wi->num_windows == 8); |
725 | desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits); |
726 | |
727 | /* NOTE: PE correction is kept simple. During initial testing it had very |
728 | * little effect on the final bitrate. Probably a good idea to come |
729 | * back and do more testing later. |
730 | */ |
731 | if (ctx->bitres.bits > 0) |
732 | desired_pe *= av_clipf(pctx->pe.previous / PSY_3GPP_BITS_TO_PE(ctx->bitres.bits), |
733 | 0.85f, 1.15f); |
734 | } |
735 | pctx->pe.previous = PSY_3GPP_BITS_TO_PE(desired_bits); |
736 | ctx->bitres.alloc = desired_bits; |
737 | |
738 | if (desired_pe < pe) { |
739 | /* 5.6.1.3.4 "First Estimation of the reduction value" */ |
740 | for (w = 0; w < wi->num_windows*16; w += 16) { |
741 | reduction = calc_reduction_3gpp(a, desired_pe, pe, active_lines); |
742 | pe = 0.0f; |
743 | a = 0.0f; |
744 | active_lines = 0.0f; |
745 | for (g = 0; g < num_bands; g++) { |
746 | AacPsyBand *band = &pch->band[w+g]; |
747 | |
748 | band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction); |
749 | /* recalculate PE */ |
750 | pe += calc_pe_3gpp(band); |
751 | a += band->pe_const; |
752 | active_lines += band->active_lines; |
753 | } |
754 | } |
755 | |
756 | /* 5.6.1.3.5 "Second Estimation of the reduction value" */ |
757 | for (i = 0; i < 2; i++) { |
758 | float pe_no_ah = 0.0f, desired_pe_no_ah; |
759 | active_lines = a = 0.0f; |
760 | for (w = 0; w < wi->num_windows*16; w += 16) { |
761 | for (g = 0; g < num_bands; g++) { |
762 | AacPsyBand *band = &pch->band[w+g]; |
763 | |
764 | if (band->avoid_holes != PSY_3GPP_AH_ACTIVE) { |
765 | pe_no_ah += band->pe; |
766 | a += band->pe_const; |
767 | active_lines += band->active_lines; |
768 | } |
769 | } |
770 | } |
771 | desired_pe_no_ah = FFMAX(desired_pe - (pe - pe_no_ah), 0.0f); |
772 | if (active_lines > 0.0f) |
773 | reduction = calc_reduction_3gpp(a, desired_pe_no_ah, pe_no_ah, active_lines); |
774 | |
775 | pe = 0.0f; |
776 | for (w = 0; w < wi->num_windows*16; w += 16) { |
777 | for (g = 0; g < num_bands; g++) { |
778 | AacPsyBand *band = &pch->band[w+g]; |
779 | |
780 | if (active_lines > 0.0f) |
781 | band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction); |
782 | pe += calc_pe_3gpp(band); |
783 | if (band->thr > 0.0f) |
784 | band->norm_fac = band->active_lines / band->thr; |
785 | else |
786 | band->norm_fac = 0.0f; |
787 | norm_fac += band->norm_fac; |
788 | } |
789 | } |
790 | delta_pe = desired_pe - pe; |
791 | if (fabs(delta_pe) > 0.05f * desired_pe) |
792 | break; |
793 | } |
794 | |
795 | if (pe < 1.15f * desired_pe) { |
796 | /* 6.6.1.3.6 "Final threshold modification by linearization" */ |
797 | norm_fac = 1.0f / norm_fac; |
798 | for (w = 0; w < wi->num_windows*16; w += 16) { |
799 | for (g = 0; g < num_bands; g++) { |
800 | AacPsyBand *band = &pch->band[w+g]; |
801 | |
802 | if (band->active_lines > 0.5f) { |
803 | float delta_sfb_pe = band->norm_fac * norm_fac * delta_pe; |
804 | float thr = band->thr; |
805 | |
806 | thr *= exp2f(delta_sfb_pe / band->active_lines); |
807 | if (thr > coeffs[g].min_snr * band->energy && band->avoid_holes == PSY_3GPP_AH_INACTIVE) |
808 | thr = FFMAX(band->thr, coeffs[g].min_snr * band->energy); |
809 | band->thr = thr; |
810 | } |
811 | } |
812 | } |
813 | } else { |
814 | /* 5.6.1.3.7 "Further perceptual entropy reduction" */ |
815 | g = num_bands; |
816 | while (pe > desired_pe && g--) { |
817 | for (w = 0; w < wi->num_windows*16; w+= 16) { |
818 | AacPsyBand *band = &pch->band[w+g]; |
819 | if (band->avoid_holes != PSY_3GPP_AH_NONE && coeffs[g].min_snr < PSY_SNR_1DB) { |
820 | coeffs[g].min_snr = PSY_SNR_1DB; |
821 | band->thr = band->energy * PSY_SNR_1DB; |
