FFmpeg
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
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 
30 #include "libavutil/channel_layout.h"
31 #include "libavutil/float_dsp.h"
32 #include "libavutil/mem.h"
33 #include "avcodec.h"
34 #include "internal.h"
35 #include "get_bits.h"
36 #include "put_bits.h"
37 #include "wmavoice_data.h"
38 #include "celp_filters.h"
39 #include "acelp_vectors.h"
40 #include "acelp_filters.h"
41 #include "lsp.h"
42 #include "dct.h"
43 #include "rdft.h"
44 #include "sinewin.h"
45 
46 #define MAX_BLOCKS 8 ///< maximum number of blocks per frame
47 #define MAX_LSPS 16 ///< maximum filter order
48 #define MAX_LSPS_ALIGN16 16 ///< same as #MAX_LSPS; needs to be multiple
49  ///< of 16 for ASM input buffer alignment
50 #define MAX_FRAMES 3 ///< maximum number of frames per superframe
51 #define MAX_FRAMESIZE 160 ///< maximum number of samples per frame
52 #define MAX_SIGNAL_HISTORY 416 ///< maximum excitation signal history
53 #define MAX_SFRAMESIZE (MAX_FRAMESIZE * MAX_FRAMES)
54  ///< maximum number of samples per superframe
55 #define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that
56  ///< was split over two packets
57 #define VLC_NBITS 6 ///< number of bits to read per VLC iteration
58 
59 /**
60  * Frame type VLC coding.
61  */
62 static VLC frame_type_vlc;
63 
64 /**
65  * Adaptive codebook types.
66  */
67 enum {
68  ACB_TYPE_NONE = 0, ///< no adaptive codebook (only hardcoded fixed)
69  ACB_TYPE_ASYMMETRIC = 1, ///< adaptive codebook with per-frame pitch, which
70  ///< we interpolate to get a per-sample pitch.
71  ///< Signal is generated using an asymmetric sinc
72  ///< window function
73  ///< @note see #wmavoice_ipol1_coeffs
74  ACB_TYPE_HAMMING = 2 ///< Per-block pitch with signal generation using
75  ///< a Hamming sinc window function
76  ///< @note see #wmavoice_ipol2_coeffs
77 };
78 
79 /**
80  * Fixed codebook types.
81  */
82 enum {
83  FCB_TYPE_SILENCE = 0, ///< comfort noise during silence
84  ///< generated from a hardcoded (fixed) codebook
85  ///< with per-frame (low) gain values
86  FCB_TYPE_HARDCODED = 1, ///< hardcoded (fixed) codebook with per-block
87  ///< gain values
88  FCB_TYPE_AW_PULSES = 2, ///< Pitch-adaptive window (AW) pulse signals,
89  ///< used in particular for low-bitrate streams
90  FCB_TYPE_EXC_PULSES = 3, ///< Innovation (fixed) codebook pulse sets in
91  ///< combinations of either single pulses or
92  ///< pulse pairs
93 };
94 
95 /**
96  * Description of frame types.
97  */
98 static const struct frame_type_desc {
99  uint8_t n_blocks; ///< amount of blocks per frame (each block
100  ///< (contains 160/#n_blocks samples)
101  uint8_t log_n_blocks; ///< log2(#n_blocks)
102  uint8_t acb_type; ///< Adaptive codebook type (ACB_TYPE_*)
103  uint8_t fcb_type; ///< Fixed codebook type (FCB_TYPE_*)
104  uint8_t dbl_pulses; ///< how many pulse vectors have pulse pairs
105  ///< (rather than just one single pulse)
106  ///< only if #fcb_type == #FCB_TYPE_EXC_PULSES
107  uint16_t frame_size; ///< the amount of bits that make up the block
108  ///< data (per frame)
109 } frame_descs[17] = {
110  { 1, 0, ACB_TYPE_NONE, FCB_TYPE_SILENCE, 0, 0 },
111  { 2, 1, ACB_TYPE_NONE, FCB_TYPE_HARDCODED, 0, 28 },
112  { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_AW_PULSES, 0, 46 },
113  { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 80 },
114  { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 104 },
115  { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 0, 108 },
116  { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 132 },
117  { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 168 },
118  { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 64 },
119  { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 80 },
120  { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 104 },
121  { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 108 },
122  { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 132 },
123  { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 168 },
124  { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 176 },
125  { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 208 },
126  { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 256 }
127 };
128 
129 /**
130  * WMA Voice decoding context.
131  */
132 typedef struct {
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  int frame_lsp_bitsize; ///< size (in bits) of LSPs, when encoded
164  ///< per-frame (independent coding)
165  int sframe_lsp_bitsize; ///< size (in bits) of LSPs, when encoded
166  ///< per superframe (residual coding)
167 
168  int min_pitch_val; ///< base value for pitch parsing code
169  int max_pitch_val; ///< max value + 1 for pitch parsing
170  int pitch_nbits; ///< number of bits used to specify the
171  ///< pitch value in the frame header
172  int block_pitch_nbits; ///< number of bits used to specify the
173  ///< first block's pitch value
174  int block_pitch_range; ///< range of the block pitch
175  int block_delta_pitch_nbits; ///< number of bits used to specify the
176  ///< delta pitch between this and the last
177  ///< block's pitch value, used in all but
178  ///< first block
179  int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is
180  ///< from -this to +this-1)
181  uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale
182  ///< conversion
183 
184  /**
185  * @}
186  *
187  * @name Packet values specified in the packet header or related to a packet.
188  *
189  * A packet is considered to be a single unit of data provided to this
190  * decoder by the demuxer.
191  * @{
192  */
193  int spillover_nbits; ///< number of bits of the previous packet's
194  ///< last superframe preceding this
195  ///< packet's first full superframe (useful
196  ///< for re-synchronization also)
197  int has_residual_lsps; ///< if set, superframes contain one set of
198  ///< LSPs that cover all frames, encoded as
199  ///< independent and residual LSPs; if not
200  ///< set, each frame contains its own, fully
201  ///< independent, LSPs
202  int skip_bits_next; ///< number of bits to skip at the next call
203  ///< to #wmavoice_decode_packet() (since
204  ///< they're part of the previous superframe)
205 
206  uint8_t sframe_cache[SFRAME_CACHE_MAXSIZE + FF_INPUT_BUFFER_PADDING_SIZE];
207  ///< cache for superframe data split over
208  ///< multiple packets
209  int sframe_cache_size; ///< set to >0 if we have data from an
210  ///< (incomplete) superframe from a previous
211  ///< packet that spilled over in the current
212  ///< packet; specifies the amount of bits in
213  ///< #sframe_cache
214  PutBitContext pb; ///< bitstream writer for #sframe_cache
215 
216  /**
217  * @}
218  *
219  * @name Frame and superframe values
220  * Superframe and frame data - these can change from frame to frame,
221  * although some of them do in that case serve as a cache / history for
222  * the next frame or superframe.
