FFmpeg
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
opus_pvq.c
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2007-2008 CSIRO
3  * Copyright (c) 2007-2009 Xiph.Org Foundation
4  * Copyright (c) 2008-2009 Gregory Maxwell
5  * Copyright (c) 2012 Andrew D'Addesio
6  * Copyright (c) 2013-2014 Mozilla Corporation
7  * Copyright (c) 2017 Rostislav Pehlivanov <atomnuker@gmail.com>
8  *
9  * This file is part of FFmpeg.
10  *
11  * FFmpeg is free software; you can redistribute it and/or
12  * modify it under the terms of the GNU Lesser General Public
13  * License as published by the Free Software Foundation; either
14  * version 2.1 of the License, or (at your option) any later version.
15  *
16  * FFmpeg is distributed in the hope that it will be useful,
17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19  * Lesser General Public License for more details.
20  *
21  * You should have received a copy of the GNU Lesser General Public
22  * License along with FFmpeg; if not, write to the Free Software
23  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24  */
25 
26 #include "opustab.h"
27 #include "opus_pvq.h"
28 
29 #define CELT_PVQ_U(n, k) (ff_celt_pvq_u_row[FFMIN(n, k)][FFMAX(n, k)])
30 #define CELT_PVQ_V(n, k) (CELT_PVQ_U(n, k) + CELT_PVQ_U(n, (k) + 1))
31 
32 static inline int16_t celt_cos(int16_t x)
33 {
34  x = (MUL16(x, x) + 4096) >> 13;
35  x = (32767-x) + ROUND_MUL16(x, (-7651 + ROUND_MUL16(x, (8277 + ROUND_MUL16(-626, x)))));
36  return x + 1;
37 }
38 
39 static inline int celt_log2tan(int isin, int icos)
40 {
41  int lc, ls;
42  lc = opus_ilog(icos);
43  ls = opus_ilog(isin);
44  icos <<= 15 - lc;
45  isin <<= 15 - ls;
46  return (ls << 11) - (lc << 11) +
47  ROUND_MUL16(isin, ROUND_MUL16(isin, -2597) + 7932) -
48  ROUND_MUL16(icos, ROUND_MUL16(icos, -2597) + 7932);
49 }
50 
51 static inline int celt_bits2pulses(const uint8_t *cache, int bits)
52 {
53  // TODO: Find the size of cache and make it into an array in the parameters list
54  int i, low = 0, high;
55 
56  high = cache[0];
57  bits--;
58 
59  for (i = 0; i < 6; i++) {
60  int center = (low + high + 1) >> 1;
61  if (cache[center] >= bits)
62  high = center;
63  else
64  low = center;
65  }
66 
67  return (bits - (low == 0 ? -1 : cache[low]) <= cache[high] - bits) ? low : high;
68 }
69 
70 static inline int celt_pulses2bits(const uint8_t *cache, int pulses)
71 {
72  // TODO: Find the size of cache and make it into an array in the parameters list
73  return (pulses == 0) ? 0 : cache[pulses] + 1;
74 }
75 
76 static inline void celt_normalize_residual(const int * av_restrict iy, float * av_restrict X,
77  int N, float g)
78 {
79  int i;
80  for (i = 0; i < N; i++)
81  X[i] = g * iy[i];
82 }
83 
84 static void celt_exp_rotation_impl(float *X, uint32_t len, uint32_t stride,
85  float c, float s)
86 {
87  float *Xptr;
88  int i;
89 
90  Xptr = X;
91  for (i = 0; i < len - stride; i++) {
92  float x1 = Xptr[0];
93  float x2 = Xptr[stride];
94  Xptr[stride] = c * x2 + s * x1;
95  *Xptr++ = c * x1 - s * x2;
96  }
97 
98  Xptr = &X[len - 2 * stride - 1];
99  for (i = len - 2 * stride - 1; i >= 0; i--) {
100  float x1 = Xptr[0];
101  float x2 = Xptr[stride];
102  Xptr[stride] = c * x2 + s * x1;
103  *Xptr-- = c * x1 - s * x2;
104  }
105 }
106 
107 static inline void celt_exp_rotation(float *X, uint32_t len,
108  uint32_t stride, uint32_t K,
109  enum CeltSpread spread, const int encode)
110 {
111  uint32_t stride2 = 0;
112  float c, s;
113  float gain, theta;
114  int i;
115 
116  if (2*K >= len || spread == CELT_SPREAD_NONE)
117  return;
118 
119  gain = (float)len / (len + (20 - 5*spread) * K);
120  theta = M_PI * gain * gain / 4;
121 
122  c = cosf(theta);
123  s = sinf(theta);
124 
125  if (len >= stride << 3) {
126  stride2 = 1;
127  /* This is just a simple (equivalent) way of computing sqrt(len/stride) with rounding.