822 | pe += band->active_lines * 1.5f - band->pe; |
823 | } |
824 | } |
825 | } |
826 | /* TODO: allow more holes (unused without mid/side) */ |
827 | } |
828 | } |
829 | |
830 | for (w = 0; w < wi->num_windows*16; w += 16) { |
831 | for (g = 0; g < num_bands; g++) { |
832 | AacPsyBand *band = &pch->band[w+g]; |
833 | FFPsyBand *psy_band = &ctx->ch[channel].psy_bands[w+g]; |
834 | |
835 | psy_band->threshold = band->thr; |
836 | psy_band->energy = band->energy; |
837 | psy_band->spread = band->active_lines * 2.0f / band_sizes[g]; |
838 | psy_band->bits = PSY_3GPP_PE_TO_BITS(band->pe); |
839 | } |
840 | } |
841 | |
842 | memcpy(pch->prev_band, pch->band, sizeof(pch->band)); |
843 | } |
844 | |
845 | static void psy_3gpp_analyze(FFPsyContext *ctx, int channel, |
846 | const float **coeffs, const FFPsyWindowInfo *wi) |
847 | { |
848 | int ch; |
849 | FFPsyChannelGroup *group = ff_psy_find_group(ctx, channel); |
850 | |
851 | for (ch = 0; ch < group->num_ch; ch++) |
852 | psy_3gpp_analyze_channel(ctx, channel + ch, coeffs[ch], &wi[ch]); |
853 | } |
854 | |
855 | static av_cold void psy_3gpp_end(FFPsyContext *apc) |
856 | { |
857 | AacPsyContext *pctx = (AacPsyContext*) apc->model_priv_data; |
858 | av_freep(&pctx->ch); |
859 | av_freep(&apc->model_priv_data); |
860 | } |
861 | |
862 | static void lame_apply_block_type(AacPsyChannel *ctx, FFPsyWindowInfo *wi, int uselongblock) |
863 | { |
864 | int blocktype = ONLY_LONG_SEQUENCE; |
865 | if (uselongblock) { |
866 | if (ctx->next_window_seq == EIGHT_SHORT_SEQUENCE) |
867 | blocktype = LONG_STOP_SEQUENCE; |
868 | } else { |
869 | blocktype = EIGHT_SHORT_SEQUENCE; |
870 | if (ctx->next_window_seq == ONLY_LONG_SEQUENCE) |
871 | ctx->next_window_seq = LONG_START_SEQUENCE; |
872 | if (ctx->next_window_seq == LONG_STOP_SEQUENCE) |
873 | ctx->next_window_seq = EIGHT_SHORT_SEQUENCE; |
874 | } |
875 | |
876 | wi->window_type[0] = ctx->next_window_seq; |
877 | ctx->next_window_seq = blocktype; |
878 | } |
879 | |
880 | static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx, const float *audio, |
881 | const float *la, int channel, int prev_type) |
882 | { |
883 | AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data; |
884 | AacPsyChannel *pch = &pctx->ch[channel]; |
885 | int grouping = 0; |
886 | int uselongblock = 1; |
887 | int attacks[AAC_NUM_BLOCKS_SHORT + 1] = { 0 }; |
888 | int i; |
889 | FFPsyWindowInfo wi = { { 0 } }; |
890 | |
891 | if (la) { |
892 | float hpfsmpl[AAC_BLOCK_SIZE_LONG]; |
893 | const float *pf = hpfsmpl; |
894 | float attack_intensity[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS]; |
895 | float energy_subshort[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS]; |
896 | float energy_short[AAC_NUM_BLOCKS_SHORT + 1] = { 0 }; |
897 | const float *firbuf = la + (AAC_BLOCK_SIZE_SHORT/4 - PSY_LAME_FIR_LEN); |
898 | int att_sum = 0; |
899 | |
900 | /* LAME comment: apply high pass filter of fs/4 */ |
901 | psy_hp_filter(firbuf, hpfsmpl, psy_fir_coeffs); |
902 | |
903 | /* Calculate the energies of each sub-shortblock */ |
904 | for (i = 0; i < PSY_LAME_NUM_SUBBLOCKS; i++) { |
905 | energy_subshort[i] = pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) * PSY_LAME_NUM_SUBBLOCKS)]; |
906 | assert(pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)] > 0); |
907 | attack_intensity[i] = energy_subshort[i] / pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)]; |
908 | energy_short[0] += energy_subshort[i]; |
909 | } |
910 | |
911 | for (i = 0; i < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; i++) { |
912 | const float *const pfe = pf + AAC_BLOCK_SIZE_LONG / (AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS); |
913 | float p = 1.0f; |
914 | for (; pf < pfe; pf++) |