223  * @{
224  */
225  double prev_lsps[MAX_LSPS]; ///< LSPs of the last frame of the previous
226  ///< superframe
227  int last_pitch_val; ///< pitch value of the previous frame
228  int last_acb_type; ///< frame type [0-2] of the previous frame
229  int pitch_diff_sh16; ///< ((cur_pitch_val - #last_pitch_val)
230  ///< << 16) / #MAX_FRAMESIZE
231  float silence_gain; ///< set for use in blocks if #ACB_TYPE_NONE
232 
233  int aw_idx_is_ext; ///< whether the AW index was encoded in
234  ///< 8 bits (instead of 6)
235  int aw_pulse_range; ///< the range over which #aw_pulse_set1()
236  ///< can apply the pulse, relative to the
237  ///< value in aw_first_pulse_off. The exact
238  ///< position of the first AW-pulse is within
239  ///< [pulse_off, pulse_off + this], and
240  ///< depends on bitstream values; [16 or 24]
241  int aw_n_pulses[2]; ///< number of AW-pulses in each block; note
242  ///< that this number can be negative (in
243  ///< which case it basically means "zero")
244  int aw_first_pulse_off[2]; ///< index of first sample to which to
245  ///< apply AW-pulses, or -0xff if unset
246  int aw_next_pulse_off_cache; ///< the position (relative to start of the
247  ///< second block) at which pulses should
248  ///< start to be positioned, serves as a
249  ///< cache for pitch-adaptive window pulses
250  ///< between blocks
251 
252  int frame_cntr; ///< current frame index [0 - 0xFFFE]; is
253  ///< only used for comfort noise in #pRNG()
254  float gain_pred_err[6]; ///< cache for gain prediction
255  float excitation_history[MAX_SIGNAL_HISTORY];
256  ///< cache of the signal of previous
257  ///< superframes, used as a history for
258  ///< signal generation
259  float synth_history[MAX_LSPS]; ///< see #excitation_history
260  /**
261  * @}
262  *
263  * @name Postfilter values
264  *
265  * Variables used for postfilter implementation, mostly history for
266  * smoothing and so on, and context variables for FFT/iFFT.
267  * @{
268  */
269  RDFTContext rdft, irdft; ///< contexts for FFT-calculation in the
270  ///< postfilter (for denoise filter)
271  DCTContext dct, dst; ///< contexts for phase shift (in Hilbert
272  ///< transform, part of postfilter)
273  float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi]
274  ///< range
275  float postfilter_agc; ///< gain control memory, used in
276  ///< #adaptive_gain_control()
277  float dcf_mem[2]; ///< DC filter history
278  float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE];
279  ///< zero filter output (i.e. excitation)
280  ///< by postfilter
281  float denoise_filter_cache[MAX_FRAMESIZE];
282  int denoise_filter_cache_size; ///< samples in #denoise_filter_cache
283  DECLARE_ALIGNED(32, float, tilted_lpcs_pf)[0x80];
284  ///< aligned buffer for LPC tilting
285  DECLARE_ALIGNED(32, float, denoise_coeffs_pf)[0x80];
286  ///< aligned buffer for denoise coefficients
287  DECLARE_ALIGNED(32, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16];
288  ///< aligned buffer for postfilter speech
289  ///< synthesis
290  /**
291  * @}
292  */
293 } WMAVoiceContext;
294 
295 /**
296  * Set up the variable bit mode (VBM) tree from container extradata.
297  * @param gb bit I/O context.
298  * The bit context (s->gb) should be loaded with byte 23-46 of the
299  * container extradata (i.e. the ones containing the VBM tree).
300  * @param vbm_tree pointer to array to which the decoded VBM tree will be
301  * written.
302  * @return 0 on success, <0 on error.
303  */
304 static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25])
305 {
306  int cntr[8] = { 0 }, n, res;
307 
308  memset(vbm_tree, 0xff, sizeof(vbm_tree[0]) * 25);
309  for (n = 0; n < 17; n++) {
310  res = get_bits(gb, 3);
311  if (cntr[res] > 3) // should be >= 3 + (res == 7))
312  return -1;
313  vbm_tree[res * 3 + cntr[res]++] = n;
314  }
315  return 0;
316 }
317 
318 static av_cold void wmavoice_init_static_data(AVCodec *codec)
319 {
320  static const uint8_t bits[] = {
321  2, 2, 2, 4, 4, 4,
322  6, 6, 6, 8, 8, 8,
323  10, 10, 10, 12, 12, 12,
324  14, 14, 14, 14
325  };
326  static const uint16_t codes[] = {
327  0x0000, 0x0001, 0x0002, // 00/01/10
328  0x000c, 0x000d, 0x000e, // 11+00/01/10
329  0x003c, 0x003d, 0x003e, // 1111+00/01/10
330  0x00fc, 0x00fd, 0x00fe, // 111111+00/01/10
331  0x03fc, 0x03fd, 0x03fe, // 11111111+00/01/10
332  0x0ffc, 0x0ffd, 0x0ffe, // 1111111111+00/01/10
333  0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx
334  };
335 
336  INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits),
337  bits, 1, 1, codes, 2, 2, 132);
338 }
339 
340 /**
341  * Set up decoder with parameters from demuxer (extradata etc.).
342  */
343 static av_cold int wmavoice_decode_init(AVCodecContext *ctx)
344 {
345  int n, flags, pitch_range, lsp16_flag;
346  WMAVoiceContext *s = ctx->priv_data;
347 
348  /**
349  * Extradata layout:
350  * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c),
351  * - byte 19-22: flags field (annoyingly in LE; see below for known
352  * values),
353  * - byte 23-46: variable bitmode tree (really just 17 * 3 bits,
354  * rest is 0).