128  It's basically incrementing long as (stride2+0.5)^2 < len/stride. */
129  while ((stride2 * stride2 + stride2) * stride + (stride >> 2) < len)
130  stride2++;
131  }
132 
133  len /= stride;
134  for (i = 0; i < stride; i++) {
135  if (encode) {
136  celt_exp_rotation_impl(X + i * len, len, 1, c, -s);
137  if (stride2)
138  celt_exp_rotation_impl(X + i * len, len, stride2, s, -c);
139  } else {
140  if (stride2)
141  celt_exp_rotation_impl(X + i * len, len, stride2, s, c);
142  celt_exp_rotation_impl(X + i * len, len, 1, c, s);
143  }
144  }
145 }
146 
147 static inline uint32_t celt_extract_collapse_mask(const int *iy, uint32_t N, uint32_t B)
148 {
149  int i, j, N0 = N / B;
150  uint32_t collapse_mask = 0;
151 
152  if (B <= 1)
153  return 1;
154 
155  for (i = 0; i < B; i++)
156  for (j = 0; j < N0; j++)
157  collapse_mask |= (!!iy[i*N0+j]) << i;
158  return collapse_mask;
159 }
160 
161 static inline void celt_stereo_merge(float *X, float *Y, float mid, int N)
162 {
163  int i;
164  float xp = 0, side = 0;
165  float E[2];
166  float mid2;
167  float gain[2];
168 
169  /* Compute the norm of X+Y and X-Y as |X|^2 + |Y|^2 +/- sum(xy) */
170  for (i = 0; i < N; i++) {
171  xp += X[i] * Y[i];
172  side += Y[i] * Y[i];
173  }
174 
175  /* Compensating for the mid normalization */
176  xp *= mid;
177  mid2 = mid;
178  E[0] = mid2 * mid2 + side - 2 * xp;
179  E[1] = mid2 * mid2 + side + 2 * xp;
180  if (E[0] < 6e-4f || E[1] < 6e-4f) {
181  for (i = 0; i < N; i++)
182  Y[i] = X[i];
183  return;
184  }
185 
186  gain[0] = 1.0f / sqrtf(E[0]);
187  gain[1] = 1.0f / sqrtf(E[1]);
188 
189  for (i = 0; i < N; i++) {
190  float value[2];
191  /* Apply mid scaling (side is already scaled) */
192  value[0] = mid * X[i];
193  value[1] = Y[i];
194  X[i] = gain[0] * (value[0] - value[1]);
195  Y[i] = gain[1] * (value[0] + value[1]);
196  }
197 }
198 
199 static void celt_interleave_hadamard(float *tmp, float *X, int N0,
200  int stride, int hadamard)
201 {
202  int i, j, N = N0*stride;
203  const uint8_t *order = &ff_celt_hadamard_order[hadamard ? stride - 2 : 30];
204 
205  for (i = 0; i < stride; i++)
206  for (j = 0; j < N0; j++)
207  tmp[j*stride+i] = X[order[i]*N0+j];
208 
209  memcpy(X, tmp, N*sizeof(float));
210 }
211 
212 static void celt_deinterleave_hadamard(float *tmp, float *X, int N0,
213  int stride, int hadamard)
214 {
215  int i, j, N = N0*stride;
216  const uint8_t *order = &ff_celt_hadamard_order[hadamard ? stride - 2 : 30];
217 
218  for (i = 0; i < stride; i++)
219  for (j = 0; j < N0; j++)
220  tmp[order[i]*N0+j] = X[j*stride+i];
221 
222  memcpy(X, tmp, N*sizeof(float));
223 }
224 
225 static void celt_haar1(float *X, int N0, int stride)
226 {
227  int i, j;
228  N0 >>= 1;
229  for (i = 0; i < stride; i++) {
230  for (j = 0; j < N0; j++) {
231  float x0 = X[stride * (2 * j + 0) + i];
232  float x1 = X[stride * (2 * j + 1) + i];
233  X[stride * (2 * j + 0) + i] = (x0 + x1) * M_SQRT1_2;
234  X[stride * (2 * j + 1) + i] = (x0 - x1) * M_SQRT1_2;
235  }
236  }
237 }
238 
239 static inline int celt_compute_qn(int N, int b, int offset, int pulse_cap,
240  int stereo)
241 {
242  int qn, qb;
243  int N2 = 2 * N - 1;
244  if (stereo && N == 2)
245  N2--;
246 
247  /* The upper limit ensures that in a stereo split with itheta==16384, we'll
248  * always have enough bits left over to code at least one pulse in the
249  * side; otherwise it would collapse, since it doesn't get folded. */
250  qb = FFMIN3(b - pulse_cap - (4 << 3), (b + N2 * offset) / N2, 8 << 3);
251  qn = (qb < (1 << 3 >> 1)) ? 1 : ((ff_celt_qn_exp2[qb & 0x7] >> (14 - (qb >> 3))) + 1) >> 1 << 1;
252  return qn;
253 }
254 
255 /* Convert the quantized vector to an index */
256 static inline uint32_t celt_icwrsi(uint32_t N, uint32_t K, const int *y)
257 {
258  int i, idx = 0, sum = 0;
259  for (i = N - 1; i >= 0; i--) {
260  const uint32_t i_s = CELT_PVQ_U(N - i, sum + FFABS(y[i]) + 1);
261  idx += CELT_PVQ_U(N - i, sum) + (y[i] < 0)*i_s;
262  sum += FFABS(y[i]);
263  }
264  return idx;
265 }
266 
267 // this code was adapted from libopus
268 static inline uint64_t celt_cwrsi(uint32_t N, uint32_t K, uint32_t i, int *y)
269 {
270  uint64_t norm = 0;
271  uint32_t q, p;
272  int s, val;
273  int k0;
274 
275  while (N > 2) {
276  /*Lots of pulses case:*/
277  if (K >= N) {
278  const uint32_t *row = ff_celt_pvq_u_row[N];
279 
280  /* Are the pulses in this dimension negative? */
281  p = row[K + 1];
282  s = -(i >= p);
283  i -= p & s;
284 
285  /*Count how many pulses were placed in this dimension.*/
286  k0 = K;
287  q = row[N];
288  if (q > i) {
289  K = N;
290  do {
291  p = ff_celt_pvq_u_row[--K][N];
292  } while (p > i);
293  } else
294  for (p = row[K]; p > i; p = row[K])
295  K--;
296 
297  i -= p;
298  val = (k0 - K + s) ^ s;
299  norm += val * val;
300  *y++ = val;
301  } else { /*Lots of dimensions case:*/
302  /*Are there any pulses in this dimension at all?*/
303  p = ff_celt_pvq_u_row[K ][N];
304  q = ff_celt_pvq_u_row[K + 1][N];
305 
306  if (p <= i && i < q) {
307  i -= p;
308  *y++ = 0;
309  } else {
310  /*Are the pulses in this dimension negative?*/
311  s = -(i >= q);
312  i -= q & s;
313 
314  /*Count how many pulses were placed in this dimension.*/
315  k0 = K;
316  do p = ff_celt_pvq_u_row[--K][N];
317  while (p > i);
318 
319  i -= p;
320  val = (k0 - K + s) ^ s;
321  norm += val * val;
322  *y++ = val;
323  }
324  }
325  N--;
326  }
327 
328  /* N == 2 */
329  p = 2 * K + 1;
330  s = -(i >= p);
331  i -= p & s;
332  k0 = K;
333  K = (i + 1) / 2;
334 
335  if (K)
336  i -= 2 * K - 1;
337 
338  val = (k0 - K + s) ^ s;
339  norm += val * val;
340  *y++ = val;
341 
342  /* N==1 */
343  s = -i;
344  val = (K + s) ^ s;
345  norm += val * val;
346  *y = val;
347 
348  return norm;
349 }
350 
351 static inline void celt_encode_pulses(OpusRangeCoder *rc, int *y, uint32_t N, uint32_t K)
352 {
353  ff_opus_rc_enc_uint(rc, celt_icwrsi(N, K, y), CELT_PVQ_V(N, K));
354 }
355 
356 static inline float celt_decode_pulses(OpusRangeCoder *rc, int *y, uint32_t N, uint32_t K)
357 {
358  const uint32_t idx = ff_opus_rc_dec_uint(rc, CELT_PVQ_V(N, K));
359  return celt_cwrsi(N, K, idx, y);
360 }
361 
362 /*
363  * Faster than libopus's search, operates entirely in the signed domain.
