FFmpeg
wmavoice.c
Go to the documentation of this file.
1 /*
2  * Windows Media Audio Voice decoder.
3  * Copyright (c) 2009 Ronald S. Bultje
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  * @brief Windows Media Audio Voice compatible decoder
25  * @author Ronald S. Bultje <rsbultje@gmail.com>
26  */
27 
28 #include <math.h>
29 
31 #include "libavutil/float_dsp.h"
32 #include "libavutil/mem_internal.h"
33 #include "libavutil/thread.h"
34 #include "avcodec.h"
35 #include "codec_internal.h"
36 #include "decode.h"
37 #include "get_bits.h"
38 #include "put_bits.h"
39 #include "wmavoice_data.h"
40 #include "celp_filters.h"
41 #include "acelp_vectors.h"
42 #include "acelp_filters.h"
43 #include "lsp.h"
44 #include "dct.h"
45 #include "rdft.h"
46 #include "sinewin.h"
47 
48 #define MAX_BLOCKS 8 ///< maximum number of blocks per frame
49 #define MAX_LSPS 16 ///< maximum filter order
50 #define MAX_LSPS_ALIGN16 16 ///< same as #MAX_LSPS; needs to be multiple
51  ///< of 16 for ASM input buffer alignment
52 #define MAX_FRAMES 3 ///< maximum number of frames per superframe
53 #define MAX_FRAMESIZE 160 ///< maximum number of samples per frame
54 #define MAX_SIGNAL_HISTORY 416 ///< maximum excitation signal history
55 #define MAX_SFRAMESIZE (MAX_FRAMESIZE * MAX_FRAMES)
56  ///< maximum number of samples per superframe
57 #define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that
58  ///< was split over two packets
59 #define VLC_NBITS 6 ///< number of bits to read per VLC iteration
60 
61 /**
62  * Frame type VLC coding.
63  */
65 
66 /**
67  * Adaptive codebook types.
68  */
69 enum {
70  ACB_TYPE_NONE = 0, ///< no adaptive codebook (only hardcoded fixed)
71  ACB_TYPE_ASYMMETRIC = 1, ///< adaptive codebook with per-frame pitch, which
72  ///< we interpolate to get a per-sample pitch.
73  ///< Signal is generated using an asymmetric sinc
74  ///< window function
75  ///< @note see #wmavoice_ipol1_coeffs
76  ACB_TYPE_HAMMING = 2 ///< Per-block pitch with signal generation using
77  ///< a Hamming sinc window function
78  ///< @note see #wmavoice_ipol2_coeffs
79 };
80 
81 /**
82  * Fixed codebook types.
83  */
84 enum {
85  FCB_TYPE_SILENCE = 0, ///< comfort noise during silence
86  ///< generated from a hardcoded (fixed) codebook
87  ///< with per-frame (low) gain values
88  FCB_TYPE_HARDCODED = 1, ///< hardcoded (fixed) codebook with per-block
89  ///< gain values
90  FCB_TYPE_AW_PULSES = 2, ///< Pitch-adaptive window (AW) pulse signals,
91  ///< used in particular for low-bitrate streams
92  FCB_TYPE_EXC_PULSES = 3, ///< Innovation (fixed) codebook pulse sets in
93  ///< combinations of either single pulses or
94  ///< pulse pairs
95 };
96 
97 /**
98  * Description of frame types.
99  */
100 static const struct frame_type_desc {
101  uint8_t n_blocks; ///< amount of blocks per frame (each block
102  ///< (contains 160/#n_blocks samples)
103  uint8_t log_n_blocks; ///< log2(#n_blocks)
104  uint8_t acb_type; ///< Adaptive codebook type (ACB_TYPE_*)
105  uint8_t fcb_type; ///< Fixed codebook type (FCB_TYPE_*)
106  uint8_t dbl_pulses; ///< how many pulse vectors have pulse pairs
107  ///< (rather than just one single pulse)
108  ///< only if #fcb_type == #FCB_TYPE_EXC_PULSES
109 } frame_descs[17] = {
110  { 1, 0, ACB_TYPE_NONE, FCB_TYPE_SILENCE, 0 },
111  { 2, 1, ACB_TYPE_NONE, FCB_TYPE_HARDCODED, 0 },
127 };
128 
129 /**
130  * WMA Voice decoding context.
131  */
132 typedef struct WMAVoiceContext {
133  /**
134  * @name Global values specified in the stream header / extradata or used all over.
135  * @{
136  */
137  GetBitContext gb; ///< packet bitreader. During decoder init,
138  ///< it contains the extradata from the
139  ///< demuxer. During decoding, it contains
140  ///< packet data.
141  int8_t vbm_tree[25]; ///< converts VLC codes to frame type
142 
143  int spillover_bitsize; ///< number of bits used to specify
144  ///< #spillover_nbits in the packet header
145  ///< = ceil(log2(ctx->block_align << 3))
146  int history_nsamples; ///< number of samples in history for signal
147  ///< prediction (through ACB)
148 
149  /* postfilter specific values */
150  int do_apf; ///< whether to apply the averaged
151  ///< projection filter (APF)
152  int denoise_strength; ///< strength of denoising in Wiener filter
153  ///< [0-11]
154  int denoise_tilt_corr; ///< Whether to apply tilt correction to the
155  ///< Wiener filter coefficients (postfilter)
156  int dc_level; ///< Predicted amount of DC noise, based
157  ///< on which a DC removal filter is used
158 
159  int lsps; ///< number of LSPs per frame [10 or 16]
160  int lsp_q_mode; ///< defines quantizer defaults [0, 1]
161  int lsp_def_mode; ///< defines different sets of LSP defaults
162  ///< [0, 1]
163 
164  int min_pitch_val; ///< base value for pitch parsing code
165  int max_pitch_val; ///< max value + 1 for pitch parsing
166  int pitch_nbits; ///< number of bits used to specify the
167  ///< pitch value in the frame header
168  int block_pitch_nbits; ///< number of bits used to specify the
169  ///< first block's pitch value
170  int block_pitch_range; ///< range of the block pitch
171  int block_delta_pitch_nbits; ///< number of bits used to specify the
172  ///< delta pitch between this and the last
173  ///< block's pitch value, used in all but
174  ///< first block
175  int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is
176  ///< from -this to +this-1)
177  uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale
178  ///< conversion
179 
180  /**
181  * @}
182  *
183  * @name Packet values specified in the packet header or related to a packet.
184  *
185  * A packet is considered to be a single unit of data provided to this
186  * decoder by the demuxer.
187  * @{
188  */
189  int spillover_nbits; ///< number of bits of the previous packet's
190  ///< last superframe preceding this
191  ///< packet's first full superframe (useful
192  ///< for re-synchronization also)
193  int has_residual_lsps; ///< if set, superframes contain one set of
194  ///< LSPs that cover all frames, encoded as
195  ///< independent and residual LSPs; if not
196  ///< set, each frame contains its own, fully
197  ///< independent, LSPs
198  int skip_bits_next; ///< number of bits to skip at the next call
199  ///< to #wmavoice_decode_packet() (since
200  ///< they're part of the previous superframe)
201 
203  ///< cache for superframe data split over
204  ///< multiple packets
205  int sframe_cache_size; ///< set to >0 if we have data from an
206  ///< (incomplete) superframe from a previous
207  ///< packet that spilled over in the current
208  ///< packet; specifies the amount of bits in
209  ///< #sframe_cache
210  PutBitContext pb; ///< bitstream writer for #sframe_cache
211 
212  /**
213  * @}
214  *
215  * @name Frame and superframe values
216  * Superframe and frame data - these can change from frame to frame,
217  * although some of them do in that case serve as a cache / history for
218  * the next frame or superframe.
219  * @{
220  */
221  double prev_lsps[MAX_LSPS]; ///< LSPs of the last frame of the previous
222  ///< superframe
223  int last_pitch_val; ///< pitch value of the previous frame
224  int last_acb_type; ///< frame type [0-2] of the previous frame
225  int pitch_diff_sh16; ///< ((cur_pitch_val - #last_pitch_val)
226  ///< << 16) / #MAX_FRAMESIZE
227  float silence_gain; ///< set for use in blocks if #ACB_TYPE_NONE
228 
229  int aw_idx_is_ext; ///< whether the AW index was encoded in
230  ///< 8 bits (instead of 6)
231  int aw_pulse_range; ///< the range over which #aw_pulse_set1()
232  ///< can apply the pulse, relative to the
233  ///< value in aw_first_pulse_off. The exact
234  ///< position of the first AW-pulse is within
235  ///< [pulse_off, pulse_off + this], and
236  ///< depends on bitstream values; [16 or 24]
237  int aw_n_pulses[2]; ///< number of AW-pulses in each block; note
238  ///< that this number can be negative (in
239  ///< which case it basically means "zero")
240  int aw_first_pulse_off[2]; ///< index of first sample to which to
241  ///< apply AW-pulses, or -0xff if unset
242  int aw_next_pulse_off_cache; ///< the position (relative to start of the
243  ///< second block) at which pulses should
244  ///< start to be positioned, serves as a
245  ///< cache for pitch-adaptive window pulses
246  ///< between blocks
247 
248  int frame_cntr; ///< current frame index [0 - 0xFFFE]; is
249  ///< only used for comfort noise in #pRNG()
250  int nb_superframes; ///< number of superframes in current packet
251  float gain_pred_err[6]; ///< cache for gain prediction
253  ///< cache of the signal of previous
254  ///< superframes, used as a history for
255  ///< signal generation
256  float synth_history[MAX_LSPS]; ///< see #excitation_history
257  /**
258  * @}
259  *
260  * @name Postfilter values
261  *
262  * Variables used for postfilter implementation, mostly history for
263  * smoothing and so on, and context variables for FFT/iFFT.
264  * @{
265  */
266  RDFTContext rdft, irdft; ///< contexts for FFT-calculation in the
267  ///< postfilter (for denoise filter)
268  DCTContext dct, dst; ///< contexts for phase shift (in Hilbert
269  ///< transform, part of postfilter)
270  float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi]
271  ///< range
272  float postfilter_agc; ///< gain control memory, used in
273  ///< #adaptive_gain_control()
274  float dcf_mem[2]; ///< DC filter history
276  ///< zero filter output (i.e. excitation)
277  ///< by postfilter
279  int denoise_filter_cache_size; ///< samples in #denoise_filter_cache
280  DECLARE_ALIGNED(32, float, tilted_lpcs_pf)[0x80];
281  ///< aligned buffer for LPC tilting
283  ///< aligned buffer for denoise coefficients
285  ///< aligned buffer for postfilter speech
286  ///< synthesis
287  /**
288  * @}
289  */
291 
292 /**
293  * Set up the variable bit mode (VBM) tree from container extradata.
294  * @param gb bit I/O context.
295  * The bit context (s->gb) should be loaded with byte 23-46 of the
296  * container extradata (i.e. the ones containing the VBM tree).
297  * @param vbm_tree pointer to array to which the decoded VBM tree will be
298  * written.
299  * @return 0 on success, <0 on error.
300  */
301 static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25])
302 {
303  int cntr[8] = { 0 }, n, res;
304 
305  memset(vbm_tree, 0xff, sizeof(vbm_tree[0]) * 25);
306  for (n = 0; n < 17; n++) {
307  res = get_bits(gb, 3);
308  if (cntr[res] > 3) // should be >= 3 + (res == 7))
309  return -1;
310  vbm_tree[res * 3 + cntr[res]++] = n;
311  }
312  return 0;
313 }
314 
316 {
317  static const uint8_t bits[] = {
318  2, 2, 2, 4, 4, 4,
319  6, 6, 6, 8, 8, 8,
320  10, 10, 10, 12, 12, 12,
321  14, 14, 14, 14
322  };
323 
326  1, NULL, 0, 0, 0, 0, 132);
327 }
328 
330 {
332  int n;
333 
334  s->postfilter_agc = 0;
335  s->sframe_cache_size = 0;
336  s->skip_bits_next = 0;
337  for (n = 0; n < s->lsps; n++)
338  s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
339  memset(s->excitation_history, 0,
340  sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY);
341  memset(s->synth_history, 0,
342  sizeof(*s->synth_history) * MAX_LSPS);
343  memset(s->gain_pred_err, 0,
344  sizeof(s->gain_pred_err));
345 
346  if (s->do_apf) {
347  memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0,
348  sizeof(*s->synth_filter_out_buf) * s->lsps);
349  memset(s->dcf_mem, 0,
350  sizeof(*s->dcf_mem) * 2);
351  memset(s->zero_exc_pf, 0,
352  sizeof(*s->zero_exc_pf) * s->history_nsamples);
353  memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache));
354  }
355 }
356 
357 /**
358  * Set up decoder with parameters from demuxer (extradata etc.).
359  */
361 {
362  static AVOnce init_static_once = AV_ONCE_INIT;
363  int n, flags, pitch_range, lsp16_flag, ret;
365 
366  ff_thread_once(&init_static_once, wmavoice_init_static_data);
367 
368  /**
369  * Extradata layout:
370  * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c),
371  * - byte 19-22: flags field (annoyingly in LE; see below for known
372  * values),
373  * - byte 23-46: variable bitmode tree (really just 17 * 3 bits,
374  * rest is 0).