355  */
356  if (ctx->extradata_size != 46) {
357  av_log(ctx, AV_LOG_ERROR,
358  "Invalid extradata size %d (should be 46)\n",
359  ctx->extradata_size);
360  return AVERROR_INVALIDDATA;
361  }
362  flags = AV_RL32(ctx->extradata + 18);
363  s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align);
364  s->do_apf = flags & 0x1;
365  if (s->do_apf) {
366  ff_rdft_init(&s->rdft, 7, DFT_R2C);
367  ff_rdft_init(&s->irdft, 7, IDFT_C2R);
368  ff_dct_init(&s->dct, 6, DCT_I);
369  ff_dct_init(&s->dst, 6, DST_I);
370 
371  ff_sine_window_init(s->cos, 256);
372  memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0]));
373  for (n = 0; n < 255; n++) {
374  s->sin[n] = -s->sin[510 - n];
375  s->cos[510 - n] = s->cos[n];
376  }
377  }
378  s->denoise_strength = (flags >> 2) & 0xF;
379  if (s->denoise_strength >= 12) {
380  av_log(ctx, AV_LOG_ERROR,
381  "Invalid denoise filter strength %d (max=11)\n",
382  s->denoise_strength);
383  return AVERROR_INVALIDDATA;
384  }
385  s->denoise_tilt_corr = !!(flags & 0x40);
386  s->dc_level = (flags >> 7) & 0xF;
387  s->lsp_q_mode = !!(flags & 0x2000);
388  s->lsp_def_mode = !!(flags & 0x4000);
389  lsp16_flag = flags & 0x1000;
390  if (lsp16_flag) {
391  s->lsps = 16;
392  s->frame_lsp_bitsize = 34;
393  s->sframe_lsp_bitsize = 60;
394  } else {
395  s->lsps = 10;
396  s->frame_lsp_bitsize = 24;
397  s->sframe_lsp_bitsize = 48;
398  }
399  for (n = 0; n < s->lsps; n++)
400  s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
401 
402  init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3);
403  if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) {
404  av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n");
405  return AVERROR_INVALIDDATA;
406  }
407 
408  s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8;
409  s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8;
410  pitch_range = s->max_pitch_val - s->min_pitch_val;
411  if (pitch_range <= 0) {
412  av_log(ctx, AV_LOG_ERROR, "Invalid pitch range; broken extradata?\n");
413  return AVERROR_INVALIDDATA;
414  }
415  s->pitch_nbits = av_ceil_log2(pitch_range);
416  s->last_pitch_val = 40;
417  s->last_acb_type = ACB_TYPE_NONE;
418  s->history_nsamples = s->max_pitch_val + 8;
419 
420  if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) {
421  int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8,
422  max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8;
423 
424  av_log(ctx, AV_LOG_ERROR,
425  "Unsupported samplerate %d (min=%d, max=%d)\n",
426  ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz
427 
428  return AVERROR(ENOSYS);
429  }
430 
431  s->block_conv_table[0] = s->min_pitch_val;
432  s->block_conv_table[1] = (pitch_range * 25) >> 6;
433  s->block_conv_table[2] = (pitch_range * 44) >> 6;
434  s->block_conv_table[3] = s->max_pitch_val - 1;
435  s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF;
436  if (s->block_delta_pitch_hrange <= 0) {
437  av_log(ctx, AV_LOG_ERROR, "Invalid delta pitch hrange; broken extradata?\n");
438  return AVERROR_INVALIDDATA;
439  }
440  s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange);
441  s->block_pitch_range = s->block_conv_table[2] +
442  s->block_conv_table[3] + 1 +
443  2 * (s->block_conv_table[1] - 2 * s->min_pitch_val);
444  s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range);
445 
446  ctx->channels = 1;
447  ctx->channel_layout = AV_CH_LAYOUT_MONO;
448  ctx->sample_fmt = AV_SAMPLE_FMT_FLT;
449 
450  return 0;
451 }
452 
453 /**
454  * @name Postfilter functions
455  * Postfilter functions (gain control, wiener denoise filter, DC filter,
456  * kalman smoothening, plus surrounding code to wrap it)
457  * @{
458  */
459 /**
460  * Adaptive gain control (as used in postfilter).
461  *
462  * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except
463  * that the energy here is calculated using sum(abs(...)), whereas the
464  * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)).
465  *
466  * @param out output buffer for filtered samples
467  * @param in input buffer containing the samples as they are after the
468  * postfilter steps so far
469  * @param speech_synth input buffer containing speech synth before postfilter
470  * @param size input buffer size
471  * @param alpha exponential filter factor
472  * @param gain_mem pointer to filter memory (single float)
473  */
474 static void adaptive_gain_control(float *out, const float *in,
475  const float *speech_synth,
476  int size, float alpha, float *gain_mem)
477 {
478  int i;
479  float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor;
480  float mem = *gain_mem;
481 
482  for (i = 0; i < size; i++) {
483  speech_energy += fabsf(speech_synth[i]);
484  postfilter_energy += fabsf(in[i]);
485  }
486  gain_scale_factor = (1.0 - alpha) * speech_energy / postfilter_energy;
487 
488  for (i = 0; i < size; i++) {
489  mem = alpha * mem + gain_scale_factor;
490  out[i] = in[i] * mem;
491  }
492 
493  *gain_mem = mem;
494 }
495 
496 /**
497  * Kalman smoothing function.
498  *
499  * This function looks back pitch +/- 3 samples back into history to find
500  * the best fitting curve (that one giving the optimal gain of the two
501  * signals, i.e. the highest dot product between the two), and then
502  * uses that signal history to smoothen the output of the speech synthesis
503  * filter.
504  *
505  * @param s WMA Voice decoding context
506  * @param pitch pitch of the speech signal
507  * @param in input speech signal
508  * @param out output pointer for smoothened signal
509  * @param size input/output buffer size
510  *
511  * @returns -1 if no smoothening took place, e.g. because no optimal
512  * fit could be found, or 0 on success.