364  * Slightly worse/better depending on N, K and the input vector.
365  */
366 static float ppp_pvq_search_c(float *X, int *y, int K, int N)
367 {
368  int i, y_norm = 0;
369  float res = 0.0f, xy_norm = 0.0f;
370 
371  for (i = 0; i < N; i++)
372  res += FFABS(X[i]);
373 
374  res = K/(res + FLT_EPSILON);
375 
376  for (i = 0; i < N; i++) {
377  y[i] = lrintf(res*X[i]);
378  y_norm += y[i]*y[i];
379  xy_norm += y[i]*X[i];
380  K -= FFABS(y[i]);
381  }
382 
383  while (K) {
384  int max_idx = 0, phase = FFSIGN(K);
385  float max_num = 0.0f;
386  float max_den = 1.0f;
387  y_norm += 1.0f;
388 
389  for (i = 0; i < N; i++) {
390  /* If the sum has been overshot and the best place has 0 pulses allocated
391  * to it, attempting to decrease it further will actually increase the
392  * sum. Prevent this by disregarding any 0 positions when decrementing. */
393  const int ca = 1 ^ ((y[i] == 0) & (phase < 0));
394  const int y_new = y_norm + 2*phase*FFABS(y[i]);
395  float xy_new = xy_norm + 1*phase*FFABS(X[i]);
396  xy_new = xy_new * xy_new;
397  if (ca && (max_den*xy_new) > (y_new*max_num)) {
398  max_den = y_new;
399  max_num = xy_new;
400  max_idx = i;
401  }
402  }
403 
404  K -= phase;
405 
406  phase *= FFSIGN(X[max_idx]);
407  xy_norm += 1*phase*X[max_idx];
408  y_norm += 2*phase*y[max_idx];
409  y[max_idx] += phase;
410  }
411 
412  return (float)y_norm;
413 }
414 
415 static uint32_t celt_alg_quant(OpusRangeCoder *rc, float *X, uint32_t N, uint32_t K,
416  enum CeltSpread spread, uint32_t blocks, float gain,
417  CeltPVQ *pvq)
418 {
419  int *y = pvq->qcoeff;
420 
421  celt_exp_rotation(X, N, blocks, K, spread, 1);
422  gain /= sqrtf(pvq->pvq_search(X, y, K, N));
423  celt_encode_pulses(rc, y, N, K);
424  celt_normalize_residual(y, X, N, gain);
425  celt_exp_rotation(X, N, blocks, K, spread, 0);
426  return celt_extract_collapse_mask(y, N, blocks);
427 }
428 
429 /** Decode pulse vector and combine the result with the pitch vector to produce
430  the final normalised signal in the current band. */
431 static uint32_t celt_alg_unquant(OpusRangeCoder *rc, float *X, uint32_t N, uint32_t K,
432  enum CeltSpread spread, uint32_t blocks, float gain,
433  CeltPVQ *pvq)
434 {
435  int *y = pvq->qcoeff;
436 
437  gain /= sqrtf(celt_decode_pulses(rc, y, N, K));
438  celt_normalize_residual(y, X, N, gain);
439  celt_exp_rotation(X, N, blocks, K, spread, 0);
440  return celt_extract_collapse_mask(y, N, blocks);
441 }
442 
443 static int celt_calc_theta(const float *X, const float *Y, int coupling, int N)
444 {
445  int i;
446  float e[2] = { 0.0f, 0.0f };
447  if (coupling) { /* Coupling case */
448  for (i = 0; i < N; i++) {
449  e[0] += (X[i] + Y[i])*(X[i] + Y[i]);
450  e[1] += (X[i] - Y[i])*(X[i] - Y[i]);
451  }
452  } else {
453  for (i = 0; i < N; i++) {
454  e[0] += X[i]*X[i];
455  e[1] += Y[i]*Y[i];
456  }
457  }
458  return lrintf(32768.0f*atan2f(sqrtf(e[1]), sqrtf(e[0]))/M_PI);
459 }
460 
461 static void celt_stereo_is_decouple(float *X, float *Y, float e_l, float e_r, int N)
462 {
463  int i;
464  const float energy_n = 1.0f/(sqrtf(e_l*e_l + e_r*e_r) + FLT_EPSILON);
465  e_l *= energy_n;
466  e_r *= energy_n;
467  for (i = 0; i < N; i++)
468  X[i] = e_l*X[i] + e_r*Y[i];
469 }
470 
471 static void celt_stereo_ms_decouple(float *X, float *Y, int N)
472 {
473  int i;
474  for (i = 0; i < N; i++) {
475  const float Xret = X[i];
476  X[i] = (X[i] + Y[i])*M_SQRT1_2;
477  Y[i] = (Y[i] - Xret)*M_SQRT1_2;
478  }
479 }
480 
482  OpusRangeCoder *rc,
483  const int band, float *X,
484  float *Y, int N, int b,
485  uint32_t blocks, float *lowband,
486  int duration, float *lowband_out,
487  int level, float gain,
488  float *lowband_scratch,
489  int fill, int quant,
490  QUANT_FN(*rec))
491 {
492  int i;
493  const uint8_t *cache;
494  int stereo = !!Y, split = stereo;
495  int imid = 0, iside = 0;
496  uint32_t N0 = N;
497  int N_B = N / blocks;
498  int N_B0 = N_B;
499  int B0 = blocks;
500  int time_divide = 0;
501  int recombine = 0;
502  int inv = 0;
503  float mid = 0, side = 0;
504  int longblocks = (B0 == 1);
505  uint32_t cm = 0;
506 
507  if (N == 1) {
508  float *x = X;
509  for (i = 0; i <= stereo; i++) {
510  int sign = 0;
511  if (f->remaining2 >= 1 << 3) {
512  if (quant) {
513  sign = x[0] < 0;
514  ff_opus_rc_put_raw(rc, sign, 1);
515  } else {
516  sign = ff_opus_rc_get_raw(rc, 1);
517  }
518  f->remaining2 -= 1 << 3;
519  }
520  x[0] = 1.0f - 2.0f*sign;
521  x = Y;
522  }
523  if (lowband_out)
524  lowband_out[0] = X[0];
525  return 1;
526  }
527 
528  if (!stereo && level == 0) {
529  int tf_change = f->tf_change[band];
530  int k;
531  if (tf_change > 0)
532  recombine = tf_change;
533  /* Band recombining to increase frequency resolution */
534 