375  */
376  if (ctx->extradata_size != 46) {
378  "Invalid extradata size %d (should be 46)\n",
379  ctx->extradata_size);
380  return AVERROR_INVALIDDATA;
381  }
382  if (ctx->block_align <= 0 || ctx->block_align > (1<<22)) {
383  av_log(ctx, AV_LOG_ERROR, "Invalid block alignment %d.\n", ctx->block_align);
384  return AVERROR_INVALIDDATA;
385  }
386 
387  flags = AV_RL32(ctx->extradata + 18);
388  s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align);
389  s->do_apf = flags & 0x1;
390  if (s->do_apf) {
391  if ((ret = ff_rdft_init(&s->rdft, 7, DFT_R2C)) < 0 ||
392  (ret = ff_rdft_init(&s->irdft, 7, IDFT_C2R)) < 0 ||
393  (ret = ff_dct_init (&s->dct, 6, DCT_I)) < 0 ||
394  (ret = ff_dct_init (&s->dst, 6, DST_I)) < 0)
395  return ret;
396 
397  ff_sine_window_init(s->cos, 256);
398  memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0]));
399  for (n = 0; n < 255; n++) {
400  s->sin[n] = -s->sin[510 - n];
401  s->cos[510 - n] = s->cos[n];
402  }
403  }
404  s->denoise_strength = (flags >> 2) & 0xF;
405  if (s->denoise_strength >= 12) {
407  "Invalid denoise filter strength %d (max=11)\n",
408  s->denoise_strength);
409  return AVERROR_INVALIDDATA;
410  }
411  s->denoise_tilt_corr = !!(flags & 0x40);
412  s->dc_level = (flags >> 7) & 0xF;
413  s->lsp_q_mode = !!(flags & 0x2000);
414  s->lsp_def_mode = !!(flags & 0x4000);
415  lsp16_flag = flags & 0x1000;
416  if (lsp16_flag) {
417  s->lsps = 16;
418  } else {
419  s->lsps = 10;
420  }
421  for (n = 0; n < s->lsps; n++)
422  s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
423 
424  init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3);
425  if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) {
426  av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n");
427  return AVERROR_INVALIDDATA;
428  }
429 
430  if (ctx->sample_rate >= INT_MAX / (256 * 37))
431  return AVERROR_INVALIDDATA;
432 
433  s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8;
434  s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8;
435  pitch_range = s->max_pitch_val - s->min_pitch_val;
436  if (pitch_range <= 0) {
437  av_log(ctx, AV_LOG_ERROR, "Invalid pitch range; broken extradata?\n");
438  return AVERROR_INVALIDDATA;
439  }
440  s->pitch_nbits = av_ceil_log2(pitch_range);
441  s->last_pitch_val = 40;
442  s->last_acb_type = ACB_TYPE_NONE;
443  s->history_nsamples = s->max_pitch_val + 8;
444 
445  if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) {
446  int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8,
447  max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8;
448 
450  "Unsupported samplerate %d (min=%d, max=%d)\n",
451  ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz
452 
453  return AVERROR(ENOSYS);
454  }
455 
456  s->block_conv_table[0] = s->min_pitch_val;
457  s->block_conv_table[1] = (pitch_range * 25) >> 6;
458  s->block_conv_table[2] = (pitch_range * 44) >> 6;
459  s->block_conv_table[3] = s->max_pitch_val - 1;
460  s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF;
461  if (s->block_delta_pitch_hrange <= 0) {
462  av_log(ctx, AV_LOG_ERROR, "Invalid delta pitch hrange; broken extradata?\n");
463  return AVERROR_INVALIDDATA;
464  }
465  s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange);
466  s->block_pitch_range = s->block_conv_table[2] +
467  s->block_conv_table[3] + 1 +
468  2 * (s->block_conv_table[1] - 2 * s->min_pitch_val);
469  s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range);
470 
471  av_channel_layout_uninit(&ctx->ch_layout);
473  ctx->sample_fmt = AV_SAMPLE_FMT_FLT;
474 
475  return 0;
476 }
477 
478 /**
479  * @name Postfilter functions
480  * Postfilter functions (gain control, wiener denoise filter, DC filter,
481  * kalman smoothening, plus surrounding code to wrap it)
482  * @{
483  */
484 /**
485  * Adaptive gain control (as used in postfilter).
486  *
487  * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except
488  * that the energy here is calculated using sum(abs(...)), whereas the
489  * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)).
490  *
491  * @param out output buffer for filtered samples
492  * @param in input buffer containing the samples as they are after the
493  * postfilter steps so far
494  * @param speech_synth input buffer containing speech synth before postfilter
495  * @param size input buffer size
496  * @param alpha exponential filter factor
497  * @param gain_mem pointer to filter memory (single float)
498  */
499 static void adaptive_gain_control(float *out, const float *in,
500  const float *speech_synth,
501  int size, float alpha, float *gain_mem)
502 {
503  int i;
504  float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor;
505  float mem = *gain_mem;
506 
507  for (i = 0; i < size; i++) {
508  speech_energy += fabsf(speech_synth[i]);
509  postfilter_energy += fabsf(in[i]);
510  }
511  gain_scale_factor = postfilter_energy == 0.0 ? 0.0 :
512  (1.0 - alpha) * speech_energy / postfilter_energy;
513 
514  for (i = 0; i < size; i++) {
515  mem = alpha * mem + gain_scale_factor;
516  out[i] = in[i] * mem;
517  }
518 
519  *gain_mem = mem;
520 }
521 
522 /**
523  * Kalman smoothing function.
524  *
525  * This function looks back pitch +/- 3 samples back into history to find
526  * the best fitting curve (that one giving the optimal gain of the two
527  * signals, i.e. the highest dot product between the two), and then
528  * uses that signal history to smoothen the output of the speech synthesis
529  * filter.
530  *
531  * @param s WMA Voice decoding context
532  * @param pitch pitch of the speech signal
533  * @param in input speech signal
534  * @param out output pointer for smoothened signal
535  * @param size input/output buffer size
536  *
537  * @returns -1 if no smoothening took place, e.g. because no optimal
538  * fit could be found, or 0 on success.
539  */
540 static int kalman_smoothen(WMAVoiceContext *s, int pitch,
541  const float *in, float *out, int size)
542 {
543  int n;
544  float optimal_gain = 0, dot;
545  const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)],
546  *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)],
547  *best_hist_ptr = NULL;
548 
549  /* find best fitting point in history */
550  do {
551  dot = avpriv_scalarproduct_float_c(in, ptr, size);
552  if (dot > optimal_gain) {
553  optimal_gain = dot;
554  best_hist_ptr = ptr;
555  }
556  } while (--ptr >= end);
557 
558  if (optimal_gain <= 0)
559  return -1;
560  dot = avpriv_scalarproduct_float_c(best_hist_ptr, best_hist_ptr, size);
561  if (dot <= 0) // would be 1.0
562  return -1;
563 
564  if (optimal_gain <= dot) {
565  dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000
566  } else
567  dot = 0.625;
568 
569  /* actual smoothing */
570  for (n = 0; n < size; n++)
571  out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]);
572 
573  return 0;
574 }
575 
576 /**
577  * Get the tilt factor of a formant filter from its transfer function
578  * @see #tilt_factor() in amrnbdec.c, which does essentially the same,
579  * but somehow (??) it does a speech synthesis filter in the
580  * middle, which is missing here
581  *
582  * @param lpcs LPC coefficients
583  * @param n_lpcs Size of LPC buffer
584  * @returns the tilt factor
585  */
586 static float tilt_factor(const float *lpcs, int n_lpcs)
587 {
588  float rh0, rh1;
589 
590  rh0 = 1.0 + avpriv_scalarproduct_float_c(lpcs, lpcs, n_lpcs);
591  rh1 = lpcs[0] + avpriv_scalarproduct_float_c(lpcs, &lpcs[1], n_lpcs - 1);
592 
593  return rh1 / rh0;
594 }
595 
596 /**
597  * Derive denoise filter coefficients (in real domain) from the LPCs.
598  */
599 static void calc_input_response(WMAVoiceContext *s, float *lpcs,
600  int fcb_type, float *coeffs, int remainder)
601 {
602  float last_coeff, min = 15.0, max = -15.0;
603  float irange, angle_mul, gain_mul, range, sq;
604  int n, idx;
605 
606  /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */
607  s->rdft.rdft_calc(&s->rdft, lpcs);
608 #define log_range(var, assign) do { \
609  float tmp = log10f(assign); var = tmp; \
610  max = FFMAX(max, tmp); min = FFMIN(min, tmp); \
611  } while (0)
612  log_range(last_coeff, lpcs[1] * lpcs[1]);
613  for (n = 1; n < 64; n++)
614  log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] +
615  lpcs[n * 2 + 1] * lpcs[n * 2 + 1]);
616  log_range(lpcs[0], lpcs[0] * lpcs[0]);
617 #undef log_range
618  range = max - min;
619  lpcs[64] = last_coeff;
620 
621  /* Now, use this spectrum to pick out these frequencies with higher
622  * (relative) power/energy (which we then take to be "not noise"),
623  * and set up a table (still in lpc[]) of (relative) gains per frequency.
624  * These frequencies will be maintained, while others ("noise") will be
625  * decreased in the filter output. */
626  irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63]
627  gain_mul = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) :
628  (5.0 / 14.7));
629  angle_mul = gain_mul * (8.0 * M_LN10 / M_PI);
630  for (n = 0; n <= 64; n++) {
631  float pwr;
632 
633  idx = lrint((max - lpcs[n]) * irange - 1);
634  idx = FFMAX(0, idx);
635  pwr = wmavoice_denoise_power_table[s->denoise_strength][idx];
636  lpcs[n] = angle_mul * pwr;
637 
638  /* 70.57 =~ 1/log10(1.0331663) */
639  idx = av_clipf((pwr * gain_mul - 0.0295) * 70.570526123, 0, INT_MAX / 2);
640 
641  if (idx > 127) { // fall back if index falls outside table range
642  coeffs[n] = wmavoice_energy_table[127] *
643  powf(1.0331663, idx - 127);
644  } else
645  coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)];
646  }
647 
648  /* calculate the Hilbert transform of the gains, which we do (since this
649  * is a sine input) by doing a phase shift (in theory, H(sin())=cos()).
650  * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the
651  * "moment" of the LPCs in this filter. */
652  s->dct.dct_calc(&s->dct, lpcs);
653  s->dst.dct_calc(&s->dst, lpcs);
654 
655  /* Split out the coefficient indexes into phase/magnitude pairs */
656  idx = 255 + av_clip(lpcs[64], -255, 255);
657  coeffs[0] = coeffs[0] * s->cos[idx];
658  idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255);
659  last_coeff = coeffs[64] * s->cos[idx];
660  for (n = 63;; n--) {
661  idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255);
662  coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
663  coeffs[n * 2] = coeffs[n] * s->cos[idx];
664 
665  if (!--n) break;
666 
667  idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255);
668  coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
669  coeffs[n * 2] = coeffs[n] * s->cos[idx];
670  }
671  coeffs[1] = last_coeff;
672 
673  /* move into real domain */
674  s->irdft.rdft_calc(&s->irdft, coeffs);
675 
676  /* tilt correction and normalize scale */
677  memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder));
678  if (s->denoise_tilt_corr) {
679  float tilt_mem = 0;
680 
681  coeffs[remainder - 1] = 0;
682  ff_tilt_compensation(&tilt_mem,
683  -1.8 * tilt_factor(coeffs, remainder - 1),
684  coeffs, remainder);
685  }
686  sq = (1.0 / 64.0) * sqrtf(1 / avpriv_scalarproduct_float_c(coeffs, coeffs,
687  remainder));
688  for (n = 0; n < remainder; n++)
689  coeffs[n] *= sq;
690 }
691 
692 /**
693  * This function applies a Wiener filter on the (noisy) speech signal as
694  * a means to denoise it.
695  *
696  * - take RDFT of LPCs to get the power spectrum of the noise + speech;
697  * - using this power spectrum, calculate (for each frequency) the Wiener
698  * filter gain, which depends on the frequency power and desired level
699  * of noise subtraction (when set too high, this leads to artifacts)
700  * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse
701  * of 4-8kHz);
702  * - by doing a phase shift, calculate the Hilbert transform of this array
703  * of per-frequency filter-gains to get the filtering coefficients;
704  * - smoothen/normalize/de-tilt these filter coefficients as desired;
705  * - take RDFT of noisy sound, apply the coefficients and take its IRDFT
706  * to get the denoised speech signal;
707  * - the leftover (i.e. output of the IRDFT on denoised speech data beyond
708  * the frame boundary) are saved and applied to subsequent frames by an
709  * overlap-add method (otherwise you get clicking-artifacts).
710  *
711  * @param s WMA Voice decoding context
712  * @param fcb_type Frame (codebook) type
713  * @param synth_pf input: the noisy speech signal, output: denoised speech
714  * data; should be 16-byte aligned (for ASM purposes)
715  * @param size size of the speech data
716  * @param lpcs LPCs used to synthesize this frame's speech data
717  */
718 static void wiener_denoise(WMAVoiceContext *s, int fcb_type,
719  float *synth_pf, int size,
720  const float *lpcs)
721 {
722  int remainder, lim, n;
723 
724  if (fcb_type != FCB_TYPE_SILENCE) {
725  float *tilted_lpcs = s->tilted_lpcs_pf,
726  *coeffs = s->denoise_coeffs_pf, tilt_mem = 0;
727 
728  tilted_lpcs[0] = 1.0;
729  memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps);
730  memset(&tilted_lpcs[s->lsps + 1], 0,
731  sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1));
732  ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps),
733  tilted_lpcs, s->lsps + 2);
734 
735  /* The IRDFT output (127 samples for 7-bit filter) beyond the frame
736  * size is applied to the next frame. All input beyond this is zero,
737  * and thus all output beyond this will go towards zero, hence we can
738  * limit to min(size-1, 127-size) as a performance consideration. */
739  remainder = FFMIN(127 - size, size - 1);
740  calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder);
741 
742  /* apply coefficients (in frequency spectrum domain), i.e. complex
743  * number multiplication */
744  memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size));
745  s->rdft.rdft_calc(&s->rdft, synth_pf);
746  s->rdft.rdft_calc(&s->rdft, coeffs);
747  synth_pf[0] *= coeffs[0];
748  synth_pf[1] *= coeffs[1];
749  for (n = 1; n < 64; n++) {
750  float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1];
751  synth_pf[n * 2] = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1];
752  synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1];
753  }
754  s->irdft.rdft_calc(&s->irdft, synth_pf);
755  }
756 
757  /* merge filter output with the history of previous runs */
758  if (s->denoise_filter_cache_size) {
759  lim = FFMIN(s->denoise_filter_cache_size, size);
760  for (n = 0; n < lim; n++)
761  synth_pf[n] += s->denoise_filter_cache[n];
762  s->denoise_filter_cache_size -= lim;
763  memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size],
764  sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size);
765  }
766 
767  /* move remainder of filter output into a cache for future runs */
768  if (fcb_type != FCB_TYPE_SILENCE) {
769  lim = FFMIN(remainder, s->denoise_filter_cache_size);
770  for (n = 0; n < lim; n++)
771  s->denoise_filter_cache[n] += synth_pf[size + n];
772  if (lim < remainder) {
773  memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim],
774  sizeof(s->denoise_filter_cache[0]) * (remainder - lim));
775  s->denoise_filter_cache_size = remainder;
776  }
777  }
778 }
779 
780 /**
781  * Averaging projection filter, the postfilter used in WMAVoice.