513  */
514 static int kalman_smoothen(WMAVoiceContext *s, int pitch,
515  const float *in, float *out, int size)
516 {
517  int n;
518  float optimal_gain = 0, dot;
519  const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)],
520  *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)],
521  *best_hist_ptr = NULL;
522 
523  /* find best fitting point in history */
524  do {
525  dot = avpriv_scalarproduct_float_c(in, ptr, size);
526  if (dot > optimal_gain) {
527  optimal_gain = dot;
528  best_hist_ptr = ptr;
529  }
530  } while (--ptr >= end);
531 
532  if (optimal_gain <= 0)
533  return -1;
534  dot = avpriv_scalarproduct_float_c(best_hist_ptr, best_hist_ptr, size);
535  if (dot <= 0) // would be 1.0
536  return -1;
537 
538  if (optimal_gain <= dot) {
539  dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000
540  } else
541  dot = 0.625;
542 
543  /* actual smoothing */
544  for (n = 0; n < size; n++)
545  out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]);
546 
547  return 0;
548 }
549 
550 /**
551  * Get the tilt factor of a formant filter from its transfer function
552  * @see #tilt_factor() in amrnbdec.c, which does essentially the same,
553  * but somehow (??) it does a speech synthesis filter in the
554  * middle, which is missing here
555  *
556  * @param lpcs LPC coefficients
557  * @param n_lpcs Size of LPC buffer
558  * @returns the tilt factor
559  */
560 static float tilt_factor(const float *lpcs, int n_lpcs)
561 {
562  float rh0, rh1;
563 
564  rh0 = 1.0 + avpriv_scalarproduct_float_c(lpcs, lpcs, n_lpcs);
565  rh1 = lpcs[0] + avpriv_scalarproduct_float_c(lpcs, &lpcs[1], n_lpcs - 1);
566 
567  return rh1 / rh0;
568 }
569 
570 /**
571  * Derive denoise filter coefficients (in real domain) from the LPCs.
572  */
573 static void calc_input_response(WMAVoiceContext *s, float *lpcs,
574  int fcb_type, float *coeffs, int remainder)
575 {
576  float last_coeff, min = 15.0, max = -15.0;
577  float irange, angle_mul, gain_mul, range, sq;
578  int n, idx;
579 
580  /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */
581  s->rdft.rdft_calc(&s->rdft, lpcs);
582 #define log_range(var, assign) do { \
583  float tmp = log10f(assign); var = tmp; \
584  max = FFMAX(max, tmp); min = FFMIN(min, tmp); \
585  } while (0)
586  log_range(last_coeff, lpcs[1] * lpcs[1]);
587  for (n = 1; n < 64; n++)
588  log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] +
589  lpcs[n * 2 + 1] * lpcs[n * 2 + 1]);
590  log_range(lpcs[0], lpcs[0] * lpcs[0]);
591 #undef log_range
592  range = max - min;
593  lpcs[64] = last_coeff;
594 
595  /* Now, use this spectrum to pick out these frequencies with higher
596  * (relative) power/energy (which we then take to be "not noise"),
597  * and set up a table (still in lpc[]) of (relative) gains per frequency.
598  * These frequencies will be maintained, while others ("noise") will be
599  * decreased in the filter output. */
600  irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63]
601  gain_mul = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) :
602  (5.0 / 14.7));
603  angle_mul = gain_mul * (8.0 * M_LN10 / M_PI);
604  for (n = 0; n <= 64; n++) {
605  float pwr;
606 
607  idx = FFMAX(0, lrint((max - lpcs[n]) * irange) - 1);
608  pwr = wmavoice_denoise_power_table[s->denoise_strength][idx];
609  lpcs[n] = angle_mul * pwr;
610 
611  /* 70.57 =~ 1/log10(1.0331663) */
612  idx = (pwr * gain_mul - 0.0295) * 70.570526123;
613  if (idx > 127) { // fall back if index falls outside table range
614  coeffs[n] = wmavoice_energy_table[127] *
615  powf(1.0331663, idx - 127);
616  } else
617  coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)];
618  }
619 
620  /* calculate the Hilbert transform of the gains, which we do (since this
621  * is a sine input) by doing a phase shift (in theory, H(sin())=cos()).
622  * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the
623  * "moment" of the LPCs in this filter. */
624  s->dct.dct_calc(&s->dct, lpcs);
625  s->dst.dct_calc(&s->dst, lpcs);
626 
627  /* Split out the coefficient indexes into phase/magnitude pairs */
628  idx = 255 + av_clip(lpcs[64], -255, 255);
629  coeffs[0] = coeffs[0] * s->cos[idx];
630  idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255);
631  last_coeff = coeffs[64] * s->cos[idx];
632  for (n = 63;; n--) {
633  idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255);
634  coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
635  coeffs[n * 2] = coeffs[n] * s->cos[idx];
636 
637  if (!--n) break;
638 
639  idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255);
640  coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
641  coeffs[n * 2] = coeffs[n] * s->cos[idx];
642  }
643  coeffs[1] = last_coeff;
644 
645  /* move into real domain */
646  s->irdft.rdft_calc(&s->irdft, coeffs);
647 
648  /* tilt correction and normalize scale */
649  memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder));
650  if (s->denoise_tilt_corr) {
651  float tilt_mem = 0;
652 
653  coeffs[remainder - 1] = 0;
654  ff_tilt_compensation(&tilt_mem,
655  -1.8 * tilt_factor(coeffs, remainder - 1),
656  coeffs, remainder);
657  }
658  sq = (1.0 / 64.0) * sqrtf(1 / avpriv_scalarproduct_float_c(coeffs, coeffs,
659  remainder));
660  for (n = 0; n < remainder; n++)
661  coeffs[n] *= sq;
662 }
663 
664 /**
665  * This function applies a Wiener filter on the (noisy) speech signal as
666  * a means to denoise it.
667  *
668  * - take RDFT of LPCs to get the power spectrum of the noise + speech;
669  * - using this power spectrum, calculate (for each frequency) the Wiener
670  * filter gain, which depends on the frequency power and desired level
671  * of noise subtraction (when set too high, this leads to artifacts)
672  * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse
673  * of 4-8kHz);
674  * - by doing a phase shift, calculate the Hilbert transform of this array
675  * of per-frequency filter-gains to get the filtering coefficients;
676  * - smoothen/normalize/de-tilt these filter coefficients as desired;
677  * - take RDFT of noisy sound, apply the coefficients and take its IRDFT
678  * to get the denoised speech signal;
679  * - the leftover (i.e. output of the IRDFT on denoised speech data beyond
680  * the frame boundary) are saved and applied to subsequent frames by an
681  * overlap-add method (otherwise you get clicking-artifacts).