535  if (lowband &&
536  (recombine || ((N_B & 1) == 0 && tf_change < 0) || B0 > 1)) {
537  for (i = 0; i < N; i++)
538  lowband_scratch[i] = lowband[i];
539  lowband = lowband_scratch;
540  }
541 
542  for (k = 0; k < recombine; k++) {
543  if (quant || lowband)
544  celt_haar1(quant ? X : lowband, N >> k, 1 << k);
545  fill = ff_celt_bit_interleave[fill & 0xF] | ff_celt_bit_interleave[fill >> 4] << 2;
546  }
547  blocks >>= recombine;
548  N_B <<= recombine;
549 
550  /* Increasing the time resolution */
551  while ((N_B & 1) == 0 && tf_change < 0) {
552  if (quant || lowband)
553  celt_haar1(quant ? X : lowband, N_B, blocks);
554  fill |= fill << blocks;
555  blocks <<= 1;
556  N_B >>= 1;
557  time_divide++;
558  tf_change++;
559  }
560  B0 = blocks;
561  N_B0 = N_B;
562 
563  /* Reorganize the samples in time order instead of frequency order */
564  if (B0 > 1 && (quant || lowband))
565  celt_deinterleave_hadamard(pvq->hadamard_tmp, quant ? X : lowband,
566  N_B >> recombine, B0 << recombine,
567  longblocks);
568  }
569 
570  /* If we need 1.5 more bit than we can produce, split the band in two. */
571  cache = ff_celt_cache_bits +
572  ff_celt_cache_index[(duration + 1) * CELT_MAX_BANDS + band];
573  if (!stereo && duration >= 0 && b > cache[cache[0]] + 12 && N > 2) {
574  N >>= 1;
575  Y = X + N;
576  split = 1;
577  duration -= 1;
578  if (blocks == 1)
579  fill = (fill & 1) | (fill << 1);
580  blocks = (blocks + 1) >> 1;
581  }
582 
583  if (split) {
584  int qn;
585  int itheta = quant ? celt_calc_theta(X, Y, stereo, N) : 0;
586  int mbits, sbits, delta;
587  int qalloc;
588  int pulse_cap;
589  int offset;
590  int orig_fill;
591  int tell;
592 
593  /* Decide on the resolution to give to the split parameter theta */
594  pulse_cap = ff_celt_log_freq_range[band] + duration * 8;
595  offset = (pulse_cap >> 1) - (stereo && N == 2 ? CELT_QTHETA_OFFSET_TWOPHASE :
597  qn = (stereo && band >= f->intensity_stereo) ? 1 :
598  celt_compute_qn(N, b, offset, pulse_cap, stereo);
599  tell = opus_rc_tell_frac(rc);
600  if (qn != 1) {
601  if (quant)
602  itheta = (itheta*qn + 8192) >> 14;
603  /* Entropy coding of the angle. We use a uniform pdf for the
604  * time split, a step for stereo, and a triangular one for the rest. */
605  if (quant) {
606  if (stereo && N > 2)
607  ff_opus_rc_enc_uint_step(rc, itheta, qn / 2);
608  else if (stereo || B0 > 1)
609  ff_opus_rc_enc_uint(rc, itheta, qn + 1);
610  else
611  ff_opus_rc_enc_uint_tri(rc, itheta, qn);
612  itheta = itheta * 16384 / qn;
613  if (stereo) {
614  if (itheta == 0)
615  celt_stereo_is_decouple(X, Y, f->block[0].lin_energy[band],
616  f->block[1].lin_energy[band], N);
617  else
618  celt_stereo_ms_decouple(X, Y, N);
619  }
620  } else {
621  if (stereo && N > 2)
622  itheta = ff_opus_rc_dec_uint_step(rc, qn / 2);
623  else if (stereo || B0 > 1)
624  itheta = ff_opus_rc_dec_uint(rc, qn+1);
625  else
626  itheta = ff_opus_rc_dec_uint_tri(rc, qn);
627  itheta = itheta * 16384 / qn;
628  }
629  } else if (stereo) {
630  if (quant) {
631  inv = itheta > 8192;
632  if (inv) {
633  for (i = 0; i < N; i++)
634  Y[i] *= -1;
635  }
636  celt_stereo_is_decouple(X, Y, f->block[0].lin_energy[band],
637  f->block[1].lin_energy[band], N);
638 
639  if (b > 2 << 3 && f->remaining2 > 2 << 3) {
640  ff_opus_rc_enc_log(rc, inv, 2);
641  } else {
642  inv = 0;
643  }
644  } else {
645  inv = (b > 2 << 3 && f->remaining2 > 2 << 3) ? ff_opus_rc_dec_log(rc, 2) : 0;
646  }
647  itheta = 0;
648  }
649  qalloc = opus_rc_tell_frac(rc) - tell;
650  b -= qalloc;
651 
652  orig_fill = fill;
653  if (itheta == 0) {
654  imid = 32767;
655  iside = 0;
656  fill = av_mod_uintp2(fill, blocks);
657  delta = -16384;
658  } else if (itheta == 16384) {
659  imid = 0;
660  iside = 32767;
661  fill &= ((1 << blocks) - 1) << blocks;
662  delta = 16384;
663  } else {
664  imid = celt_cos(itheta);
665  iside = celt_cos(16384-itheta);
666  /* This is the mid vs side allocation that minimizes squared error
667  in that band. */
668  delta = ROUND_MUL16((N - 1) << 7, celt_log2tan(iside, imid));
669  }
670 
671  mid = imid / 32768.0f;
672  side = iside / 32768.0f;
673 
674  /* This is a special case for N=2 that only works for stereo and takes
675  advantage of the fact that mid and side are orthogonal to encode
676  the side with just one bit. */
677  if (N == 2 && stereo) {
678  int c;
679  int sign = 0;
680  float tmp;
681  float *x2, *y2;
682  mbits = b;
683  /* Only need one bit for the side */
684  sbits = (itheta != 0 && itheta != 16384) ? 1 << 3 : 0;
685  mbits -= sbits;
686  c = (itheta > 8192);
687  f->remaining2 -= qalloc+sbits;
688 
689  x2 = c ? Y : X;
690  y2 = c ? X : Y;
691  if (sbits) {
692  if (quant) {
693  sign = x2[0]*y2[1] - x2[1]*y2[0] < 0;