782  *
783  * This uses the following steps:
784  * - A zero-synthesis filter (generate excitation from synth signal)
785  * - Kalman smoothing on excitation, based on pitch
786  * - Re-synthesized smoothened output
787  * - Iterative Wiener denoise filter
788  * - Adaptive gain filter
789  * - DC filter
790  *
791  * @param s WMAVoice decoding context
792  * @param synth Speech synthesis output (before postfilter)
793  * @param samples Output buffer for filtered samples
794  * @param size Buffer size of synth & samples
795  * @param lpcs Generated LPCs used for speech synthesis
796  * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned)
797  * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses)
798  * @param pitch Pitch of the input signal
799  */
800 static void postfilter(WMAVoiceContext *s, const float *synth,
801  float *samples, int size,
802  const float *lpcs, float *zero_exc_pf,
803  int fcb_type, int pitch)
804 {
805  float synth_filter_in_buf[MAX_FRAMESIZE / 2],
806  *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16],
807  *synth_filter_in = zero_exc_pf;
808 
809  av_assert0(size <= MAX_FRAMESIZE / 2);
810 
811  /* generate excitation from input signal */
812  ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps);
813 
814  if (fcb_type >= FCB_TYPE_AW_PULSES &&
815  !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size))
816  synth_filter_in = synth_filter_in_buf;
817 
818  /* re-synthesize speech after smoothening, and keep history */
819  ff_celp_lp_synthesis_filterf(synth_pf, lpcs,
820  synth_filter_in, size, s->lsps);
821  memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps],
822  sizeof(synth_pf[0]) * s->lsps);
823 
824  wiener_denoise(s, fcb_type, synth_pf, size, lpcs);
825 
826  adaptive_gain_control(samples, synth_pf, synth, size, 0.99,
827  &s->postfilter_agc);
828 
829  if (s->dc_level > 8) {
830  /* remove ultra-low frequency DC noise / highpass filter;
831  * coefficients are identical to those used in SIPR decoding,
832  * and very closely resemble those used in AMR-NB decoding. */
834  (const float[2]) { -1.99997, 1.0 },
835  (const float[2]) { -1.9330735188, 0.93589198496 },
836  0.93980580475, s->dcf_mem, size);
837  }
838 }
839 /**
840  * @}
841  */
842 
843 /**
844  * Dequantize LSPs
845  * @param lsps output pointer to the array that will hold the LSPs
846  * @param num number of LSPs to be dequantized
847  * @param values quantized values, contains n_stages values
848  * @param sizes range (i.e. max value) of each quantized value
849  * @param n_stages number of dequantization runs
850  * @param table dequantization table to be used
851  * @param mul_q LSF multiplier
852  * @param base_q base (lowest) LSF values
853  */
854 static void dequant_lsps(double *lsps, int num,
855  const uint16_t *values,
856  const uint16_t *sizes,
857  int n_stages, const uint8_t *table,
858  const double *mul_q,
859  const double *base_q)
860 {
861  int n, m;
862 
863  memset(lsps, 0, num * sizeof(*lsps));
864  for (n = 0; n < n_stages; n++) {
865  const uint8_t *t_off = &table[values[n] * num];
866  double base = base_q[n], mul = mul_q[n];
867 
868  for (m = 0; m < num; m++)
869  lsps[m] += base + mul * t_off[m];
870 
871  table += sizes[n] * num;
872  }
873 }
874 
875 /**
876  * @name LSP dequantization routines
877  * LSP dequantization routines, for 10/16LSPs and independent/residual coding.
878  * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;
879  * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits.
880  * @{
881  */
882 /**
883  * Parse 10 independently-coded LSPs.
884  */
885 static void dequant_lsp10i(GetBitContext *gb, double *lsps)
886 {
887  static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 };
888  static const double mul_lsf[4] = {
889  5.2187144800e-3, 1.4626986422e-3,
890  9.6179549166e-4, 1.1325736225e-3
891  };
892  static const double base_lsf[4] = {
893  M_PI * -2.15522e-1, M_PI * -6.1646e-2,
894  M_PI * -3.3486e-2, M_PI * -5.7408e-2
895  };
896  uint16_t v[4];
897 
898  v[0] = get_bits(gb, 8);
899  v[1] = get_bits(gb, 6);
900  v[2] = get_bits(gb, 5);
901  v[3] = get_bits(gb, 5);
902 
903  dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i,
904  mul_lsf, base_lsf);
905 }
906 
907 /**
908  * Parse 10 independently-coded LSPs, and then derive the tables to
909  * generate LSPs for the other frames from them (residual coding).
910  */
912  double *i_lsps, const double *old,
913  double *a1, double *a2, int q_mode)
914 {
915  static const uint16_t vec_sizes[3] = { 128, 64, 64 };
916  static const double mul_lsf[3] = {
917  2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3
918  };
919  static const double base_lsf[3] = {
920  M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2
921  };
922  const float (*ipol_tab)[2][10] = q_mode ?
924  uint16_t interpol, v[3];
925  int n;
926 
927  dequant_lsp10i(gb, i_lsps);
928 
929  interpol = get_bits(gb, 5);
930  v[0] = get_bits(gb, 7);
931  v[1] = get_bits(gb, 6);
932  v[2] = get_bits(gb, 6);
933 
934  for (n = 0; n < 10; n++) {
935  double delta = old[n] - i_lsps[n];
936  a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
937  a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
938  }
939 
940  dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r,
941  mul_lsf, base_lsf);
942 }
943 
944 /**
945  * Parse 16 independently-coded LSPs.
946  */
947 static void dequant_lsp16i(GetBitContext *gb, double *lsps)
948 {
949  static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 };
950  static const double mul_lsf[5] = {
951  3.3439586280e-3, 6.9908173703e-4,
952  3.3216608306e-3, 1.0334960326e-3,
953  3.1899104283e-3
954  };
955  static const double base_lsf[5] = {
956  M_PI * -1.27576e-1, M_PI * -2.4292e-2,
957  M_PI * -1.28094e-1, M_PI * -3.2128e-2,
958  M_PI * -1.29816e-1
959  };
960  uint16_t v[5];
961 
962  v[0] = get_bits(gb, 8);
963  v[1] = get_bits(gb, 6);
964  v[2] = get_bits(gb, 7);
965  v[3] = get_bits(gb, 6);
966  v[4] = get_bits(gb, 7);
967 
968  dequant_lsps( lsps, 5, v, vec_sizes, 2,
969  wmavoice_dq_lsp16i1, mul_lsf, base_lsf);
970  dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2,
971  wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]);
972  dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1,
973  wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]);
974 }
975 
976 /**
977  * Parse 16 independently-coded LSPs, and then derive the tables to
978  * generate LSPs for the other frames from them (residual coding).
979  */
981  double *i_lsps, const double *old,
982  double *a1, double *a2, int q_mode)
983 {
984  static const uint16_t vec_sizes[3] = { 128, 128, 128 };
985  static const double mul_lsf[3] = {
986  1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3
987  };
988  static const double base_lsf[3] = {
989  M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2
990  };
991  const float (*ipol_tab)[2][16] = q_mode ?
993  uint16_t interpol, v[3];
994  int n;
995 
996  dequant_lsp16i(gb, i_lsps);
997 
998  interpol = get_bits(gb, 5);
999  v[0] = get_bits(gb, 7);
1000  v[1] = get_bits(gb, 7);
1001  v[2] = get_bits(gb, 7);
1002 
1003  for (n = 0; n < 16; n++) {
1004  double delta = old[n] - i_lsps[n];
1005  a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
1006  a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
1007  }
1008 
1009  dequant_lsps( a2, 10, v, vec_sizes, 1,
1010  wmavoice_dq_lsp16r1, mul_lsf, base_lsf);
1011  dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1,
1012  wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]);
1013  dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1,
1014  wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]);
1015 }
1016 
1017 /**
1018  * @}
1019  * @name Pitch-adaptive window coding functions
1020  * The next few functions are for pitch-adaptive window coding.
1021  * @{
1022  */
1023 /**
1024  * Parse the offset of the first pitch-adaptive window pulses, and
1025  * the distribution of pulses between the two blocks in this frame.
1026  * @param s WMA Voice decoding context private data
1027  * @param gb bit I/O context
1028  * @param pitch pitch for each block in this frame
1029  */
1031  const int *pitch)
1032 {
1033  static const int16_t start_offset[94] = {
1034  -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11,
1035  13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26,
1036  27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43,
1037  45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67,
1038  69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91,
1039  93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115,
1040  117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139,
1041  141, 143, 145, 147, 149, 151, 153, 155, 157, 159
1042  };
1043  int bits, offset;
1044 
1045  /* position of pulse */
1046  s->aw_idx_is_ext = 0;
1047  if ((bits = get_bits(gb, 6)) >= 54) {
1048  s->aw_idx_is_ext = 1;
1049  bits += (bits - 54) * 3 + get_bits(gb, 2);
1050  }
1051 
1052  /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count
1053  * the distribution of the pulses in each block contained in this frame. */
1054  s->aw_pulse_range = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16;
1055  for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ;
1056  s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0];
1057  s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2;
1058  offset += s->aw_n_pulses[0] * pitch[0];
1059  s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1];
1060  s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2;
1061 
1062  /* if continuing from a position before the block, reset position to
1063  * start of block (when corrected for the range over which it can be
1064  * spread in aw_pulse_set1()). */
1065  if (start_offset[bits] < MAX_FRAMESIZE / 2) {
1066  while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0)
1067  s->aw_first_pulse_off[1] -= pitch[1];
1068  if (start_offset[bits] < 0)
1069  while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0)
1070  s->aw_first_pulse_off[0] -= pitch[0];
1071  }
1072 }
1073 
1074 /**
1075  * Apply second set of pitch-adaptive window pulses.
1076  * @param s WMA Voice decoding context private data
1077  * @param gb bit I/O context
1078  * @param block_idx block index in frame [0, 1]
1079  * @param fcb structure containing fixed codebook vector info
1080  * @return -1 on error, 0 otherwise
1081  */
1083  int block_idx, AMRFixed *fcb)
1084 {
1085  uint16_t use_mask_mem[9]; // only 5 are used, rest is padding
1086  uint16_t *use_mask = use_mask_mem + 2;
1087  /* in this function, idx is the index in the 80-bit (+ padding) use_mask
1088  * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits
1089  * of idx are the position of the bit within a particular item in the
1090  * array (0 being the most significant bit, and 15 being the least
1091  * significant bit), and the remainder (>> 4) is the index in the
1092  * use_mask[]-array. This is faster and uses less memory than using a
1093  * 80-byte/80-int array. */
1094  int pulse_off = s->aw_first_pulse_off[block_idx],
1095  pulse_start, n, idx, range, aidx, start_off = 0;
1096 
1097  /* set offset of first pulse to within this block */
1098  if (s->aw_n_pulses[block_idx] > 0)
1099  while (pulse_off + s->aw_pulse_range < 1)
1100  pulse_off += fcb->pitch_lag;
1101 
1102  /* find range per pulse */
1103  if (s->aw_n_pulses[0] > 0) {
1104  if (block_idx == 0) {
1105  range = 32;
1106  } else /* block_idx = 1 */ {
1107  range = 8;
1108  if (s->aw_n_pulses[block_idx] > 0)
1109  pulse_off = s->aw_next_pulse_off_cache;
1110  }
1111  } else
1112  range = 16;
1113  pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0;
1114 
1115  /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,
1116  * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus
1117  * we exclude that range from being pulsed again in this function. */
1118  memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0]));
1119  memset( use_mask, -1, 5 * sizeof(use_mask[0]));
1120  memset(&use_mask[5], 0, 2 * sizeof(use_mask[0]));
1121  if (s->aw_n_pulses[block_idx] > 0)
1122  for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) {
1123  int excl_range = s->aw_pulse_range; // always 16 or 24
1124  uint16_t *use_mask_ptr = &use_mask[idx >> 4];
1125  int first_sh = 16 - (idx & 15);
1126  *use_mask_ptr++ &= 0xFFFFu << first_sh;
1127  excl_range -= first_sh;
1128  if (excl_range >= 16) {
1129  *use_mask_ptr++ = 0;
1130  *use_mask_ptr &= 0xFFFF >> (excl_range - 16);
1131  } else
1132  *use_mask_ptr &= 0xFFFF >> excl_range;
1133  }
1134 
1135  /* find the 'aidx'th offset that is not excluded */
1136  aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4);
1137  for (n = 0; n <= aidx; pulse_start++) {
1138  for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ;
1139  if (idx >= MAX_FRAMESIZE / 2) { // find from zero
1140  if (use_mask[0]) idx = 0x0F;
1141  else if (use_mask[1]) idx = 0x1F;
1142  else if (use_mask[2]) idx = 0x2F;
1143  else if (use_mask[3]) idx = 0x3F;
1144  else if (use_mask[4]) idx = 0x4F;
1145  else return -1;
1146  idx -= av_log2_16bit(use_mask[idx >> 4]);
1147  }
1148  if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) {
1149  use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15));
1150  n++;
1151  start_off = idx;
1152  }
1153  }
1154 
1155  fcb->x[fcb->n] = start_off;
1156  fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0;
1157  fcb->n++;
1158 
1159  /* set offset for next block, relative to start of that block */
1160  n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag;
1161  s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0;
1162  return 0;
1163 }
1164 
1165 /**
1166  * Apply first set of pitch-adaptive window pulses.