682  *
683  * @param s WMA Voice decoding context
684  * @param fcb_type Frame (codebook) type
685  * @param synth_pf input: the noisy speech signal, output: denoised speech
686  * data; should be 16-byte aligned (for ASM purposes)
687  * @param size size of the speech data
688  * @param lpcs LPCs used to synthesize this frame's speech data
689  */
690 static void wiener_denoise(WMAVoiceContext *s, int fcb_type,
691  float *synth_pf, int size,
692  const float *lpcs)
693 {
694  int remainder, lim, n;
695 
696  if (fcb_type != FCB_TYPE_SILENCE) {
697  float *tilted_lpcs = s->tilted_lpcs_pf,
698  *coeffs = s->denoise_coeffs_pf, tilt_mem = 0;
699 
700  tilted_lpcs[0] = 1.0;
701  memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps);
702  memset(&tilted_lpcs[s->lsps + 1], 0,
703  sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1));
704  ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps),
705  tilted_lpcs, s->lsps + 2);
706 
707  /* The IRDFT output (127 samples for 7-bit filter) beyond the frame
708  * size is applied to the next frame. All input beyond this is zero,
709  * and thus all output beyond this will go towards zero, hence we can
710  * limit to min(size-1, 127-size) as a performance consideration. */
711  remainder = FFMIN(127 - size, size - 1);
712  calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder);
713 
714  /* apply coefficients (in frequency spectrum domain), i.e. complex
715  * number multiplication */
716  memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size));
717  s->rdft.rdft_calc(&s->rdft, synth_pf);
718  s->rdft.rdft_calc(&s->rdft, coeffs);
719  synth_pf[0] *= coeffs[0];
720  synth_pf[1] *= coeffs[1];
721  for (n = 1; n < 64; n++) {
722  float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1];
723  synth_pf[n * 2] = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1];
724  synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1];
725  }
726  s->irdft.rdft_calc(&s->irdft, synth_pf);
727  }
728 
729  /* merge filter output with the history of previous runs */
730  if (s->denoise_filter_cache_size) {
731  lim = FFMIN(s->denoise_filter_cache_size, size);
732  for (n = 0; n < lim; n++)
733  synth_pf[n] += s->denoise_filter_cache[n];
734  s->denoise_filter_cache_size -= lim;
735  memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size],
736  sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size);
737  }
738 
739  /* move remainder of filter output into a cache for future runs */
740  if (fcb_type != FCB_TYPE_SILENCE) {
741  lim = FFMIN(remainder, s->denoise_filter_cache_size);
742  for (n = 0; n < lim; n++)
743  s->denoise_filter_cache[n] += synth_pf[size + n];
744  if (lim < remainder) {
745  memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim],
746  sizeof(s->denoise_filter_cache[0]) * (remainder - lim));
747  s->denoise_filter_cache_size = remainder;
748  }
749  }
750 }
751 
752 /**
753  * Averaging projection filter, the postfilter used in WMAVoice.
754  *
755  * This uses the following steps:
756  * - A zero-synthesis filter (generate excitation from synth signal)
757  * - Kalman smoothing on excitation, based on pitch
758  * - Re-synthesized smoothened output
759  * - Iterative Wiener denoise filter
760  * - Adaptive gain filter
761  * - DC filter
762  *
763  * @param s WMAVoice decoding context
764  * @param synth Speech synthesis output (before postfilter)
765  * @param samples Output buffer for filtered samples
766  * @param size Buffer size of synth & samples
767  * @param lpcs Generated LPCs used for speech synthesis
768  * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned)
769  * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses)
770  * @param pitch Pitch of the input signal
771  */
772 static void postfilter(WMAVoiceContext *s, const float *synth,
773  float *samples, int size,
774  const float *lpcs, float *zero_exc_pf,
775  int fcb_type, int pitch)
776 {
777  float synth_filter_in_buf[MAX_FRAMESIZE / 2],
778  *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16],
779  *synth_filter_in = zero_exc_pf;
780 
781  av_assert0(size <= MAX_FRAMESIZE / 2);
782 
783  /* generate excitation from input signal */
784  ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps);
785 
786  if (fcb_type >= FCB_TYPE_AW_PULSES &&
787  !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size))
788  synth_filter_in = synth_filter_in_buf;
789 
790  /* re-synthesize speech after smoothening, and keep history */
791  ff_celp_lp_synthesis_filterf(synth_pf, lpcs,
792  synth_filter_in, size, s->lsps);
793  memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps],
794  sizeof(synth_pf[0]) * s->lsps);
795 
796  wiener_denoise(s, fcb_type, synth_pf, size, lpcs);
797 
798  adaptive_gain_control(samples, synth_pf, synth, size, 0.99,
799  &s->postfilter_agc);
800 
801  if (s->dc_level > 8) {
802  /* remove ultra-low frequency DC noise / highpass filter;
803  * coefficients are identical to those used in SIPR decoding,
804  * and very closely resemble those used in AMR-NB decoding. */
805  ff_acelp_apply_order_2_transfer_function(samples, samples,
806  (const float[2]) { -1.99997, 1.0 },
807  (const float[2]) { -1.9330735188, 0.93589198496 },
808  0.93980580475, s->dcf_mem, size);
809  }
810 }
811 /**
812  * @}
813  */
814 
815 /**
816  * Dequantize LSPs
817  * @param lsps output pointer to the array that will hold the LSPs
818  * @param num number of LSPs to be dequantized
819  * @param values quantized values, contains n_stages values
820  * @param sizes range (i.e. max value) of each quantized value
821  * @param n_stages number of dequantization runs
822  * @param table dequantization table to be used
823  * @param mul_q LSF multiplier
824  * @param base_q base (lowest) LSF values
825  */
826 static void dequant_lsps(double *lsps, int num,
827  const uint16_t *values,
828  const uint16_t *sizes,
829  int n_stages, const uint8_t *table,
830  const double *mul_q,
831  const double *base_q)
832 {
833  int n, m;
834 
835  memset(lsps, 0, num * sizeof(*lsps));
836  for (n = 0; n < n_stages; n++) {
837  const uint8_t *t_off = &table[values[n] * num];
838  double base = base_q[n], mul = mul_q[n];
839 
840  for (m = 0; m < num; m++)
841  lsps[m] += base + mul * t_off[m];
842 
843  table += sizes[n] * num;
844  }
845 }
846 
847 /**
848  * @name LSP dequantization routines
849  * LSP dequantization routines, for 10/16LSPs and independent/residual coding.
850  * @note we assume enough bits are available, caller should check.
851  * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;
852  * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits.
853  * @{
854  */
855 /**
856  * Parse 10 independently-coded LSPs.