694  ff_opus_rc_put_raw(rc, sign, 1);
695  } else {
696  sign = ff_opus_rc_get_raw(rc, 1);
697  }
698  }
699  sign = 1 - 2 * sign;
700  /* We use orig_fill here because we want to fold the side, but if
701  itheta==16384, we'll have cleared the low bits of fill. */
702  cm = rec(pvq, f, rc, band, x2, NULL, N, mbits, blocks, lowband, duration,
703  lowband_out, level, gain, lowband_scratch, orig_fill);
704  /* We don't split N=2 bands, so cm is either 1 or 0 (for a fold-collapse),
705  and there's no need to worry about mixing with the other channel. */
706  y2[0] = -sign * x2[1];
707  y2[1] = sign * x2[0];
708  X[0] *= mid;
709  X[1] *= mid;
710  Y[0] *= side;
711  Y[1] *= side;
712  tmp = X[0];
713  X[0] = tmp - Y[0];
714  Y[0] = tmp + Y[0];
715  tmp = X[1];
716  X[1] = tmp - Y[1];
717  Y[1] = tmp + Y[1];
718  } else {
719  /* "Normal" split code */
720  float *next_lowband2 = NULL;
721  float *next_lowband_out1 = NULL;
722  int next_level = 0;
723  int rebalance;
724  uint32_t cmt;
725 
726  /* Give more bits to low-energy MDCTs than they would
727  * otherwise deserve */
728  if (B0 > 1 && !stereo && (itheta & 0x3fff)) {
729  if (itheta > 8192)
730  /* Rough approximation for pre-echo masking */
731  delta -= delta >> (4 - duration);
732  else
733  /* Corresponds to a forward-masking slope of
734  * 1.5 dB per 10 ms */
735  delta = FFMIN(0, delta + (N << 3 >> (5 - duration)));
736  }
737  mbits = av_clip((b - delta) / 2, 0, b);
738  sbits = b - mbits;
739  f->remaining2 -= qalloc;
740 
741  if (lowband && !stereo)
742  next_lowband2 = lowband + N; /* >32-bit split case */
743 
744  /* Only stereo needs to pass on lowband_out.
745  * Otherwise, it's handled at the end */
746  if (stereo)
747  next_lowband_out1 = lowband_out;
748  else
749  next_level = level + 1;
750 
751  rebalance = f->remaining2;
752  if (mbits >= sbits) {
753  /* In stereo mode, we do not apply a scaling to the mid
754  * because we need the normalized mid for folding later */
755  cm = rec(pvq, f, rc, band, X, NULL, N, mbits, blocks, lowband,
756  duration, next_lowband_out1, next_level,
757  stereo ? 1.0f : (gain * mid), lowband_scratch, fill);
758  rebalance = mbits - (rebalance - f->remaining2);
759  if (rebalance > 3 << 3 && itheta != 0)
760  sbits += rebalance - (3 << 3);
761 
762  /* For a stereo split, the high bits of fill are always zero,
763  * so no folding will be done to the side. */
764  cmt = rec(pvq, f, rc, band, Y, NULL, N, sbits, blocks, next_lowband2,
765  duration, NULL, next_level, gain * side, NULL,
766  fill >> blocks);
767  cm |= cmt << ((B0 >> 1) & (stereo - 1));
768  } else {
769  /* For a stereo split, the high bits of fill are always zero,
770  * so no folding will be done to the side. */
771  cm = rec(pvq, f, rc, band, Y, NULL, N, sbits, blocks, next_lowband2,
772  duration, NULL, next_level, gain * side, NULL, fill >> blocks);
773  cm <<= ((B0 >> 1) & (stereo - 1));
774  rebalance = sbits - (rebalance - f->remaining2);
775  if (rebalance > 3 << 3 && itheta != 16384)
776  mbits += rebalance - (3 << 3);
777 
778  /* In stereo mode, we do not apply a scaling to the mid because
779  * we need the normalized mid for folding later */
780  cm |= rec(pvq, f, rc, band, X, NULL, N, mbits, blocks, lowband, duration,
781  next_lowband_out1, next_level, stereo ? 1.0f : (gain * mid),
782  lowband_scratch, fill);
783  }
784  }
785  } else {
786  /* This is the basic no-split case */
787  uint32_t q = celt_bits2pulses(cache, b);
788  uint32_t curr_bits = celt_pulses2bits(cache, q);
789  f->remaining2 -= curr_bits;
790 
791  /* Ensures we can never bust the budget */
792  while (f->remaining2 < 0 && q > 0) {
793  f->remaining2 += curr_bits;
794  curr_bits = celt_pulses2bits(cache, --q);
795  f->remaining2 -= curr_bits;
796  }
797 
798  if (q != 0) {
799  /* Finally do the actual (de)quantization */
800  if (quant) {
801  cm = celt_alg_quant(rc, X, N, (q < 8) ? q : (8 + (q & 7)) << ((q >> 3) - 1),
802  f->spread, blocks, gain, pvq);
803  } else {
804  cm = celt_alg_unquant(rc, X, N, (q < 8) ? q : (8 + (q & 7)) << ((q >> 3) - 1),
805  f->spread, blocks, gain, pvq);
806  }
807  } else {
808  /* If there's no pulse, fill the band anyway */
809  uint32_t cm_mask = (1 << blocks) - 1;
810  fill &= cm_mask;
811  if (fill) {
812  if (!lowband) {
813  /* Noise */
814  for (i = 0; i < N; i++)
815  X[i] = (((int32_t)celt_rng(f)) >> 20);
816  cm = cm_mask;
817  } else {
818  /* Folded spectrum */
819  for (i = 0; i < N; i++) {
820  /* About 48 dB below the "normal" folding level */
821  X[i] = lowband[i] + (((celt_rng(f)) & 0x8000) ? 1.0f / 256 : -1.0f / 256);
822  }
823  cm = fill;
824  }
825  celt_renormalize_vector(X, N, gain);
826  } else {
827  memset(X, 0, N*sizeof(float));
828  }
829  }