915 | p = FFMAX(p, fabsf(*pf)); |
916 | pch->prev_energy_subshort[i] = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS] = p; |
917 | energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p; |
918 | /* NOTE: The indexes below are [i + 3 - 2] in the LAME source. |
919 | * Obviously the 3 and 2 have some significance, or this would be just [i + 1] |
920 | * (which is what we use here). What the 3 stands for is ambiguous, as it is both |
921 | * number of short blocks, and the number of sub-short blocks. |
922 | * It seems that LAME is comparing each sub-block to sub-block + 1 in the |
923 | * previous block. |
924 | */ |
925 | if (p > energy_subshort[i + 1]) |
926 | p = p / energy_subshort[i + 1]; |
927 | else if (energy_subshort[i + 1] > p * 10.0f) |
928 | p = energy_subshort[i + 1] / (p * 10.0f); |
929 | else |
930 | p = 0.0; |
931 | attack_intensity[i + PSY_LAME_NUM_SUBBLOCKS] = p; |
932 | } |
933 | |
934 | /* compare energy between sub-short blocks */ |
935 | for (i = 0; i < (AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS; i++) |
936 | if (!attacks[i / PSY_LAME_NUM_SUBBLOCKS]) |
937 | if (attack_intensity[i] > pch->attack_threshold) |
938 | attacks[i / PSY_LAME_NUM_SUBBLOCKS] = (i % PSY_LAME_NUM_SUBBLOCKS) + 1; |
939 | |
940 | /* should have energy change between short blocks, in order to avoid periodic signals */ |
941 | /* Good samples to show the effect are Trumpet test songs */ |
942 | /* GB: tuned (1) to avoid too many short blocks for test sample TRUMPET */ |
943 | /* RH: tuned (2) to let enough short blocks through for test sample FSOL and SNAPS */ |
944 | for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) { |
945 | const float u = energy_short[i - 1]; |
946 | const float v = energy_short[i]; |
947 | const float m = FFMAX(u, v); |
948 | if (m < 40000) { /* (2) */ |
949 | if (u < 1.7f * v && v < 1.7f * u) { /* (1) */ |
950 | if (i == 1 && attacks[0] < attacks[i]) |
951 | attacks[0] = 0; |
952 | attacks[i] = 0; |
953 | } |
954 | } |
955 | att_sum += attacks[i]; |
956 | } |
957 | |
958 | if (attacks[0] <= pch->prev_attack) |
959 | attacks[0] = 0; |
960 | |
961 | att_sum += attacks[0]; |
962 | /* 3 below indicates the previous attack happened in the last sub-block of the previous sequence */ |
963 | if (pch->prev_attack == 3 || att_sum) { |
964 | uselongblock = 0; |
965 | |
966 | for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) |
967 | if (attacks[i] && attacks[i-1]) |
968 | attacks[i] = 0; |
969 | } |
970 | } else { |
971 | /* We have no lookahead info, so just use same type as the previous sequence. */ |
972 | uselongblock = !(prev_type == EIGHT_SHORT_SEQUENCE); |
973 | } |
974 | |
975 | lame_apply_block_type(pch, &wi, uselongblock); |
976 | |
977 | wi.window_type[1] = prev_type; |
978 | if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) { |
979 | |
980 | wi.num_windows = 1; |
981 | wi.grouping[0] = 1; |
982 | if (wi.window_type[0] == LONG_START_SEQUENCE) |
983 | wi.window_shape = 0; |
984 | else |
985 | wi.window_shape = 1; |
986 | |
987 | } else { |
988 | int lastgrp = 0; |
989 | |
990 | wi.num_windows = 8; |
991 | wi.window_shape = 0; |
992 | for (i = 0; i < 8; i++) { |
993 | if (!((pch->next_grouping >> i) & 1)) |
994 | lastgrp = i; |
995 | wi.grouping[lastgrp]++; |
996 | } |
997 | } |
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 | for (i = 0; i < 9; i++) { |
1006 | if (attacks[i]) { |
1007 | grouping = i; |
1008 | break; |
1009 | } |
1010 | } |
1011 | pch->next_grouping = window_grouping[grouping]; |
1012 | |
1013 | pch->prev_attack = attacks[8]; |
1014 | |
1015 | return wi; |
1016 | } |
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 | }; |
1026 |