1167  * @param s WMA Voice decoding context private data
1168  * @param gb bit I/O context
1169  * @param block_idx block index in frame [0, 1]
1170  * @param fcb storage location for fixed codebook pulse info
1171  */
1173  int block_idx, AMRFixed *fcb)
1174 {
1175  int val = get_bits(gb, 12 - 2 * (s->aw_idx_is_ext && !block_idx));
1176  float v;
1177 
1178  if (s->aw_n_pulses[block_idx] > 0) {
1179  int n, v_mask, i_mask, sh, n_pulses;
1180 
1181  if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each
1182  n_pulses = 3;
1183  v_mask = 8;
1184  i_mask = 7;
1185  sh = 4;
1186  } else { // 4 pulses, 1:sign + 2:index each
1187  n_pulses = 4;
1188  v_mask = 4;
1189  i_mask = 3;
1190  sh = 3;
1191  }
1192 
1193  for (n = n_pulses - 1; n >= 0; n--, val >>= sh) {
1194  fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0;
1195  fcb->x[fcb->n] = (val & i_mask) * n_pulses + n +
1196  s->aw_first_pulse_off[block_idx];
1197  while (fcb->x[fcb->n] < 0)
1198  fcb->x[fcb->n] += fcb->pitch_lag;
1199  if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2)
1200  fcb->n++;
1201  }
1202  } else {
1203  int num2 = (val & 0x1FF) >> 1, delta, idx;
1204 
1205  if (num2 < 1 * 79) { delta = 1; idx = num2 + 1; }
1206  else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; }
1207  else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; }
1208  else { delta = 7; idx = num2 + 1 - 3 * 75; }
1209  v = (val & 0x200) ? -1.0 : 1.0;
1210 
1211  fcb->no_repeat_mask |= 3 << fcb->n;
1212  fcb->x[fcb->n] = idx - delta;
1213  fcb->y[fcb->n] = v;
1214  fcb->x[fcb->n + 1] = idx;
1215  fcb->y[fcb->n + 1] = (val & 1) ? -v : v;
1216  fcb->n += 2;
1217  }
1218 }
1219 
1220 /**
1221  * @}
1222  *
1223  * Generate a random number from frame_cntr and block_idx, which will live
1224  * in the range [0, 1000 - block_size] (so it can be used as an index in a
1225  * table of size 1000 of which you want to read block_size entries).
1226  *
1227  * @param frame_cntr current frame number
1228  * @param block_num current block index
1229  * @param block_size amount of entries we want to read from a table
1230  * that has 1000 entries
1231  * @return a (non-)random number in the [0, 1000 - block_size] range.
1232  */
1233 static int pRNG(int frame_cntr, int block_num, int block_size)
1234 {
1235  /* array to simplify the calculation of z:
1236  * y = (x % 9) * 5 + 6;
1237  * z = (49995 * x) / y;
1238  * Since y only has 9 values, we can remove the division by using a
1239  * LUT and using FASTDIV-style divisions. For each of the 9 values
1240  * of y, we can rewrite z as:
1241  * z = x * (49995 / y) + x * ((49995 % y) / y)
1242  * In this table, each col represents one possible value of y, the
1243  * first number is 49995 / y, and the second is the FASTDIV variant
1244  * of 49995 % y / y. */
1245  static const unsigned int div_tbl[9][2] = {
1246  { 8332, 3 * 715827883U }, // y = 6
1247  { 4545, 0 * 390451573U }, // y = 11
1248  { 3124, 11 * 268435456U }, // y = 16
1249  { 2380, 15 * 204522253U }, // y = 21
1250  { 1922, 23 * 165191050U }, // y = 26
1251  { 1612, 23 * 138547333U }, // y = 31
1252  { 1388, 27 * 119304648U }, // y = 36
1253  { 1219, 16 * 104755300U }, // y = 41
1254  { 1086, 39 * 93368855U } // y = 46
1255  };
1256  unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr;
1257  if (x >= 0xFFFF) x -= 0xFFFF; // max value of x is 8*1877+0xFFFE=0x13AA6,
1258  // so this is effectively a modulo (%)
1259  y = x - 9 * MULH(477218589, x); // x % 9
1260  z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1]));
1261  // z = x * 49995 / (y * 5 + 6)
1262  return z % (1000 - block_size);
1263 }
1264 
1265 /**
1266  * Parse hardcoded signal for a single block.
1267  * @note see #synth_block().
1268  */
1270  int block_idx, int size,
1271  const struct frame_type_desc *frame_desc,
1272  float *excitation)
1273 {
1274  float gain;
1275  int n, r_idx;
1276 
1278 
1279  /* Set the offset from which we start reading wmavoice_std_codebook */
1280  if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1281  r_idx = pRNG(s->frame_cntr, block_idx, size);
1282  gain = s->silence_gain;
1283  } else /* FCB_TYPE_HARDCODED */ {
1284  r_idx = get_bits(gb, 8);
1285  gain = wmavoice_gain_universal[get_bits(gb, 6)];
1286  }
1287 
1288  /* Clear gain prediction parameters */
1289  memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err));
1290 
1291  /* Apply gain to hardcoded codebook and use that as excitation signal */
1292  for (n = 0; n < size; n++)
1293  excitation[n] = wmavoice_std_codebook[r_idx + n] * gain;
1294 }
1295 
1296 /**
1297  * Parse FCB/ACB signal for a single block.
1298  * @note see #synth_block().
1299  */
1301  int block_idx, int size,
1302  int block_pitch_sh2,
1303  const struct frame_type_desc *frame_desc,
1304  float *excitation)
1305 {
1306  static const float gain_coeff[6] = {
1307  0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458
1308  };
1309  float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain;
1310  int n, idx, gain_weight;
1311  AMRFixed fcb;
1312 
1313  av_assert0(size <= MAX_FRAMESIZE / 2);
1314  memset(pulses, 0, sizeof(*pulses) * size);
1315 
1316  fcb.pitch_lag = block_pitch_sh2 >> 2;
1317  fcb.pitch_fac = 1.0;
1318  fcb.no_repeat_mask = 0;
1319  fcb.n = 0;
1320 
1321  /* For the other frame types, this is where we apply the innovation
1322  * (fixed) codebook pulses of the speech signal. */
1323  if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1324  aw_pulse_set1(s, gb, block_idx, &fcb);
1325  if (aw_pulse_set2(s, gb, block_idx, &fcb)) {
1326  /* Conceal the block with silence and return.
1327  * Skip the correct amount of bits to read the next
1328  * block from the correct offset. */
1329  int r_idx = pRNG(s->frame_cntr, block_idx, size);
1330 
1331  for (n = 0; n < size; n++)
1332  excitation[n] =
1333  wmavoice_std_codebook[r_idx + n] * s->silence_gain;
1334  skip_bits(gb, 7 + 1);
1335  return;
1336  }
1337  } else /* FCB_TYPE_EXC_PULSES */ {
1338  int offset_nbits = 5 - frame_desc->log_n_blocks;
1339 
1340  fcb.no_repeat_mask = -1;
1341  /* similar to ff_decode_10_pulses_35bits(), but with single pulses
1342  * (instead of double) for a subset of pulses */
1343  for (n = 0; n < 5; n++) {
1344  float sign;
1345  int pos1, pos2;
1346 
1347  sign = get_bits1(gb) ? 1.0 : -1.0;
1348  pos1 = get_bits(gb, offset_nbits);
1349  fcb.x[fcb.n] = n + 5 * pos1;
1350  fcb.y[fcb.n++] = sign;
1351  if (n < frame_desc->dbl_pulses) {
1352  pos2 = get_bits(gb, offset_nbits);
1353  fcb.x[fcb.n] = n + 5 * pos2;
1354  fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign;
1355  }
1356  }
1357  }
1358  ff_set_fixed_vector(pulses, &fcb, 1.0, size);
1359 
1360  /* Calculate gain for adaptive & fixed codebook signal.
1361  * see ff_amr_set_fixed_gain(). */
1362  idx = get_bits(gb, 7);
1363  fcb_gain = expf(avpriv_scalarproduct_float_c(s->gain_pred_err,
1364  gain_coeff, 6) -
1365  5.2409161640 + wmavoice_gain_codebook_fcb[idx]);
1366  acb_gain = wmavoice_gain_codebook_acb[idx];
1367  pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx],
1368  -2.9957322736 /* log(0.05) */,
1369  1.6094379124 /* log(5.0) */);
1370 
1371  gain_weight = 8 >> frame_desc->log_n_blocks;
1372  memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err,
1373  sizeof(*s->gain_pred_err) * (6 - gain_weight));
1374  for (n = 0; n < gain_weight; n++)
1375  s->gain_pred_err[n] = pred_err;
1376 
1377  /* Calculation of adaptive codebook */
1378  if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1379  int len;
1380  for (n = 0; n < size; n += len) {
1381  int next_idx_sh16;
1382  int abs_idx = block_idx * size + n;
1383  int pitch_sh16 = (s->last_pitch_val << 16) +
1384  s->pitch_diff_sh16 * abs_idx;
1385  int pitch = (pitch_sh16 + 0x6FFF) >> 16;
1386  int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000;
1387  idx = idx_sh16 >> 16;
1388  if (s->pitch_diff_sh16) {
1389  if (s->pitch_diff_sh16 > 0) {
1390  next_idx_sh16 = (idx_sh16) &~ 0xFFFF;
1391  } else
1392  next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF;
1393  len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8,
1394  1, size - n);
1395  } else
1396  len = size;
1397 
1398  ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch],
1400  idx, 9, len);
1401  }
1402  } else /* ACB_TYPE_HAMMING */ {
1403  int block_pitch = block_pitch_sh2 >> 2;
1404  idx = block_pitch_sh2 & 3;
1405  if (idx) {
1406  ff_acelp_interpolatef(excitation, &excitation[-block_pitch],
1408  idx, 8, size);
1409  } else
1410  av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch,
1411  sizeof(float) * size);
1412  }
1413 
1414  /* Interpolate ACB/FCB and use as excitation signal */
1415  ff_weighted_vector_sumf(excitation, excitation, pulses,
1416  acb_gain, fcb_gain, size);
1417 }
1418 
1419 /**
1420  * Parse data in a single block.
1421  *
1422  * @param s WMA Voice decoding context private data
1423  * @param gb bit I/O context
1424  * @param block_idx index of the to-be-read block
1425  * @param size amount of samples to be read in this block
1426  * @param block_pitch_sh2 pitch for this block << 2
1427  * @param lsps LSPs for (the end of) this frame
1428  * @param prev_lsps LSPs for the last frame
1429  * @param frame_desc frame type descriptor
1430  * @param excitation target memory for the ACB+FCB interpolated signal
1431  * @param synth target memory for the speech synthesis filter output
1432  * @return 0 on success, <0 on error.
1433  */
1435  int block_idx, int size,
1436  int block_pitch_sh2,
1437  const double *lsps, const double *prev_lsps,
1438  const struct frame_type_desc *frame_desc,
1439  float *excitation, float *synth)
1440 {
1441  double i_lsps[MAX_LSPS];
1442  float lpcs[MAX_LSPS];
1443  float fac;
1444  int n;
1445 
1446  if (frame_desc->acb_type == ACB_TYPE_NONE)
1447  synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation);
1448  else
1449  synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2,
1450  frame_desc, excitation);
1451 
1452  /* convert interpolated LSPs to LPCs */
1453  fac = (block_idx + 0.5) / frame_desc->n_blocks;
1454  for (n = 0; n < s->lsps; n++) // LSF -> LSP
1455  i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n]));
1456  ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1457 
1458  /* Speech synthesis */
1459  ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps);
1460 }
1461 
1462 /**
1463  * Synthesize output samples for a single frame.
1464  *
1465  * @param ctx WMA Voice decoder context
1466  * @param gb bit I/O context (s->gb or one for cross-packet superframes)
1467  * @param frame_idx Frame number within superframe [0-2]
1468  * @param samples pointer to output sample buffer, has space for at least 160
1469  * samples
1470  * @param lsps LSP array
1471  * @param prev_lsps array of previous frame's LSPs
1472  * @param excitation target buffer for excitation signal
1473  * @param synth target buffer for synthesized speech data
1474  * @return 0 on success, <0 on error.