830  }
831 
832  /* This code is used by the decoder and by the resynthesis-enabled encoder */
833  if (stereo) {
834  if (N > 2)
835  celt_stereo_merge(X, Y, mid, N);
836  if (inv) {
837  for (i = 0; i < N; i++)
838  Y[i] *= -1;
839  }
840  } else if (level == 0) {
841  int k;
842 
843  /* Undo the sample reorganization going from time order to frequency order */
844  if (B0 > 1)
845  celt_interleave_hadamard(pvq->hadamard_tmp, X, N_B >> recombine,
846  B0 << recombine, longblocks);
847 
848  /* Undo time-freq changes that we did earlier */
849  N_B = N_B0;
850  blocks = B0;
851  for (k = 0; k < time_divide; k++) {
852  blocks >>= 1;
853  N_B <<= 1;
854  cm |= cm >> blocks;
855  celt_haar1(X, N_B, blocks);
856  }
857 
858  for (k = 0; k < recombine; k++) {
860  celt_haar1(X, N0>>k, 1<<k);
861  }
862  blocks <<= recombine;
863 
864  /* Scale output for later folding */
865  if (lowband_out) {
866  float n = sqrtf(N0);
867  for (i = 0; i < N0; i++)
868  lowband_out[i] = n * X[i];
869  }
870  cm = av_mod_uintp2(cm, blocks);
871  }
872 
873  return cm;
874 }
875 
876 
877 static QUANT_FN(pvq_decode_band)
878 {
879  return quant_band_template(pvq, f, rc, band, X, Y, N, b, blocks, lowband, duration,
880  lowband_out, level, gain, lowband_scratch, fill, 0,
881  pvq->decode_band);
882 }
883 
884 static QUANT_FN(pvq_encode_band)
885 {
886  return quant_band_template(pvq, f, rc, band, X, Y, N, b, blocks, lowband, duration,
887  lowband_out, level, gain, lowband_scratch, fill, 1,
888  pvq->encode_band);
889 }
890 
891 static float pvq_band_cost(CeltPVQ *pvq, CeltFrame *f, OpusRangeCoder *rc, int band,
892  float *bits, float lambda)
893 {
894  int i, b = 0;
895  uint32_t cm[2] = { (1 << f->blocks) - 1, (1 << f->blocks) - 1 };
896  const int band_size = ff_celt_freq_range[band] << f->size;
897  float buf[176 * 2], lowband_scratch[176], norm1[176], norm2[176];
898  float dist, cost, err_x = 0.0f, err_y = 0.0f;
899  float *X = buf;
900  float *X_orig = f->block[0].coeffs + (ff_celt_freq_bands[band] << f->size);
901  float *Y = (f->channels == 2) ? &buf[176] : NULL;
902  float *Y_orig = f->block[1].coeffs + (ff_celt_freq_bands[band] << f->size);
904 
905  memcpy(X, X_orig, band_size*sizeof(float));
906  if (Y)
907  memcpy(Y, Y_orig, band_size*sizeof(float));
908 
909  f->remaining2 = ((f->framebits << 3) - f->anticollapse_needed) - opus_rc_tell_frac(rc) - 1;
910  if (band <= f->coded_bands - 1) {
911  int curr_balance = f->remaining / FFMIN(3, f->coded_bands - band);
912  b = av_clip_uintp2(FFMIN(f->remaining2 + 1, f->pulses[band] + curr_balance), 14);
913  }
914 
915  if (f->dual_stereo) {
916  pvq->encode_band(pvq, f, rc, band, X, NULL, band_size, b / 2, f->blocks, NULL,
917  f->size, norm1, 0, 1.0f, lowband_scratch, cm[0]);
918 
919  pvq->encode_band(pvq, f, rc, band, Y, NULL, band_size, b / 2, f->blocks, NULL,
920  f->size, norm2, 0, 1.0f, lowband_scratch, cm[1]);
921  } else {
922  pvq->encode_band(pvq, f, rc, band, X, Y, band_size, b, f->blocks, NULL, f->size,
923  norm1, 0, 1.0f, lowband_scratch, cm[0] | cm[1]);
924  }
925 
926  for (i = 0; i < band_size; i++) {
927  err_x += (X[i] - X_orig[i])*(X[i] - X_orig[i]);
928  err_y += (Y[i] - Y_orig[i])*(Y[i] - Y_orig[i]);
929  }
930 
931  dist = sqrtf(err_x) + sqrtf(err_y);
932  cost = OPUS_RC_CHECKPOINT_BITS(rc)/8.0f;
933  *bits += cost;
934 
936 
937  return lambda*dist*cost;
938 }
939 
941 {
942  CeltPVQ *s = av_malloc(sizeof(CeltPVQ));
943  if (!s)
944  return AVERROR(ENOMEM);
945 
947  s->decode_band = pvq_decode_band;
948  s->encode_band = pvq_encode_band;
950 
951  if (ARCH_X86)
953 
954  *pvq = s;
955 
956  return 0;
957 }
958 
960 {
961  av_freep(pvq);
962 }
int channels
Definition: opus_celt.h:99
static void celt_stereo_merge(float *X, float *Y, float mid, int N)
Definition: opus_pvq.c:161
#define NULL
Definition: coverity.c:32
const char const char void * val
Definition: avisynth_c.h:771
const char * s
Definition: avisynth_c.h:768
const uint8_t ff_celt_cache_bits[392]
Definition: opustab.c:885
const uint32_t *const ff_celt_pvq_u_row[15]
Definition: opustab.c:1152
#define OPUS_RC_CHECKPOINT_SPAWN(rc)
Definition: opus_rc.h:116
void ff_opus_dsp_init_x86(struct CeltPVQ *s)
Definition: opus_dsp_init.c:31
static float pvq_band_cost(CeltPVQ *pvq, CeltFrame *f, OpusRangeCoder *rc, int band, float *bits, float lambda)
Definition: opus_pvq.c:891
static int celt_compute_qn(int N, int b, int offset, int pulse_cap, int stereo)
Definition: opus_pvq.c:239
int framebits
Definition: opus_celt.h:130
const uint8_t ff_celt_log_freq_range[]
Definition: opustab.c:771
int remaining2
Definition: opus_celt.h:132
float coeffs[CELT_MAX_FRAME_SIZE]
Definition: opus_celt.h:75
const char * g
Definition: vf_curves.c:112
Definition: vf_geq.c:47