857  */
858 static void dequant_lsp10i(GetBitContext *gb, double *lsps)
859 {
860  static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 };
861  static const double mul_lsf[4] = {
862  5.2187144800e-3, 1.4626986422e-3,
863  9.6179549166e-4, 1.1325736225e-3
864  };
865  static const double base_lsf[4] = {
866  M_PI * -2.15522e-1, M_PI * -6.1646e-2,
867  M_PI * -3.3486e-2, M_PI * -5.7408e-2
868  };
869  uint16_t v[4];
870 
871  v[0] = get_bits(gb, 8);
872  v[1] = get_bits(gb, 6);
873  v[2] = get_bits(gb, 5);
874  v[3] = get_bits(gb, 5);
875 
876  dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i,
877  mul_lsf, base_lsf);
878 }
879 
880 /**
881  * Parse 10 independently-coded LSPs, and then derive the tables to
882  * generate LSPs for the other frames from them (residual coding).
883  */
884 static void dequant_lsp10r(GetBitContext *gb,
885  double *i_lsps, const double *old,
886  double *a1, double *a2, int q_mode)
887 {
888  static const uint16_t vec_sizes[3] = { 128, 64, 64 };
889  static const double mul_lsf[3] = {
890  2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3
891  };
892  static const double base_lsf[3] = {
893  M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2
894  };
895  const float (*ipol_tab)[2][10] = q_mode ?
896  wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a;
897  uint16_t interpol, v[3];
898  int n;
899 
900  dequant_lsp10i(gb, i_lsps);
901 
902  interpol = get_bits(gb, 5);
903  v[0] = get_bits(gb, 7);
904  v[1] = get_bits(gb, 6);
905  v[2] = get_bits(gb, 6);
906 
907  for (n = 0; n < 10; n++) {
908  double delta = old[n] - i_lsps[n];
909  a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
910  a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
911  }
912 
913  dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r,
914  mul_lsf, base_lsf);
915 }
916 
917 /**
918  * Parse 16 independently-coded LSPs.
919  */
920 static void dequant_lsp16i(GetBitContext *gb, double *lsps)
921 {
922  static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 };
923  static const double mul_lsf[5] = {
924  3.3439586280e-3, 6.9908173703e-4,
925  3.3216608306e-3, 1.0334960326e-3,
926  3.1899104283e-3
927  };
928  static const double base_lsf[5] = {
929  M_PI * -1.27576e-1, M_PI * -2.4292e-2,
930  M_PI * -1.28094e-1, M_PI * -3.2128e-2,
931  M_PI * -1.29816e-1
932  };
933  uint16_t v[5];
934 
935  v[0] = get_bits(gb, 8);
936  v[1] = get_bits(gb, 6);
937  v[2] = get_bits(gb, 7);
938  v[3] = get_bits(gb, 6);
939  v[4] = get_bits(gb, 7);
940 
941  dequant_lsps( lsps, 5, v, vec_sizes, 2,
942  wmavoice_dq_lsp16i1, mul_lsf, base_lsf);
943  dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2,
944  wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]);
945  dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1,
946  wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]);
947 }
948 
949 /**
950  * Parse 16 independently-coded LSPs, and then derive the tables to
951  * generate LSPs for the other frames from them (residual coding).
952  */
953 static void dequant_lsp16r(GetBitContext *gb,
954  double *i_lsps, const double *old,
955  double *a1, double *a2, int q_mode)
956 {
957  static const uint16_t vec_sizes[3] = { 128, 128, 128 };
958  static const double mul_lsf[3] = {
959  1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3
960  };
961  static const double base_lsf[3] = {
962  M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2
963  };
964  const float (*ipol_tab)[2][16] = q_mode ?
965  wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a;
966  uint16_t interpol, v[3];
967  int n;
968 
969  dequant_lsp16i(gb, i_lsps);
970 
971  interpol = get_bits(gb, 5);
972  v[0] = get_bits(gb, 7);
973  v[1] = get_bits(gb, 7);
974  v[2] = get_bits(gb, 7);
975 
976  for (n = 0; n < 16; n++) {
977  double delta = old[n] - i_lsps[n];
978  a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
979  a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
980  }
981 
982  dequant_lsps( a2, 10, v, vec_sizes, 1,
983  wmavoice_dq_lsp16r1, mul_lsf, base_lsf);
984  dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1,
985  wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]);
986  dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1,
987  wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]);
988 }
989 
990 /**
991  * @}
992  * @name Pitch-adaptive window coding functions
993  * The next few functions are for pitch-adaptive window coding.
994  * @{
995  */
996 /**
997  * Parse the offset of the first pitch-adaptive window pulses, and
998  * the distribution of pulses between the two blocks in this frame.
999  * @param s WMA Voice decoding context private data
1000  * @param gb bit I/O context
1001  * @param pitch pitch for each block in this frame
1002  */
1003 static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb,
1004  const int *pitch)
1005 {
1006  static const int16_t start_offset[94] = {
1007  -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11,
1008  13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26,
1009  27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43,
1010  45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67,
1011  69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91,
1012  93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115,
1013  117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139,
1014  141, 143, 145, 147, 149, 151, 153, 155, 157, 159
1015  };
1016  int bits, offset;
1017 
1018  /* position of pulse */
1019  s->aw_idx_is_ext = 0;
1020  if ((bits = get_bits(gb, 6)) >= 54) {
1021  s->aw_idx_is_ext = 1;
1022  bits += (bits - 54) * 3 + get_bits(gb, 2);
1023  }
1024 
1025  /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count
1026  * the distribution of the pulses in each block contained in this frame. */
1027  s->aw_pulse_range = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16;
1028  for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ;
1029  s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0];
1030  s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2;
1031  offset += s->aw_n_pulses[0] * pitch[0];
1032  s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1];
1033  s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2;
1034 
1035  /* if continuing from a position before the block, reset position to
1036  * start of block (when corrected for the range over which it can be
1037  * spread in aw_pulse_set1()). */
1038  if (start_offset[bits] < MAX_FRAMESIZE / 2) {
1039  while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0)
1040  s->aw_first_pulse_off[1] -= pitch[1];
1041  if (start_offset[bits] < 0)
1042  while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0)
1043  s->aw_first_pulse_off[0] -= pitch[0];
1044  }
1045 }
1046 
1047 /**
1048  * Apply second set of pitch-adaptive window pulses.