1475  */
1476 static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx,
1477  float *samples,
1478  const double *lsps, const double *prev_lsps,
1479  float *excitation, float *synth)
1480 {
1482  int n, n_blocks_x2, log_n_blocks_x2, av_uninit(cur_pitch_val);
1483  int pitch[MAX_BLOCKS], av_uninit(last_block_pitch);
1484 
1485  /* Parse frame type ("frame header"), see frame_descs */
1486  int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)], block_nsamples;
1487 
1488  if (bd_idx < 0) {
1490  "Invalid frame type VLC code, skipping\n");
1491  return AVERROR_INVALIDDATA;
1492  }
1493 
1494  block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks;
1495 
1496  /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */
1497  if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) {
1498  /* Pitch is provided per frame, which is interpreted as the pitch of
1499  * the last sample of the last block of this frame. We can interpolate
1500  * the pitch of other blocks (and even pitch-per-sample) by gradually
1501  * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */
1502  n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1;
1503  log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1;
1504  cur_pitch_val = s->min_pitch_val + get_bits(gb, s->pitch_nbits);
1505  cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1);
1506  if (s->last_acb_type == ACB_TYPE_NONE ||
1507  20 * abs(cur_pitch_val - s->last_pitch_val) >
1508  (cur_pitch_val + s->last_pitch_val))
1509  s->last_pitch_val = cur_pitch_val;
1510 
1511  /* pitch per block */
1512  for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1513  int fac = n * 2 + 1;
1514 
1515  pitch[n] = (MUL16(fac, cur_pitch_val) +
1516  MUL16((n_blocks_x2 - fac), s->last_pitch_val) +
1517  frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2;
1518  }
1519 
1520  /* "pitch-diff-per-sample" for calculation of pitch per sample */
1521  s->pitch_diff_sh16 =
1522  (cur_pitch_val - s->last_pitch_val) * (1 << 16) / MAX_FRAMESIZE;
1523  }
1524 
1525  /* Global gain (if silence) and pitch-adaptive window coordinates */
1526  switch (frame_descs[bd_idx].fcb_type) {
1527  case FCB_TYPE_SILENCE:
1528  s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)];
1529  break;
1530  case FCB_TYPE_AW_PULSES:
1531  aw_parse_coords(s, gb, pitch);
1532  break;
1533  }
1534 
1535  for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1536  int bl_pitch_sh2;
1537 
1538  /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */
1539  switch (frame_descs[bd_idx].acb_type) {
1540  case ACB_TYPE_HAMMING: {
1541  /* Pitch is given per block. Per-block pitches are encoded as an
1542  * absolute value for the first block, and then delta values
1543  * relative to this value) for all subsequent blocks. The scale of
1544  * this pitch value is semi-logarithmic compared to its use in the
1545  * decoder, so we convert it to normal scale also. */
1546  int block_pitch,
1547  t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2,
1548  t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1,
1549  t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1;
1550 
1551  if (n == 0) {
1552  block_pitch = get_bits(gb, s->block_pitch_nbits);
1553  } else
1554  block_pitch = last_block_pitch - s->block_delta_pitch_hrange +
1555  get_bits(gb, s->block_delta_pitch_nbits);
1556  /* Convert last_ so that any next delta is within _range */
1557  last_block_pitch = av_clip(block_pitch,
1558  s->block_delta_pitch_hrange,
1559  s->block_pitch_range -
1560  s->block_delta_pitch_hrange);
1561 
1562  /* Convert semi-log-style scale back to normal scale */
1563  if (block_pitch < t1) {
1564  bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch;
1565  } else {
1566  block_pitch -= t1;
1567  if (block_pitch < t2) {
1568  bl_pitch_sh2 =
1569  (s->block_conv_table[1] << 2) + (block_pitch << 1);
1570  } else {
1571  block_pitch -= t2;
1572  if (block_pitch < t3) {
1573  bl_pitch_sh2 =
1574  (s->block_conv_table[2] + block_pitch) << 2;
1575  } else
1576  bl_pitch_sh2 = s->block_conv_table[3] << 2;
1577  }
1578  }
1579  pitch[n] = bl_pitch_sh2 >> 2;
1580  break;
1581  }
1582 
1583  case ACB_TYPE_ASYMMETRIC: {
1584  bl_pitch_sh2 = pitch[n] << 2;
1585  break;
1586  }
1587 
1588  default: // ACB_TYPE_NONE has no pitch
1589  bl_pitch_sh2 = 0;
1590  break;
1591  }
1592 
1593  synth_block(s, gb, n, block_nsamples, bl_pitch_sh2,
1594  lsps, prev_lsps, &frame_descs[bd_idx],
1595  &excitation[n * block_nsamples],
1596  &synth[n * block_nsamples]);
1597  }
1598 
1599  /* Averaging projection filter, if applicable. Else, just copy samples
1600  * from synthesis buffer */
1601  if (s->do_apf) {
1602  double i_lsps[MAX_LSPS];
1603  float lpcs[MAX_LSPS];
1604 
1605  for (n = 0; n < s->lsps; n++) // LSF -> LSP
1606  i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n]));
1607  ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1608  postfilter(s, synth, samples, 80, lpcs,
1609  &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx],
1610  frame_descs[bd_idx].fcb_type, pitch[0]);
1611 
1612  for (n = 0; n < s->lsps; n++) // LSF -> LSP
1613  i_lsps[n] = cos(lsps[n]);
1614  ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1615  postfilter(s, &synth[80], &samples[80], 80, lpcs,
1616  &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80],
1617  frame_descs[bd_idx].fcb_type, pitch[0]);
1618  } else
1619  memcpy(samples, synth, 160 * sizeof(synth[0]));
1620 
1621  /* Cache values for next frame */
1622  s->frame_cntr++;
1623  if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%)
1624  s->last_acb_type = frame_descs[bd_idx].acb_type;
1625  switch (frame_descs[bd_idx].acb_type) {
1626  case ACB_TYPE_NONE:
1627  s->last_pitch_val = 0;
1628  break;
1629  case ACB_TYPE_ASYMMETRIC:
1630  s->last_pitch_val = cur_pitch_val;
1631  break;
1632  case ACB_TYPE_HAMMING:
1633  s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1];
1634  break;
1635  }
1636 
1637  return 0;
1638 }
1639 
1640 /**
1641  * Ensure minimum value for first item, maximum value for last value,
1642  * proper spacing between each value and proper ordering.
1643  *
1644  * @param lsps array of LSPs
1645  * @param num size of LSP array
1646  *
1647  * @note basically a double version of #ff_acelp_reorder_lsf(), might be
1648  * useful to put in a generic location later on. Parts are also
1649  * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(),
1650  * which is in float.
1651  */
1652 static void stabilize_lsps(double *lsps, int num)
1653 {
1654  int n, m, l;
1655 
1656  /* set minimum value for first, maximum value for last and minimum
1657  * spacing between LSF values.
1658  * Very similar to ff_set_min_dist_lsf(), but in double. */
1659  lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI);
1660  for (n = 1; n < num; n++)
1661  lsps[n] = FFMAX(lsps[n], lsps[n - 1] + 0.0125 * M_PI);
1662  lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI);
1663 
1664  /* reorder (looks like one-time / non-recursed bubblesort).
1665  * Very similar to ff_sort_nearly_sorted_floats(), but in double. */
1666  for (n = 1; n < num; n++) {
1667  if (lsps[n] < lsps[n - 1]) {
1668  for (m = 1; m < num; m++) {
1669  double tmp = lsps[m];
1670  for (l = m - 1; l >= 0; l--) {
1671  if (lsps[l] <= tmp) break;
1672  lsps[l + 1] = lsps[l];
1673  }
1674  lsps[l + 1] = tmp;
1675  }
1676  break;
1677  }
1678  }
1679 }
1680 
1681 /**
1682  * Synthesize output samples for a single superframe. If we have any data
1683  * cached in s->sframe_cache, that will be used instead of whatever is loaded
1684  * in s->gb.
1685  *
1686  * WMA Voice superframes contain 3 frames, each containing 160 audio samples,
1687  * to give a total of 480 samples per frame. See #synth_frame() for frame
1688  * parsing. In addition to 3 frames, superframes can also contain the LSPs
1689  * (if these are globally specified for all frames (residually); they can
1690  * also be specified individually per-frame. See the s->has_residual_lsps
1691  * option), and can specify the number of samples encoded in this superframe
1692  * (if less than 480), usually used to prevent blanks at track boundaries.
1693  *
1694  * @param ctx WMA Voice decoder context
1695  * @return 0 on success, <0 on error or 1 if there was not enough data to
1696  * fully parse the superframe
1697  */
1699  int *got_frame_ptr)
1700 {
1702  GetBitContext *gb = &s->gb, s_gb;
1703  int n, res, n_samples = MAX_SFRAMESIZE;
1704  double lsps[MAX_FRAMES][MAX_LSPS];
1705  const double *mean_lsf = s->lsps == 16 ?
1706  wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode];
1707  float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12];
1708  float synth[MAX_LSPS + MAX_SFRAMESIZE];
1709  float *samples;
1710 
1711  memcpy(synth, s->synth_history,
1712  s->lsps * sizeof(*synth));
1713  memcpy(excitation, s->excitation_history,
1714  s->history_nsamples * sizeof(*excitation));
1715 
1716  if (s->sframe_cache_size > 0) {
1717  gb = &s_gb;
1718  init_get_bits(gb, s->sframe_cache, s->sframe_cache_size);
1719  s->sframe_cache_size = 0;
1720  }
1721 
1722  /* First bit is speech/music bit, it differentiates between WMAVoice
1723  * speech samples (the actual codec) and WMAVoice music samples, which
1724  * are really WMAPro-in-WMAVoice-superframes. I've never seen those in
1725  * the wild yet. */
1726  if (!get_bits1(gb)) {
1727  avpriv_request_sample(ctx, "WMAPro-in-WMAVoice");
1728  return AVERROR_PATCHWELCOME;
1729  }
1730 
1731  /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */
1732  if (get_bits1(gb)) {
1733  if ((n_samples = get_bits(gb, 12)) > MAX_SFRAMESIZE) {
1735  "Superframe encodes > %d samples (%d), not allowed\n",
1736  MAX_SFRAMESIZE, n_samples);
1737  return AVERROR_INVALIDDATA;
1738  }
1739  }
1740 
1741  /* Parse LSPs, if global for the superframe (can also be per-frame). */
1742  if (s->has_residual_lsps) {
1743  double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2];
1744 
1745  for (n = 0; n < s->lsps; n++)
1746  prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n];
1747 
1748  if (s->lsps == 10) {
1749  dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1750  } else /* s->lsps == 16 */
1751  dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1752 
1753  for (n = 0; n < s->lsps; n++) {
1754  lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]);
1755  lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]);
1756  lsps[2][n] += mean_lsf[n];
1757  }
1758  for (n = 0; n < 3; n++)
1759  stabilize_lsps(lsps[n], s->lsps);
1760  }
1761 
1762  /* synth_superframe can run multiple times per packet
1763  * free potential previous frame */
1765 
1766  /* get output buffer */
1767  frame->nb_samples = MAX_SFRAMESIZE;
1768  if ((res = ff_get_buffer(ctx, frame, 0)) < 0)
1769  return res;
1770  frame->nb_samples = n_samples;
1771  samples = (float *)frame->data[0];
1772 
1773  /* Parse frames, optionally preceded by per-frame (independent) LSPs. */
1774  for (n = 0; n < 3; n++) {
1775  if (!s->has_residual_lsps) {
1776  int m;
1777 
1778  if (s->lsps == 10) {
1779  dequant_lsp10i(gb, lsps[n]);
1780  } else /* s->lsps == 16 */
1781  dequant_lsp16i(gb, lsps[n]);
1782 
1783  for (m = 0; m < s->lsps; m++)
1784  lsps[n][m] += mean_lsf[m];
1785  stabilize_lsps(lsps[n], s->lsps);
1786  }
1787 
1788  if ((res = synth_frame(ctx, gb, n,
1789  &samples[n * MAX_FRAMESIZE],
1790  lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1],
1791  &excitation[s->history_nsamples + n * MAX_FRAMESIZE],
1792  &synth[s->lsps + n * MAX_FRAMESIZE]))) {
1793  *got_frame_ptr = 0;
1794  return res;
1795  }
1796  }
1797 
1798  /* Statistics? FIXME - we don't check for length, a slight overrun
1799  * will be caught by internal buffer padding, and anything else
1800  * will be skipped, not read. */
1801  if (get_bits1(gb)) {
1802  res = get_bits(gb, 4);
1803  skip_bits(gb, 10 * (res + 1));
1804  }
1805 
1806  if (get_bits_left(gb) < 0) {
1808  return AVERROR_INVALIDDATA;
1809  }
1810 
1811  *got_frame_ptr = 1;
1812 
1813  /* Update history */
1814  memcpy(s->prev_lsps, lsps[2],
1815  s->lsps * sizeof(*s->prev_lsps));
1816  memcpy(s->synth_history, &synth[MAX_SFRAMESIZE],
1817  s->lsps * sizeof(*synth));
1818  memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE],
1819  s->history_nsamples * sizeof(*excitation));
1820  if (s->do_apf)
1821  memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE],
1822  s->history_nsamples * sizeof(*s->zero_exc_pf));
1823 
1824  return 0;
1825 }
1826 
1827 /**
1828  * Parse the packet header at the start of each packet (input data to this
1829  * decoder).
1830  *
1831  * @param s WMA Voice decoding context private data
1832  * @return <0 on error, nb_superframes on success.
1833  */
1835 {
1836  GetBitContext *gb = &s->gb;
1837  unsigned int res, n_superframes = 0;
1838 
1839  skip_bits(gb, 4); // packet sequence number
1840  s->has_residual_lsps = get_bits1(gb);
1841  do {
1842  if (get_bits_left(gb) < 6 + s->spillover_bitsize)
1843  return AVERROR_INVALIDDATA;
1844 
1845  res = get_bits(gb, 6); // number of superframes per packet
1846  // (minus first one if there is spillover)
1847  n_superframes += res;
1848  } while (res == 0x3F);
1849  s->spillover_nbits = get_bits(gb, s->spillover_bitsize);
1850 
1851  return get_bits_left(gb) >= 0 ? n_superframes : AVERROR_INVALIDDATA;
1852 }
1853 
1854 /**
1855  * Copy (unaligned) bits from gb/data/size to pb.
1856  *
1857  * @param pb target buffer to copy bits into
1858  * @param data source buffer to copy bits from
1859  * @param size size of the source data, in bytes
1860  * @param gb bit I/O context specifying the current position in the source.
1861  * data. This function might use this to align the bit position to
1862  * a whole-byte boundary before calling #ff_copy_bits() on aligned
1863  * source data
1864  * @param nbits the amount of bits to copy from source to target
1865  *
1866  * @note after calling this function, the current position in the input bit
1867  * I/O context is undefined.
1868  */
1869 static void copy_bits(PutBitContext *pb,
1870  const uint8_t *data, int size,
1871  GetBitContext *gb, int nbits)
1872 {
1873  int rmn_bytes, rmn_bits;
1874 
1875  rmn_bits = rmn_bytes = get_bits_left(gb);
1876  if (rmn_bits < nbits)
1877  return;
1878  if (nbits > put_bits_left(pb))
1879  return;
1880  rmn_bits &= 7; rmn_bytes >>= 3;
1881  if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0)
1882  put_bits(pb, rmn_bits, get_bits(gb, rmn_bits));
1883  ff_copy_bits(pb, data + size - rmn_bytes,
1884  FFMIN(nbits - rmn_bits, rmn_bytes << 3));
1885 }
1886 
1887 /**
1888  * Packet decoding: a packet is anything that the (ASF) demuxer contains,
1889  * and we expect that the demuxer / application provides it to us as such
1890  * (else you'll probably get garbage as output). Every packet has a size of
1891  * ctx->block_align bytes, starts with a packet header (see
1892  * #parse_packet_header()), and then a series of superframes. Superframe
1893  * boundaries may exceed packets, i.e. superframes can split data over
1894  * multiple (two) packets.
1895  *
1896  * For more information about frames, see #synth_superframe().