static int celt_calc_theta(const float *X, const float *Y, int coupling, int N)
Definition: opus_pvq.c:443
#define M_SQRT1_2
Definition: mathematics.h:58
static float ppp_pvq_search_c(float *X, int *y, int K, int N)
Definition: opus_pvq.c:366
const uint8_t ff_celt_freq_bands[]
Definition: opustab.c:763
static const int8_t pulses[4]
Number of non-zero pulses in the MP-MLQ excitation.
Definition: g723_1.h:720
const char * b
Definition: vf_curves.c:113
void ff_opus_rc_enc_log(OpusRangeCoder *rc, int val, uint32_t bits)
Definition: opus_rc.c:131
uint32_t ff_opus_rc_dec_log(OpusRangeCoder *rc, uint32_t bits)
Definition: opus_rc.c:114
const uint8_t ff_celt_bit_deinterleave[]
Definition: opustab.c:933
static uint32_t celt_icwrsi(uint32_t N, uint32_t K, const int *y)
Definition: opus_pvq.c:256
const uint16_t ff_celt_qn_exp2[]
Definition: opustab.c:946
void ff_opus_rc_enc_uint(OpusRangeCoder *rc, uint32_t val, uint32_t size)
CELT: write a uniformly distributed integer.
Definition: opus_rc.c:204
static uint32_t celt_alg_unquant(OpusRangeCoder *rc, float *X, uint32_t N, uint32_t K, enum CeltSpread spread, uint32_t blocks, float gain, CeltPVQ *pvq)
Decode pulse vector and combine the result with the pitch vector to produce the final normalised sign...
Definition: opus_pvq.c:431
static void celt_interleave_hadamard(float *tmp, float *X, int N0, int stride, int hadamard)
Definition: opus_pvq.c:199
const uint8_t ff_celt_hadamard_order[]
Definition: opustab.c:938
uint32_t ff_opus_rc_dec_uint_tri(OpusRangeCoder *rc, int qn)
Definition: opus_rc.c:234
#define opus_ilog(i)
Definition: opus_rc.h:31
CeltBlock block[2]
Definition: opus_celt.h:97
uint8_t bits
Definition: crc.c:296
uint8_t
static void celt_stereo_ms_decouple(float *X, float *Y, int N)
Definition: opus_pvq.c:471
#define av_cold
Definition: attributes.h:82
#define av_malloc(s)
float delta
#define Y
Definition: vf_boxblur.c:76
int qcoeff[256]
Definition: opus_pvq.h:36
QUANT_FN * encode_band
Definition: opus_pvq.h:42
int av_cold ff_celt_pvq_init(CeltPVQ **pvq)
Definition: opus_pvq.c:940
#define cosf(x)
Definition: libm.h:78
int64_t duration
Definition: movenc.c:63
#define N
Definition: vf_pp7.c:73
static void celt_exp_rotation_impl(float *X, uint32_t len, uint32_t stride, float c, float s)
Definition: opus_pvq.c:84
int dual_stereo
Definition: opus_celt.h:118
static av_always_inline uint32_t quant_band_template(CeltPVQ *pvq, CeltFrame *f, OpusRangeCoder *rc, const int band, float *X, float *Y, int N, int b, uint32_t blocks, float *lowband, int duration, float *lowband_out, int level, float gain, float *lowband_scratch, int fill, int quant, QUANT_FN(*rec))
Definition: opus_pvq.c:481
#define FFMIN3(a, b, c)
Definition: common.h:97
#define atan2f(y, x)
Definition: libm.h:45
static void celt_deinterleave_hadamard(float *tmp, float *X, int N0, int stride, int hadamard)
Definition: opus_pvq.c:212
static float celt_decode_pulses(OpusRangeCoder *rc, int *y, uint32_t N, uint32_t K)
Definition: opus_pvq.c:356
#define lrintf(x)
Definition: libm_mips.h:70
#define N2
Definition: vf_pp7.c:69
static void celt_haar1(float *X, int N0, int stride)
Definition: opus_pvq.c:225
int coded_bands
Definition: opus_celt.h:105
float lin_energy[CELT_MAX_BANDS]
Definition: opus_celt.h:67
#define cm
Definition: dvbsubdec.c:37
void ff_opus_rc_enc_uint_tri(OpusRangeCoder *rc, uint32_t k, int qn)
Definition: opus_rc.c:258
QUANT_FN * decode_band
Definition: opus_pvq.h:41
const int16_t ff_celt_cache_index[105]
Definition: opustab.c:915
static int celt_log2tan(int isin, int icos)
Definition: opus_pvq.c:39
static void celt_encode_pulses(OpusRangeCoder *rc, int *y, uint32_t N, uint32_t K)
Definition: opus_pvq.c:351
#define AVERROR(e)
Definition: error.h:43
#define OPUS_RC_CHECKPOINT_ROLLBACK(rc)
Definition: opus_rc.h:123
int tf_change[CELT_MAX_BANDS]
Definition: opus_celt.h:137
int pulses[CELT_MAX_BANDS]
Definition: opus_celt.h:136
static const uint8_t offset[127][2]
Definition: vf_spp.c:92
int anticollapse_needed
Definition: opus_celt.h:115
#define MUL16(ra, rb)
Definition: mathops.h:88
float hadamard_tmp[256]
Definition: opus_pvq.h:37
static char * split(char *message, char delim)
Definition: af_channelmap.c:81
#define E
Definition: avdct.c:32
#define FFMIN(a, b)
Definition: common.h:96
#define FFSIGN(a)
Definition: common.h:73
static int16_t celt_cos(int16_t x)
Definition: opus_pvq.c:32
#define OPUS_RC_CHECKPOINT_BITS(rc)
Definition: opus_rc.h:120
GLsizei GLboolean const GLfloat * value
Definition: opengl_enc.c:109
#define N0
Definition: vf_pp7.c:67
int32_t
static uint64_t celt_cwrsi(uint32_t N, uint32_t K, uint32_t i, int *y)
Definition: opus_pvq.c:268
#define FFABS(a)
Absolute value, Note, INT_MIN / INT64_MIN result in undefined behavior as they are not representable ...