1049  * @param s WMA Voice decoding context private data
1050  * @param gb bit I/O context
1051  * @param block_idx block index in frame [0, 1]
1052  * @param fcb structure containing fixed codebook vector info
1053  * @return -1 on error, 0 otherwise
1054  */
1055 static int aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb,
1056  int block_idx, AMRFixed *fcb)
1057 {
1058  uint16_t use_mask_mem[9]; // only 5 are used, rest is padding
1059  uint16_t *use_mask = use_mask_mem + 2;
1060  /* in this function, idx is the index in the 80-bit (+ padding) use_mask
1061  * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits
1062  * of idx are the position of the bit within a particular item in the
1063  * array (0 being the most significant bit, and 15 being the least
1064  * significant bit), and the remainder (>> 4) is the index in the
1065  * use_mask[]-array. This is faster and uses less memory than using a
1066  * 80-byte/80-int array. */
1067  int pulse_off = s->aw_first_pulse_off[block_idx],
1068  pulse_start, n, idx, range, aidx, start_off = 0;
1069 
1070  /* set offset of first pulse to within this block */
1071  if (s->aw_n_pulses[block_idx] > 0)
1072  while (pulse_off + s->aw_pulse_range < 1)
1073  pulse_off += fcb->pitch_lag;
1074 
1075  /* find range per pulse */
1076  if (s->aw_n_pulses[0] > 0) {
1077  if (block_idx == 0) {
1078  range = 32;
1079  } else /* block_idx = 1 */ {
1080  range = 8;
1081  if (s->aw_n_pulses[block_idx] > 0)
1082  pulse_off = s->aw_next_pulse_off_cache;
1083  }
1084  } else
1085  range = 16;
1086  pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0;
1087 
1088  /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,
1089  * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus
1090  * we exclude that range from being pulsed again in this function. */
1091  memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0]));
1092  memset( use_mask, -1, 5 * sizeof(use_mask[0]));
1093  memset(&use_mask[5], 0, 2 * sizeof(use_mask[0]));
1094  if (s->aw_n_pulses[block_idx] > 0)
1095  for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) {
1096  int excl_range = s->aw_pulse_range; // always 16 or 24
1097  uint16_t *use_mask_ptr = &use_mask[idx >> 4];
1098  int first_sh = 16 - (idx & 15);
1099  *use_mask_ptr++ &= 0xFFFFu << first_sh;
1100  excl_range -= first_sh;
1101  if (excl_range >= 16) {
1102  *use_mask_ptr++ = 0;
1103  *use_mask_ptr &= 0xFFFF >> (excl_range - 16);
1104  } else
1105  *use_mask_ptr &= 0xFFFF >> excl_range;
1106  }
1107 
1108  /* find the 'aidx'th offset that is not excluded */
1109  aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4);
1110  for (n = 0; n <= aidx; pulse_start++) {
1111  for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ;
1112  if (idx >= MAX_FRAMESIZE / 2) { // find from zero
1113  if (use_mask[0]) idx = 0x0F;
1114  else if (use_mask[1]) idx = 0x1F;
1115  else if (use_mask[2]) idx = 0x2F;
1116  else if (use_mask[3]) idx = 0x3F;
1117  else if (use_mask[4]) idx = 0x4F;
1118  else return -1;
1119  idx -= av_log2_16bit(use_mask[idx >> 4]);
1120  }
1121  if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) {
1122  use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15));
1123  n++;
1124  start_off = idx;
1125  }
1126  }
1127 
1128  fcb->x[fcb->n] = start_off;
1129  fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0;
1130  fcb->n++;
1131 
1132  /* set offset for next block, relative to start of that block */
1133  n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag;
1134  s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0;
1135  return 0;
1136 }
1137 
1138 /**
1139  * Apply first set of pitch-adaptive window pulses.
1140  * @param s WMA Voice decoding context private data
1141  * @param gb bit I/O context
1142  * @param block_idx block index in frame [0, 1]
1143  * @param fcb storage location for fixed codebook pulse info
1144  */
1145 static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb,
1146  int block_idx, AMRFixed *fcb)
1147 {
1148  int val = get_bits(gb, 12 - 2 * (s->aw_idx_is_ext && !block_idx));
1149  float v;
1150 
1151  if (s->aw_n_pulses[block_idx] > 0) {
1152  int n, v_mask, i_mask, sh, n_pulses;
1153 
1154  if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each
1155  n_pulses = 3;
1156  v_mask = 8;
1157  i_mask = 7;
1158  sh = 4;
1159  } else { // 4 pulses, 1:sign + 2:index each
1160  n_pulses = 4;
1161  v_mask = 4;
1162  i_mask = 3;
1163  sh = 3;
1164  }
1165 
1166  for (n = n_pulses - 1; n >= 0; n--, val >>= sh) {
1167  fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0;
1168  fcb->x[fcb->n] = (val & i_mask) * n_pulses + n +
1169  s->aw_first_pulse_off[block_idx];
1170  while (fcb->x[fcb->n] < 0)
1171  fcb->x[fcb->n] += fcb->pitch_lag;
1172  if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2)
1173  fcb->n++;
1174  }
1175  } else {
1176  int num2 = (val & 0x1FF) >> 1, delta, idx;
1177 
1178  if (num2 < 1 * 79) { delta = 1; idx = num2 + 1; }
1179  else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; }
1180  else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; }
1181  else { delta = 7; idx = num2 + 1 - 3 * 75; }
1182  v = (val & 0x200) ? -1.0 : 1.0;
1183 
1184  fcb->no_repeat_mask |= 3 << fcb->n;
1185  fcb->x[fcb->n] = idx - delta;
1186  fcb->y[fcb->n] = v;
1187  fcb->x[fcb->n + 1] = idx;
1188  fcb->y[fcb->n + 1] = (val & 1) ? -v : v;
1189  fcb->n += 2;
1190  }
1191 }
1192 
1193 /**
1194  * @}
1195  *
1196  * Generate a random number from frame_cntr and block_idx, which will lief
1197  * in the range [0, 1000 - block_size] (so it can be used as an index in a
1198  * table of size 1000 of which you want to read block_size entries).
1199  *
1200  * @param frame_cntr current frame number
1201  * @param block_num current block index
1202  * @param block_size amount of entries we want to read from a table
1203  * that has 1000 entries
1204  * @return a (non-)random number in the [0, 1000 - block_size] range.