1897  */
1899  int *got_frame_ptr, AVPacket *avpkt)
1900 {
1902  GetBitContext *gb = &s->gb;
1903  const uint8_t *buf = avpkt->data;
1904  uint8_t dummy[1];
1905  int size, res, pos;
1906 
1907  /* Packets are sometimes a multiple of ctx->block_align, with a packet
1908  * header at each ctx->block_align bytes. However, FFmpeg's ASF demuxer
1909  * feeds us ASF packets, which may concatenate multiple "codec" packets
1910  * in a single "muxer" packet, so we artificially emulate that by
1911  * capping the packet size at ctx->block_align. */
1912  for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align);
1913  buf = size ? buf : dummy;
1914  res = init_get_bits8(&s->gb, buf, size);
1915  if (res < 0)
1916  return res;
1917 
1918  /* size == ctx->block_align is used to indicate whether we are dealing with
1919  * a new packet or a packet of which we already read the packet header
1920  * previously. */
1921  if (!(size % ctx->block_align)) { // new packet header
1922  if (!size) {
1923  s->spillover_nbits = 0;
1924  s->nb_superframes = 0;
1925  } else {
1926  if ((res = parse_packet_header(s)) < 0)
1927  return res;
1928  s->nb_superframes = res;
1929  }
1930 
1931  /* If the packet header specifies a s->spillover_nbits, then we want
1932  * to push out all data of the previous packet (+ spillover) before
1933  * continuing to parse new superframes in the current packet. */
1934  if (s->sframe_cache_size > 0) {
1935  int cnt = get_bits_count(gb);
1936  if (cnt + s->spillover_nbits > avpkt->size * 8) {
1937  s->spillover_nbits = avpkt->size * 8 - cnt;
1938  }
1939  copy_bits(&s->pb, buf, size, gb, s->spillover_nbits);
1940  flush_put_bits(&s->pb);
1941  s->sframe_cache_size += s->spillover_nbits;
1942  if ((res = synth_superframe(ctx, frame, got_frame_ptr)) == 0 &&
1943  *got_frame_ptr) {
1944  cnt += s->spillover_nbits;
1945  s->skip_bits_next = cnt & 7;
1946  res = cnt >> 3;
1947  return res;
1948  } else
1949  skip_bits_long (gb, s->spillover_nbits - cnt +
1950  get_bits_count(gb)); // resync
1951  } else if (s->spillover_nbits) {
1952  skip_bits_long(gb, s->spillover_nbits); // resync
1953  }
1954  } else if (s->skip_bits_next)
1955  skip_bits(gb, s->skip_bits_next);
1956 
1957  /* Try parsing superframes in current packet */
1958  s->sframe_cache_size = 0;
1959  s->skip_bits_next = 0;
1960  pos = get_bits_left(gb);
1961  if (s->nb_superframes-- == 0) {
1962  *got_frame_ptr = 0;
1963  return size;
1964  } else if (s->nb_superframes > 0) {
1965  if ((res = synth_superframe(ctx, frame, got_frame_ptr)) < 0) {
1966  return res;
1967  } else if (*got_frame_ptr) {
1968  int cnt = get_bits_count(gb);
1969  s->skip_bits_next = cnt & 7;
1970  res = cnt >> 3;
1971  return res;
1972  }
1973  } else if ((s->sframe_cache_size = pos) > 0) {
1974  /* ... cache it for spillover in next packet */
1975  init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE);
1976  copy_bits(&s->pb, buf, size, gb, s->sframe_cache_size);
1977  // FIXME bad - just copy bytes as whole and add use the
1978  // skip_bits_next field
1979  }
1980 
1981  return size;
1982 }
1983 
1985 {
1987 
1988  if (s->do_apf) {
1989  ff_rdft_end(&s->rdft);
1990  ff_rdft_end(&s->irdft);
1991  ff_dct_end(&s->dct);
1992  ff_dct_end(&s->dst);
1993  }
1994 
1995  return 0;
1996 }
1997 
1999  .p.name = "wmavoice",
2000  CODEC_LONG_NAME("Windows Media Audio Voice"),
2001  .p.type = AVMEDIA_TYPE_AUDIO,
2002  .p.id = AV_CODEC_ID_WMAVOICE,
2003  .priv_data_size = sizeof(WMAVoiceContext),
2005  .close = wmavoice_decode_end,
2008  .caps_internal = FF_CODEC_CAP_INIT_CLEANUP,
2009  .flush = wmavoice_flush,
2010 };
WMAVoiceContext::has_residual_lsps
int has_residual_lsps
if set, superframes contain one set of LSPs that cover all frames, encoded as independent and residua...
Definition: wmavoice.c:193
skip_bits_long
static void skip_bits_long(GetBitContext *s, int n)
Skips the specified number of bits.
Definition: get_bits.h:268
AMRFixed::x
int x[10]
Definition: acelp_vectors.h:55
DCT_I
@ DCT_I
Definition: avfft.h:96
wmavoice_std_codebook
static const float wmavoice_std_codebook[1000]
Definition: wmavoice_data.h:2585
interpol
static int interpol(MBContext *s, uint32_t *color, int x, int y, int linesize)
Definition: vsrc_mandelbrot.c:186
MAX_LSPS
#define MAX_LSPS
maximum filter order
Definition: wmavoice.c:49
WMAVoiceContext::aw_next_pulse_off_cache
int aw_next_pulse_off_cache
the position (relative to start of the second block) at which pulses should start to be positioned,...
Definition: wmavoice.c:242
WMAVoiceContext::max_pitch_val
int max_pitch_val
max value + 1 for pitch parsing
Definition: wmavoice.c:165
av_clip
#define av_clip
Definition: common.h:95
aw_pulse_set2
static int aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb, int block_idx, AMRFixed *fcb)
Apply second set of pitch-adaptive window pulses.
Definition: wmavoice.c:1082
FF_CODEC_CAP_INIT_CLEANUP
#define FF_CODEC_CAP_INIT_CLEANUP
The codec allows calling the close function for deallocation even if the init function returned a fai...
Definition: codec_internal.h:42
acelp_vectors.h
get_bits_left
static int get_bits_left(GetBitContext *gb)
Definition: get_bits.h:664
AVERROR
Filter the word “frame” indicates either a video frame or a group of audio as stored in an AVFrame structure Format for each input and each output the list of supported formats For video that means pixel format For audio that means channel sample they are references to shared objects When the negotiation mechanism computes the intersection of the formats supported at each end of a all references to both lists are replaced with a reference to the intersection And when a single format is eventually chosen for a link amongst the remaining all references to the list are updated That means that if a filter requires that its input and output have the same format amongst a supported all it has to do is use a reference to the same list of formats query_formats can leave some formats unset and return AVERROR(EAGAIN) to cause the negotiation mechanism toagain later. That can be used by filters with complex requirements to use the format negotiated on one link to set the formats supported on another. Frame references ownership and permissions
FCB_TYPE_SILENCE
@ FCB_TYPE_SILENCE
comfort noise during silence generated from a hardcoded (fixed) codebook with per-frame (low) gain va...
Definition: wmavoice.c:85
wmavoice_dq_lsp10i
static const uint8_t wmavoice_dq_lsp10i[0xf00]
Definition: wmavoice_data.h:33
mem_internal.h
out
FILE * out
Definition: movenc.c:54
u
#define u(width, name, range_min, range_max)
Definition: cbs_h2645.c:262
thread.h
ACB_TYPE_NONE
@ ACB_TYPE_NONE
no adaptive codebook (only hardcoded fixed)
Definition: wmavoice.c:70
frame_descs
static const struct frame_type_desc frame_descs[17]
rdft.h
wmavoice_dq_lsp16r3
static const uint8_t wmavoice_dq_lsp16r3[0x600]
Definition: wmavoice_data.h:1526
dequant_lsps
static void dequant_lsps(double *lsps, int num, const uint16_t *values, const uint16_t *sizes, int n_stages, const uint8_t *table, const double *mul_q, const double *base_q)
Dequantize LSPs.
Definition: wmavoice.c:854
init_put_bits
static void init_put_bits(PutBitContext *s, uint8_t *buffer, int buffer_size)
Initialize the PutBitContext s.
Definition: put_bits.h:62
WMAVoiceContext::excitation_history
float excitation_history[MAX_SIGNAL_HISTORY]
cache of the signal of previous superframes, used as a history for signal generation
Definition: wmavoice.c:252
get_bits_count
static int get_bits_count(const GetBitContext *s)
Definition: get_bits.h:256
av_log2_16bit
int av_log2_16bit(unsigned v)
Definition: intmath.c:31
AVFrame
This structure describes decoded (raw) audio or video data.
Definition: frame.h:330
put_bits
static void put_bits(Jpeg2000EncoderContext *s, int val, int n)
put n times val bit
Definition: j2kenc.c:221
tmp
static uint8_t tmp[11]
Definition: aes_ctr.c:28
aw_pulse_set1
static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb, int block_idx, AMRFixed *fcb)
Apply first set of pitch-adaptive window pulses.
Definition: wmavoice.c:1172
ff_acelp_apply_order_2_transfer_function
void ff_acelp_apply_order_2_transfer_function(float *out, const float *in, const float zero_coeffs[2], const float pole_coeffs[2], float gain, float mem[2], int n)
Apply an order 2 rational transfer function in-place.
Definition: acelp_filters.c:121
WMAVoiceContext::irdft
RDFTContext irdft
contexts for FFT-calculation in the postfilter (for denoise filter)
Definition: wmavoice.c:266
AVPacket::data
uint8_t * data
Definition: packet.h:374
pRNG
static int pRNG(int frame_cntr, int block_num, int block_size)
Generate a random number from frame_cntr and block_idx, which will live in the range [0,...
Definition: wmavoice.c:1233
ff_wmavoice_decoder
const FFCodec ff_wmavoice_decoder
Definition: wmavoice.c:1998
table
static const uint16_t table[]
Definition: prosumer.c:205
data
const char data[16]
Definition: mxf.c:146
WMAVoiceContext::silence_gain
float silence_gain
set for use in blocks if ACB_TYPE_NONE
Definition: wmavoice.c:227
expf
#define expf(x)
Definition: libm.h:283
WMAVoiceContext::denoise_filter_cache_size
int denoise_filter_cache_size
samples in denoise_filter_cache
Definition: wmavoice.c:279
wmavoice_denoise_power_table
static const float wmavoice_denoise_power_table[12][64]
LUT for f(x,y) = pow((y + 6.9) / 64, 0.025 * (x + 1)).
Definition: wmavoice_data.h:3064
wmavoice_gain_codebook_acb
static const float wmavoice_gain_codebook_acb[128]
Definition: wmavoice_data.h:2874
FFCodec
Definition: codec_internal.h:127
base
uint8_t base
Definition: vp3data.h:128
t1
#define t1
Definition: regdef.h:29
max
#define max(a, b)
Definition: cuda_runtime.h:33
FFMAX
#define FFMAX(a, b)
Definition: macros.h:47
ff_celp_lp_synthesis_filterf
void ff_celp_lp_synthesis_filterf(float *out, const float *filter_coeffs, const float *in, int buffer_length, int filter_length)
LP synthesis filter.
Definition: celp_filters.c:85
WMAVoiceContext::aw_idx_is_ext
int aw_idx_is_ext
whether the AW index was encoded in 8 bits (instead of 6)
Definition: wmavoice.c:229
init_get_bits
static int init_get_bits(GetBitContext *s, const uint8_t *buffer, int bit_size)
Initialize GetBitContext.
Definition: get_bits.h:493
WMAVoiceContext::dc_level
int dc_level
Predicted amount of DC noise, based on which a DC removal filter is used.
Definition: wmavoice.c:156
wmavoice_dq_lsp16i1
static const uint8_t wmavoice_dq_lsp16i1[0x640]
Definition: wmavoice_data.h:420
WMAVoiceContext::block_conv_table
uint16_t block_conv_table[4]
boundaries for block pitch unit/scale conversion
Definition: wmavoice.c:177
frame_type_desc::log_n_blocks
uint8_t log_n_blocks
log2(n_blocks)
Definition: wmavoice.c:103
FCB_TYPE_HARDCODED
@ FCB_TYPE_HARDCODED
hardcoded (fixed) codebook with per-block gain values
Definition: wmavoice.c:88
skip_bits
static void skip_bits(GetBitContext *s, int n)
Definition: get_bits.h:371
WMAVoiceContext::aw_pulse_range
int aw_pulse_range
the range over which aw_pulse_set1() can apply the pulse, relative to the value in aw_first_pulse_off...
Definition: wmavoice.c:231
get_bits
static unsigned int get_bits(GetBitContext *s, int n)
Read 1-25 bits.
Definition: get_bits.h:325
ff_copy_bits
void ff_copy_bits(PutBitContext *pb, const uint8_t *src, int length)
Copy the content of src to the bitstream.
Definition: bitstream.c:49
FFCodec::p
AVCodec p
The public AVCodec.
Definition: codec_internal.h:131
av_ceil_log2
#define av_ceil_log2
Definition: common.h:92
WMAVoiceContext::tilted_lpcs_pf
float tilted_lpcs_pf[0x80]
aligned buffer for LPC tilting
Definition: wmavoice.c:280
ACB_TYPE_HAMMING
@ ACB_TYPE_HAMMING
Per-block pitch with signal generation using a Hamming sinc window function.
Definition: wmavoice.c:76
AMRFixed::pitch_fac
float pitch_fac
Definition: acelp_vectors.h:59
dummy
int dummy
Definition: motion.c:65
GetBitContext
Definition: get_bits.h:107
MULH
#define MULH
Definition: mathops.h:42
wmavoice_flush
static av_cold void wmavoice_flush(AVCodecContext *ctx)
Definition: wmavoice.c:329
put_bits_left
static int put_bits_left(PutBitContext *s)
Definition: put_bits.h:125
IDFT_C2R
@ IDFT_C2R
Definition: avfft.h:73
ff_rdft_end
av_cold void ff_rdft_end(RDFTContext *s)
Definition: rdft.c:117
calc_input_response
static void calc_input_response(WMAVoiceContext *s, float *lpcs, int fcb_type, float *coeffs, int remainder)
Derive denoise filter coefficients (in real domain) from the LPCs.
Definition: wmavoice.c:599
frame_type_desc::n_blocks
uint8_t n_blocks
amount of blocks per frame (each block (contains 160/n_blocks samples)
Definition: wmavoice.c:101
val
static double val(void *priv, double ch)
Definition: aeval.c:77
dequant_lsp10i
static void dequant_lsp10i(GetBitContext *gb, double *lsps)
Parse 10 independently-coded LSPs.
Definition: wmavoice.c:885
synth_block
static void synth_block(WMAVoiceContext *s, GetBitContext *gb, int block_idx, int size, int block_pitch_sh2, const double *lsps, const double *prev_lsps, const struct frame_type_desc *frame_desc, float *excitation, float *synth)
Parse data in a single block.