Definition: common.h:72
static uint32_t celt_alg_quant(OpusRangeCoder *rc, float *X, uint32_t N, uint32_t K, enum CeltSpread spread, uint32_t blocks, float gain, CeltPVQ *pvq)
Definition: opus_pvq.c:415
int blocks
Definition: opus_celt.h:112
int n
Definition: avisynth_c.h:684
const uint8_t ff_celt_freq_range[]
Definition: opustab.c:767
float(* band_cost)(struct CeltPVQ *pvq, CeltFrame *f, OpusRangeCoder *rc, int band, float *bits, float lambda)
Definition: opus_pvq.h:43
#define sinf(x)
Definition: libm.h:419
#define CELT_MAX_BANDS
Definition: opus.h:45
static void encode(AVCodecContext *ctx, AVFrame *frame, AVPacket *pkt, FILE *output)
Definition: encode_audio.c:95
void * buf
Definition: avisynth_c.h:690
#define CELT_PVQ_U(n, k)
Definition: opus_pvq.c:29
float(* pvq_search)(float *X, int *y, int K, int N)
Definition: opus_pvq.h:39
uint32_t ff_opus_rc_get_raw(OpusRangeCoder *rc, uint32_t count)
CELT: read 1-25 raw bits at the end of the frame, backwards byte-wise.
Definition: opus_rc.c:140
uint32_t ff_opus_rc_dec_uint_step(OpusRangeCoder *rc, int k0)
Definition: opus_rc.c:211
CeltSpread
Definition: opus_celt.h:49
static av_always_inline void celt_renormalize_vector(float *X, int N, float gain)
Definition: opus_celt.h:147
static uint32_t celt_extract_collapse_mask(const int *iy, uint32_t N, uint32_t B)
Definition: opus_pvq.c:147
static int celt_pulses2bits(const uint8_t *cache, int pulses)
Definition: opus_pvq.c:70
const uint8_t * quant
#define CELT_PVQ_V(n, k)
Definition: opus_pvq.c:30
uint8_t level
Definition: svq3.c:207
#define CELT_QTHETA_OFFSET
Definition: opus_celt.h:41
int remaining
Definition: opus_celt.h:131
static void celt_stereo_is_decouple(float *X, float *Y, float e_l, float e_r, int N)
Definition: opus_pvq.c:461
GLint GLenum GLboolean GLsizei stride
Definition: opengl_enc.c:105
#define ROUND_MUL16(a, b)
Definition: opus.h:51
void ff_opus_rc_put_raw(OpusRangeCoder *rc, uint32_t val, uint32_t count)
CELT: write 0 - 31 bits to the rawbits buffer.
Definition: opus_rc.c:161
static int celt_bits2pulses(const uint8_t *cache, int bits)
Definition: opus_pvq.c:51
static double c[64]
static av_always_inline uint32_t celt_rng(CeltFrame *f)
Definition: opus_celt.h:141
enum CeltSpread spread
Definition: opus_celt.h:121
const uint8_t ff_celt_bit_interleave[]
Definition: opustab.c:929
int len
static void celt_exp_rotation(float *X, uint32_t len, uint32_t stride, uint32_t K, enum CeltSpread spread, const int encode)
Definition: opus_pvq.c:107
static void celt_normalize_residual(const int *av_restrict iy, float *av_restrict X, int N, float g)
Definition: opus_pvq.c:76
uint32_t ff_opus_rc_dec_uint(OpusRangeCoder *rc, uint32_t size)
CELT: read a uniform distribution.
Definition: opus_rc.c:182
#define av_freep(p)
#define av_always_inline
Definition: attributes.h:39
#define M_PI
Definition: mathematics.h:52
void ff_opus_rc_enc_uint_step(OpusRangeCoder *rc, uint32_t val, int k0)
Definition: opus_rc.c:226
enum CeltBlockSize size
Definition: opus_celt.h:102
#define stride
static QUANT_FN(pvq_decode_band)
Definition: opus_pvq.c:877
#define B0
Definition: faandct.c:40
void av_cold ff_celt_pvq_uninit(CeltPVQ **pvq)
Definition: opus_pvq.c:959
int intensity_stereo
Definition: opus_celt.h:117
#define CELT_QTHETA_OFFSET_TWOPHASE
Definition: opus_celt.h:42
static av_always_inline uint32_t opus_rc_tell_frac(const OpusRangeCoder *rc)
Definition: opus_rc.h:66
static uint8_t tmp[11]
Definition: aes_ctr.c:26