Definition: wmavoice.c:1434
MAX_SFRAMESIZE
#define MAX_SFRAMESIZE
maximum number of samples per superframe
Definition: wmavoice.c:55
wmavoice_gain_codebook_fcb
static const float wmavoice_gain_codebook_fcb[128]
Definition: wmavoice_data.h:2893
WMAVoiceContext::denoise_filter_cache
float denoise_filter_cache[MAX_FRAMESIZE]
Definition: wmavoice.c:278
fabsf
static __device__ float fabsf(float a)
Definition: cuda_runtime.h:181
WMAVoiceContext::sin
float sin[511]
Definition: wmavoice.c:270
a1
#define a1
Definition: regdef.h:47
AV_CODEC_ID_WMAVOICE
@ AV_CODEC_ID_WMAVOICE
Definition: codec_id.h:474
lrint
#define lrint
Definition: tablegen.h:53
MUL16
#define MUL16(ra, rb)
Definition: mathops.h:89
ff_thread_once
static int ff_thread_once(char *control, void(*routine)(void))
Definition: thread.h:184
AV_LOG_ERROR
#define AV_LOG_ERROR
Something went wrong and cannot losslessly be recovered.
Definition: log.h:180
FF_ARRAY_ELEMS
#define FF_ARRAY_ELEMS(a)
Definition: sinewin_tablegen.c:29
av_cold
#define av_cold
Definition: attributes.h:90
dct.h
init_get_bits8
static int init_get_bits8(GetBitContext *s, const uint8_t *buffer, int byte_size)
Initialize GetBitContext.
Definition: get_bits.h:524
FCB_TYPE_EXC_PULSES
@ FCB_TYPE_EXC_PULSES
Innovation (fixed) codebook pulse sets in combinations of either single pulses or pulse pairs.
Definition: wmavoice.c:92
MAX_LSPS_ALIGN16
#define MAX_LSPS_ALIGN16
same as MAX_LSPS; needs to be multiple
Definition: wmavoice.c:50
av_memcpy_backptr
void av_memcpy_backptr(uint8_t *dst, int back, int cnt)
Overlapping memcpy() implementation.
Definition: mem.c:445
float
float
Definition: af_crystalizer.c:122
wmavoice_dq_lsp10r
static const uint8_t wmavoice_dq_lsp10r[0x1400]
Definition: wmavoice_data.h:749
FF_CODEC_DECODE_CB
#define FF_CODEC_DECODE_CB(func)
Definition: codec_internal.h:306
WMAVoiceContext::sframe_cache_size
int sframe_cache_size
set to >0 if we have data from an (incomplete) superframe from a previous packet that spilled over in...
Definition: wmavoice.c:205
s
#define s(width, name)
Definition: cbs_vp9.c:256
WMAVoiceContext::lsp_q_mode
int lsp_q_mode
defines quantizer defaults [0, 1]
Definition: wmavoice.c:160
frame_type_desc::fcb_type
uint8_t fcb_type
Fixed codebook type (FCB_TYPE_*)
Definition: wmavoice.c:105
log_range
#define log_range(var, assign)
WMAVoiceContext::prev_lsps
double prev_lsps[MAX_LSPS]
LSPs of the last frame of the previous superframe.
Definition: wmavoice.c:221
AVMEDIA_TYPE_AUDIO
@ AVMEDIA_TYPE_AUDIO
Definition: avutil.h:202
WMAVoiceContext::aw_n_pulses
int aw_n_pulses[2]
number of AW-pulses in each block; note that this number can be negative (in which case it basically ...
Definition: wmavoice.c:237
AMRFixed
Sparse representation for the algebraic codebook (fixed) vector.
Definition: acelp_vectors.h:53
init
int(* init)(AVBSFContext *ctx)
Definition: dts2pts_bsf.c:365
bits
uint8_t bits
Definition: vp3data.h:128
adaptive_gain_control
static void adaptive_gain_control(float *out, const float *in, const float *speech_synth, int size, float alpha, float *gain_mem)
Adaptive gain control (as used in postfilter).
Definition: wmavoice.c:499
av_assert0
#define av_assert0(cond)
assert() equivalent, that is always enabled.
Definition: avassert.h:37
wmavoice_lsp16_intercoeff_a
static const float wmavoice_lsp16_intercoeff_a[32][2][16]
Definition: wmavoice_data.h:2047
ctx
AVFormatContext * ctx
Definition: movenc.c:48
decode.h
get_bits.h
wmavoice_mean_lsf10
static const double wmavoice_mean_lsf10[2][10]
Definition: wmavoice_data.h:2565
WMAVoiceContext::spillover_nbits
int spillover_nbits
number of bits of the previous packet's last superframe preceding this packet's first full superframe...
Definition: wmavoice.c:189
UMULH
static av_always_inline unsigned UMULH(unsigned a, unsigned b)
Definition: mathops.h:69
AMRFixed::y
float y[10]
Definition: acelp_vectors.h:56
wmavoice_gain_silence
static const float wmavoice_gain_silence[256]
Definition: wmavoice_data.h:2788
FCB_TYPE_AW_PULSES
@ FCB_TYPE_AW_PULSES
Pitch-adaptive window (AW) pulse signals, used in particular for low-bitrate streams.
Definition: wmavoice.c:90
PutBitContext
Definition: put_bits.h:50
WMAVoiceContext::vbm_tree
int8_t vbm_tree[25]
converts VLC codes to frame type
Definition: wmavoice.c:141
CODEC_LONG_NAME
#define CODEC_LONG_NAME(str)
Definition: codec_internal.h:272
mul
static float mul(float src0, float src1)
Definition: dnn_backend_native_layer_mathbinary.c:39
wmavoice_dq_lsp16i3
static const uint8_t wmavoice_dq_lsp16i3[0x300]
Definition: wmavoice_data.h:682
if
if(ret)
Definition: filter_design.txt:179
AMRFixed::no_repeat_mask
int no_repeat_mask
Definition: acelp_vectors.h:57
postfilter
static void postfilter(WMAVoiceContext *s, const float *synth, float *samples, int size, const float *lpcs, float *zero_exc_pf, int fcb_type, int pitch)
Averaging projection filter, the postfilter used in WMAVoice.
Definition: wmavoice.c:800
AV_ONCE_INIT
#define AV_ONCE_INIT
Definition: thread.h:182
NULL
#define NULL
Definition: coverity.c:32
sizes
static const int sizes[][2]
Definition: img2dec.c:57
WMAVoiceContext::history_nsamples
int history_nsamples
number of samples in history for signal prediction (through ACB)
Definition: wmavoice.c:146
WMAVoiceContext::synth_history
float synth_history[MAX_LSPS]
see excitation_history
Definition: wmavoice.c:256
AVERROR_PATCHWELCOME
#define AVERROR_PATCHWELCOME
Not yet implemented in FFmpeg, patches welcome.
Definition: error.h:64
last_coeff
static const uint8_t last_coeff[3]
Definition: qdm2data.h:187
INIT_VLC_STATIC_FROM_LENGTHS
#define INIT_VLC_STATIC_FROM_LENGTHS(vlc, bits, nb_codes, lens, len_wrap, symbols, symbols_wrap, symbols_size, offset, flags, static_size)
Definition: vlc.h:131
WMAVoiceContext::denoise_strength
int denoise_strength
strength of denoising in Wiener filter [0-11]
Definition: wmavoice.c:152
MAX_SIGNAL_HISTORY
#define MAX_SIGNAL_HISTORY
maximum excitation signal history
Definition: wmavoice.c:54
WMAVoiceContext::sframe_cache
uint8_t sframe_cache[SFRAME_CACHE_MAXSIZE+AV_INPUT_BUFFER_PADDING_SIZE]
cache for superframe data split over multiple packets
Definition: wmavoice.c:202
get_bits1
static unsigned int get_bits1(GetBitContext *s)
Definition: get_bits.h:378
dequant_lsp10r
static void dequant_lsp10r(GetBitContext *gb, double *i_lsps, const double *old, double *a1, double *a2, int q_mode)
Parse 10 independently-coded LSPs, and then derive the tables to generate LSPs for the other frames f...
Definition: wmavoice.c:911
WMAVoiceContext::pitch_nbits
int pitch_nbits
number of bits used to specify the pitch value in the frame header
Definition: wmavoice.c:166
WMAVoiceContext::block_delta_pitch_nbits
int block_delta_pitch_nbits
number of bits used to specify the delta pitch between this and the last block's pitch value,...
Definition: wmavoice.c:171
kalman_smoothen
static int kalman_smoothen(WMAVoiceContext *s, int pitch, const float *in, float *out, int size)
Kalman smoothing function.
Definition: wmavoice.c:540
WMAVoiceContext::denoise_coeffs_pf
float denoise_coeffs_pf[0x80]
aligned buffer for denoise coefficients
Definition: wmavoice.c:282
WMAVoiceContext::skip_bits_next
int skip_bits_next
number of bits to skip at the next call to wmavoice_decode_packet() (since they're part of the previo...
Definition: wmavoice.c:198
DFT_R2C
@ DFT_R2C
Definition: avfft.h:72
sqrtf
static __device__ float sqrtf(float a)
Definition: cuda_runtime.h:184
abs
#define abs(x)
Definition: cuda_runtime.h:35
celp_filters.h
MAX_FRAMESIZE
#define MAX_FRAMESIZE
maximum number of samples per frame
Definition: wmavoice.c:53
av_clipf
av_clipf
Definition: af_crystalizer.c:122
MAX_FRAMES
#define MAX_FRAMES
maximum number of frames per superframe
Definition: wmavoice.c:52
get_vlc2
static av_always_inline int get_vlc2(GetBitContext *s, const VLCElem *table, int bits, int max_depth)
Parse a vlc code.
Definition: get_bits.h:631
decode_vbmtree
static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25])
Set up the variable bit mode (VBM) tree from container extradata.
Definition: wmavoice.c:301
AVOnce
#define AVOnce
Definition: thread.h:181
aw_parse_coords
static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb, const int *pitch)
Parse the offset of the first pitch-adaptive window pulses, and the distribution of pulses between th...
Definition: wmavoice.c:1030
wmavoice_init_static_data
static av_cold void wmavoice_init_static_data(void)
Definition: wmavoice.c:315
float_dsp.h
WMAVoiceContext::dcf_mem
float dcf_mem[2]
DC filter history.
Definition: wmavoice.c:274
ff_get_buffer
int ff_get_buffer(AVCodecContext *avctx, AVFrame *frame, int flags)
Get a buffer for a frame.
Definition: decode.c:1473
AV_CODEC_CAP_DR1
#define AV_CODEC_CAP_DR1
Codec uses get_buffer() or get_encode_buffer() for allocating buffers and supports custom allocators.
Definition: codec.h:52
parse_packet_header
static int parse_packet_header(WMAVoiceContext *s)
Parse the packet header at the start of each packet (input data to this decoder).
Definition: wmavoice.c:1834
AVPacket::size
int size
Definition: packet.h:375
powf
#define powf(x, y)
Definition: libm.h:50
AVChannelLayout
An AVChannelLayout holds information about the channel layout of audio data.
Definition: channel_layout.h:301
codec_internal.h
DECLARE_ALIGNED
#define DECLARE_ALIGNED(n, t, v)
Definition: mem_internal.h:87
frame_type_vlc
static VLC frame_type_vlc
Frame type VLC coding.
Definition: wmavoice.c:64
WMAVoiceContext::spillover_bitsize
int spillover_bitsize
number of bits used to specify spillover_nbits in the packet header = ceil(log2(ctx->block_align << 3...
Definition: wmavoice.c:143
for
for(k=2;k<=8;++k)
Definition: h264pred_template.c:425
WMAVoiceContext::pb
PutBitContext pb
bitstream writer for sframe_cache
Definition: wmavoice.c:210
WMAVoiceContext::last_pitch_val
int last_pitch_val
pitch value of the previous frame
Definition: wmavoice.c:223
size
int size
Definition: twinvq_data.h:10344
wiener_denoise
static void wiener_denoise(WMAVoiceContext *s, int fcb_type, float *synth_pf, int size, const float *lpcs)
This function applies a Wiener filter on the (noisy) speech signal as a means to denoise it.
Definition: wmavoice.c:718
wmavoice_lsp10_intercoeff_b
static const float wmavoice_lsp10_intercoeff_b[32][2][10]
Definition: wmavoice_data.h:1852
WMAVoiceContext::dct
DCTContext dct
Definition: wmavoice.c:268
dequant_lsp16i
static void dequant_lsp16i(GetBitContext *gb, double *lsps)
Parse 16 independently-coded LSPs.
Definition: wmavoice.c:947
wmavoice_dq_lsp16r1
static const uint8_t wmavoice_dq_lsp16r1[0x500]
Definition: wmavoice_data.h:1264
WMAVoiceContext::aw_first_pulse_off
int aw_first_pulse_off[2]
index of first sample to which to apply AW-pulses, or -0xff if unset
Definition: wmavoice.c:240
WMAVoiceContext::zero_exc_pf
float zero_exc_pf[MAX_SIGNAL_HISTORY+MAX_SFRAMESIZE]
zero filter output (i.e.
Definition: wmavoice.c:275
sinewin.h
wmavoice_dq_lsp16r2
static const uint8_t wmavoice_dq_lsp16r2[0x500]
Definition: wmavoice_data.h:1395
offset
it s the only field you need to keep assuming you have a context There is some magic you don t need to care about around this just let it vf offset
Definition: writing_filters.txt:86
ACB_TYPE_ASYMMETRIC
@ ACB_TYPE_ASYMMETRIC
adaptive codebook with per-frame pitch, which we interpolate to get a per-sample pitch.
Definition: wmavoice.c:71
frame_type_desc
Description of frame types.
Definition: wmavoice.c:100
WMAVoiceContext::block_pitch_range
int block_pitch_range
range of the block pitch
Definition: wmavoice.c:170
stabilize_lsps
static void stabilize_lsps(double *lsps, int num)
Ensure minimum value for first item, maximum value for last value, proper spacing between each value ...
Definition: wmavoice.c:1652
ff_dct_end
av_cold void ff_dct_end(DCTContext *s)
Definition: dct.c:224
M_PI
#define M_PI
Definition: mathematics.h:52
ff_tilt_compensation
void ff_tilt_compensation(float *mem, float tilt, float *samples, int size)
Apply tilt compensation filter, 1 - tilt * z-1.
Definition: acelp_filters.c:138
wmavoice_energy_table
static const float wmavoice_energy_table[128]
LUT for 1.071575641632 * pow(1.0331663, n - 127)
Definition: wmavoice_data.h:3026
ff_sine_window_init
void ff_sine_window_init(float *window, int n)
Generate a sine window.
Definition: sinewin_tablegen.h:59
wmavoice_decode_init
static av_cold int wmavoice_decode_init(AVCodecContext *ctx)
Set up decoder with parameters from demuxer (extradata etc.).
Definition: wmavoice.c:360
WMAVoiceContext::block_delta_pitch_hrange
int block_delta_pitch_hrange
1/2 range of the delta (full range is from -this to +this-1)
Definition: wmavoice.c:175
DST_I
@ DST_I
Definition: avfft.h:97
wmavoice_ipol2_coeffs
static const float wmavoice_ipol2_coeffs[32]
Hamming-window sinc function (num = 32, x = [ 0, 31 ]): (0.54 + 0.46 * cos(2 * M_PI * x / (num - 1)))...
Definition: wmavoice_data.h:3012
ff_rdft_init
av_cold int ff_rdft_init(RDFTContext *s, int nbits, enum RDFTransformType trans)
Set up a real FFT.
Definition: rdft.c:89
WMAVoiceContext::pitch_diff_sh16
int pitch_diff_sh16
((cur_pitch_val - last_pitch_val) << 16) / MAX_FRAMESIZE
Definition: wmavoice.c:225
WMAVoiceContext::gain_pred_err
float gain_pred_err[6]
cache for gain prediction
Definition: wmavoice.c:251
ff_dct_init
av_cold int ff_dct_init(DCTContext *s, int nbits, enum DCTTransformType inverse)
Set up DCT.
Definition: dct.c:179
i
#define i(width, name, range_min, range_max)
Definition: cbs_h2645.c:269
WMAVoiceContext::nb_superframes
int nb_superframes
number of superframes in current packet
Definition: wmavoice.c:250
t3
#define t3
Definition: regdef.h:31
WMAVoiceContext::cos
float cos[511]
8-bit cosine/sine windows over [-pi,pi] range
Definition: wmavoice.c:270
RDFTContext
Definition: rdft.h:28
a2
#define a2
Definition: regdef.h:48
WMAVoiceContext::denoise_tilt_corr
int denoise_tilt_corr
Whether to apply tilt correction to the Wiener filter coefficients (postfilter)
Definition: wmavoice.c:154
delta
float delta
Definition: vorbis_enc_data.h:430
wmavoice_lsp16_intercoeff_b
static const float wmavoice_lsp16_intercoeff_b[32][2][16]
Definition: wmavoice_data.h:2306
FFMIN
#define FFMIN(a, b)
Definition: macros.h:49
av_frame_unref
void av_frame_unref(AVFrame *frame)
Unreference all the buffers referenced by frame and reset the frame fields.
Definition: frame.c:478
acelp_filters.h
ff_weighted_vector_sumf
void ff_weighted_vector_sumf(float *out, const float *in_a, const float *in_b, float weight_coeff_a, float weight_coeff_b, int length)
float implementation of weighted sum of two vectors.
Definition: acelp_vectors.c:182
DCTContext
Definition: dct.h:32
WMAVoiceContext::lsp_def_mode
int lsp_def_mode
defines different sets of LSP defaults [0, 1]
Definition: wmavoice.c:161
wmavoice_gain_universal
static const float wmavoice_gain_universal[64]
Definition: wmavoice_data.h:2855
AVCodec::name
const char * name
Name of the codec implementation.
Definition: codec.h:191
len
int len
Definition: vorbis_enc_data.h:426
WMAVoiceContext::synth_filter_out_buf
float synth_filter_out_buf[0x80+MAX_LSPS_ALIGN16]
aligned buffer for postfilter speech synthesis
Definition: wmavoice.c:284
tilt_factor
static float tilt_factor(const float *lpcs, int n_lpcs)
Get the tilt factor of a formant filter from its transfer function.
Definition: wmavoice.c:586
WMAVoiceContext::rdft
RDFTContext rdft
Definition: wmavoice.c:266
VLC_NBITS
#define VLC_NBITS
number of bits to read per VLC iteration
Definition: wmavoice.c:59
wmavoice_data.h
Windows Media Voice (WMAVoice) tables.
avcodec.h
WMAVoiceContext::min_pitch_val
int min_pitch_val
base value for pitch parsing code
Definition: wmavoice.c:164
WMAVoiceContext::last_acb_type
int last_acb_type
frame type [0-2] of the previous frame
Definition: wmavoice.c:224
av_uninit
#define av_uninit(x)
Definition: attributes.h:154
ret
ret
Definition: filter_design.txt:187
frame
these buffered frames must be flushed immediately if a new input produces new the filter must not call request_frame to get more It must just process the frame or queue it The task of requesting more frames is left to the filter s request_frame method or the application If a filter has several the filter must be ready for frames arriving randomly on any input any filter with several inputs will most likely require some kind of queuing mechanism It is perfectly acceptable to have a limited queue and to drop frames when the inputs are too unbalanced request_frame For filters that do not use the this method is called when a frame is wanted on an output For a it should directly call filter_frame on the corresponding output For a if there are queued frames already one of these frames should be pushed If the filter should request a frame on one of its repeatedly until at least one frame has been pushed Return or at least make progress towards producing a frame
Definition: filter_design.txt:264
lsp.h
ff_celp_lp_zero_synthesis_filterf
void ff_celp_lp_zero_synthesis_filterf(float *out, const float *filter_coeffs, const float *in, int buffer_length, int filter_length)
LP zero synthesis filter.
Definition: celp_filters.c:200
WMAVoiceContext::do_apf
int do_apf
whether to apply the averaged projection filter (APF)
Definition: wmavoice.c:150
pos
unsigned int pos
Definition: spdifenc.c:413
AMRFixed::n
int n
Definition: acelp_vectors.h:54
wmavoice_dq_lsp16i2
static const uint8_t wmavoice_dq_lsp16i2[0x3c0]
Definition: wmavoice_data.h:583
AV_INPUT_BUFFER_PADDING_SIZE
#define AV_INPUT_BUFFER_PADDING_SIZE
Definition: defs.h:40
wmavoice_mean_lsf16
static const double wmavoice_mean_lsf16[2][16]
Definition: wmavoice_data.h:2574
AV_RL32
uint64_t_TMPL AV_WL64 unsigned int_TMPL AV_RL32
Definition: bytestream.h:92
U
#define U(x)
Definition: vpx_arith.h:37
wmavoice_decode_end
static av_cold int wmavoice_decode_end(AVCodecContext *ctx)
Definition: wmavoice.c:1984
WMAVoiceContext::lsps
int lsps
number of LSPs per frame [10 or 16]
Definition: wmavoice.c:159
AVCodecContext
main external API structure.
Definition: avcodec.h:426
wmavoice_decode_packet
static int wmavoice_decode_packet(AVCodecContext *ctx, AVFrame *frame, int *got_frame_ptr, AVPacket *avpkt)
Packet decoding: a packet is anything that the (ASF) demuxer contains, and we expect that the demuxer...
Definition: wmavoice.c:1898
channel_layout.h
t2
#define t2
Definition: regdef.h:30
WMAVoiceContext::dst
DCTContext dst
contexts for phase shift (in Hilbert transform, part of postfilter)
Definition: wmavoice.c:268
WMAVoiceContext::block_pitch_nbits
int block_pitch_nbits
number of bits used to specify the first block's pitch value
Definition: wmavoice.c:168
VLC
Definition: vlc.h:31
synth_superframe
static int synth_superframe(AVCodecContext *ctx, AVFrame *frame, int *got_frame_ptr)
Synthesize output samples for a single superframe.
Definition: wmavoice.c:1698
av_channel_layout_uninit
void av_channel_layout_uninit(AVChannelLayout *channel_layout)
Free any allocated data in the channel layout and reset the channel count to 0.
Definition: channel_layout.c:632
WMAVoiceContext::frame_cntr
int frame_cntr
current frame index [0 - 0xFFFE]; is only used for comfort noise in pRNG()
Definition: wmavoice.c:248
wmavoice_ipol1_coeffs
static const float wmavoice_ipol1_coeffs[17 *9]
Definition: wmavoice_data.h:2960
values
these buffered frames must be flushed immediately if a new input produces new the filter must not call request_frame to get more It must just process the frame or queue it The task of requesting more frames is left to the filter s request_frame method or the application If a filter has several the filter must be ready for frames arriving randomly on any input any filter with several inputs will most likely require some kind of queuing mechanism It is perfectly acceptable to have a limited queue and to drop frames when the inputs are too unbalanced request_frame For filters that do not use the this method is called when a frame is wanted on an output For a it should directly call filter_frame on the corresponding output For a if there are queued frames already one of these frames should be pushed If the filter should request a frame on one of its repeatedly until at least one frame has been pushed Return values
Definition: filter_design.txt:263
ff_set_fixed_vector
void ff_set_fixed_vector(float *out, const AMRFixed *in, float scale, int size)
Add fixed vector to an array from a sparse representation.
Definition: acelp_vectors.c:224
mean_lsf
static const float mean_lsf[10]
Definition: siprdata.h:27
AV_CODEC_CAP_DELAY
#define AV_CODEC_CAP_DELAY
Encoder or decoder requires flushing with NULL input at the end in order to give the complete and cor...
Definition: codec.h:76
VLC::table
VLCElem * table
Definition: vlc.h:33
samples
Filter the word “frame” indicates either a video frame or a group of audio samples
Definition: filter_design.txt:8
copy_bits
static void copy_bits(PutBitContext *pb, const uint8_t *data, int size, GetBitContext *gb, int nbits)
Copy (unaligned) bits from gb/data/size to pb.
Definition: wmavoice.c:1869
avpriv_scalarproduct_float_c
float avpriv_scalarproduct_float_c(const float *v1, const float *v2, int len)
Return the scalar product of two vectors.
Definition: float_dsp.c:124
synth_frame
static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx, float *samples, const double *lsps, const double *prev_lsps, float *excitation, float *synth)
Synthesize output samples for a single frame.
Definition: wmavoice.c:1476
AV_CODEC_CAP_SUBFRAMES
#define AV_CODEC_CAP_SUBFRAMES
Codec can output multiple frames per AVPacket Normally demuxers return one frame at a time,...
Definition: codec.h:94
M_LN10
#define M_LN10
Definition: mathematics.h:43
WMAVoiceContext::gb
GetBitContext gb
packet bitreader.
Definition: wmavoice.c:137
avpriv_request_sample
#define avpriv_request_sample(...)
Definition: tableprint_vlc.h:36
synth_block_fcb_acb
static void synth_block_fcb_acb(WMAVoiceContext *s, GetBitContext *gb, int block_idx, int size, int block_pitch_sh2, const struct frame_type_desc *frame_desc, float *excitation)
Parse FCB/ACB signal for a single block.
Definition: wmavoice.c:1300
flush_put_bits
static void flush_put_bits(PutBitContext *s)
Pad the end of the output stream with zeros.
Definition: put_bits.h:143
AV_CHANNEL_LAYOUT_MONO
#define AV_CHANNEL_LAYOUT_MONO
Definition: channel_layout.h:368
alpha
static const int16_t alpha[]
Definition: ilbcdata.h:55
AVPacket
This structure stores compressed data.
Definition: packet.h:351
synth_block_hardcoded
static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb, int block_idx, int size, const struct frame_type_desc *frame_desc, float *excitation)
Parse hardcoded signal for a single block.
Definition: wmavoice.c:1269
SFRAME_CACHE_MAXSIZE
#define SFRAME_CACHE_MAXSIZE
maximum cache size for frame data that
Definition: wmavoice.c:57
AMRFixed::pitch_lag
int pitch_lag
Definition: acelp_vectors.h:58
flags
#define flags(name, subs,...)
Definition: cbs_av1.c:561
av_log
#define av_log(a,...)
Definition: tableprint_vlc.h:27
wmavoice_lsp10_intercoeff_a
static const float wmavoice_lsp10_intercoeff_a[32][2][10]
Definition: wmavoice_data.h:1657
AVERROR_INVALIDDATA
#define AVERROR_INVALIDDATA
Invalid data found when processing input.
Definition: error.h:61
dequant_lsp16r
static void dequant_lsp16r(GetBitContext *gb, double *i_lsps, const double *old, double *a1, double *a2, int q_mode)
Parse 16 independently-coded LSPs, and then derive the tables to generate LSPs for the other frames f...
Definition: wmavoice.c:980
frame_type_desc::dbl_pulses
uint8_t dbl_pulses
how many pulse vectors have pulse pairs (rather than just one single pulse) only if fcb_type == FCB_T...
Definition: wmavoice.c:106
frame_type_desc::acb_type
uint8_t acb_type
Adaptive codebook type (ACB_TYPE_*)
Definition: wmavoice.c:104
MAX_BLOCKS
#define MAX_BLOCKS
maximum number of blocks per frame
Definition: wmavoice.c:48
ff_acelp_lspd2lpc
void ff_acelp_lspd2lpc(const double *lsp, float *lpc, int lp_half_order)
Reconstruct LPC coefficients from the line spectral pair frequencies.
Definition: lsp.c:220
WMAVoiceContext::postfilter_agc
float postfilter_agc
gain control memory, used in adaptive_gain_control()
Definition: wmavoice.c:272
put_bits.h
pulses
static const int8_t pulses[4]
Number of non-zero pulses in the MP-MLQ excitation.
Definition: g723_1.h:260
ff_acelp_interpolatef
void ff_acelp_interpolatef(float *out, const float *in, const float *filter_coeffs, int precision, int frac_pos, int filter_length, int length)
Floating point version of ff_acelp_interpolate()
Definition: acelp_filters.c:80
AVFormatContext::priv_data
void * priv_data
Format private data.
Definition: avformat.h:1132
AV_SAMPLE_FMT_FLT
@ AV_SAMPLE_FMT_FLT
float
Definition: samplefmt.h:60
min
float min
Definition: vorbis_enc_data.h:429
WMAVoiceContext
WMA Voice decoding context.
Definition: wmavoice.c:132