1205  */
1206 static int pRNG(int frame_cntr, int block_num, int block_size)
1207 {
1208  /* array to simplify the calculation of z:
1209  * y = (x % 9) * 5 + 6;
1210  * z = (49995 * x) / y;
1211  * Since y only has 9 values, we can remove the division by using a
1212  * LUT and using FASTDIV-style divisions. For each of the 9 values
1213  * of y, we can rewrite z as:
1214  * z = x * (49995 / y) + x * ((49995 % y) / y)
1215  * In this table, each col represents one possible value of y, the
1216  * first number is 49995 / y, and the second is the FASTDIV variant
1217  * of 49995 % y / y. */
1218  static const unsigned int div_tbl[9][2] = {
1219  { 8332, 3 * 715827883U }, // y = 6
1220  { 4545, 0 * 390451573U }, // y = 11
1221  { 3124, 11 * 268435456U }, // y = 16
1222  { 2380, 15 * 204522253U }, // y = 21
1223  { 1922, 23 * 165191050U }, // y = 26
1224  { 1612, 23 * 138547333U }, // y = 31
1225  { 1388, 27 * 119304648U }, // y = 36
1226  { 1219, 16 * 104755300U }, // y = 41
1227  { 1086, 39 * 93368855U } // y = 46
1228  };
1229  unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr;
1230  if (x >= 0xFFFF) x -= 0xFFFF; // max value of x is 8*1877+0xFFFE=0x13AA6,
1231  // so this is effectively a modulo (%)
1232  y = x - 9 * MULH(477218589, x); // x % 9
1233  z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1]));
1234  // z = x * 49995 / (y * 5 + 6)
1235  return z % (1000 - block_size);
1236 }
1237 
1238 /**
1239  * Parse hardcoded signal for a single block.
1240  * @note see #synth_block().
1241  */
1242 static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb,
1243  int block_idx, int size,
1244  const struct frame_type_desc *frame_desc,
1245  float *excitation)
1246 {
1247  float gain;
1248  int n, r_idx;
1249 
1250  av_assert0(size <= MAX_FRAMESIZE);
1251 
1252  /* Set the offset from which we start reading wmavoice_std_codebook */
1253  if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1254  r_idx = pRNG(s->frame_cntr, block_idx, size);
1255  gain = s->silence_gain;
1256  } else /* FCB_TYPE_HARDCODED */ {
1257  r_idx = get_bits(gb, 8);
1258  gain = wmavoice_gain_universal[get_bits(gb, 6)];
1259  }
1260 
1261  /* Clear gain prediction parameters */
1262  memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err));
1263 
1264  /* Apply gain to hardcoded codebook and use that as excitation signal */
1265  for (n = 0; n < size; n++)
1266  excitation[n] = wmavoice_std_codebook[r_idx + n] * gain;
1267 }
1268 
1269 /**
1270  * Parse FCB/ACB signal for a single block.
1271  * @note see #synth_block().
1272  */
1273 static void synth_block_fcb_acb(WMAVoiceContext *s, GetBitContext *gb,
1274  int block_idx, int size,
1275  int block_pitch_sh2,
1276  const struct frame_type_desc *frame_desc,
1277  float *excitation)
1278 {
1279  static const float gain_coeff[6] = {
1280  0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458
1281  };
1282  float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain;
1283  int n, idx, gain_weight;
1284  AMRFixed fcb;
1285 
1286  av_assert0(size <= MAX_FRAMESIZE / 2);
1287  memset(pulses, 0, sizeof(*pulses) * size);
1288 
1289  fcb.pitch_lag = block_pitch_sh2 >> 2;
1290  fcb.pitch_fac = 1.0;
1291  fcb.no_repeat_mask = 0;
1292  fcb.n = 0;
1293 
1294  /* For the other frame types, this is where we apply the innovation
1295  * (fixed) codebook pulses of the speech signal. */
1296  if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1297  aw_pulse_set1(s, gb, block_idx, &fcb);
1298  if (aw_pulse_set2(s, gb, block_idx, &fcb)) {
1299  /* Conceal the block with silence and return.
1300  * Skip the correct amount of bits to read the next
1301  * block from the correct offset. */
1302  int r_idx = pRNG(s->frame_cntr, block_idx, size);
1303 
1304  for (n = 0; n < size; n++)
1305  excitation[n] =
1306  wmavoice_std_codebook[r_idx + n] * s->silence_gain;
1307  skip_bits(gb, 7 + 1);
1308  return;
1309  }
1310  } else /* FCB_TYPE_EXC_PULSES */ {
1311  int offset_nbits = 5 - frame_desc->log_n_blocks;
1312 
1313  fcb.no_repeat_mask = -1;
1314  /* similar to ff_decode_10_pulses_35bits(), but with single pulses
1315  * (instead of double) for a subset of pulses */
1316  for (n = 0; n < 5; n++) {
1317  float sign;
1318  int pos1, pos2;
1319 
1320  sign = get_bits1(gb) ? 1.0 : -1.0;
1321  pos1 = get_bits(gb, offset_nbits);
1322  fcb.x[fcb.n] = n + 5 * pos1;
1323  fcb.y[fcb.n++] = sign;
1324  if (n < frame_desc->dbl_pulses) {
1325  pos2 = get_bits(gb, offset_nbits);
1326  fcb.x[fcb.n] = n + 5 * pos2;
1327  fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign;
1328  }
1329  }
1330  }
1331  ff_set_fixed_vector(pulses, &fcb, 1.0, size);
1332 
1333  /* Calculate gain for adaptive & fixed codebook signal.
1334  * see ff_amr_set_fixed_gain(). */
1335  idx = get_bits(gb, 7);
1336  fcb_gain = expf(avpriv_scalarproduct_float_c(s->gain_pred_err,
1337  gain_coeff, 6) -
1338  5.2409161640 + wmavoice_gain_codebook_fcb[idx]);
1339  acb_gain = wmavoice_gain_codebook_acb[idx];
1340  pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx],
1341  -2.9957322736 /* log(0.05) */,
1342  1.6094379124 /* log(5.0) */);
1343 
1344  gain_weight = 8 >> frame_desc->log_n_blocks;
1345  memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err,
1346  sizeof(*s->gain_pred_err) * (6 - gain_weight));
1347  for (n = 0; n < gain_weight; n++)
1348  s->gain_pred_err[n] = pred_err;
1349 
1350  /* Calculation of adaptive codebook */
1351  if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1352  int len;
1353  for (n = 0; n < size; n += len) {
1354  int next_idx_sh16;
1355  int abs_idx = block_idx * size + n;
1356  int pitch_sh16 = (s->last_pitch_val << 16) +
1357  s->pitch_diff_sh16 * abs_idx;
1358  int pitch = (pitch_sh16 + 0x6FFF) >> 16;
1359  int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000;
1360  idx = idx_sh16 >> 16;
1361  if (s->pitch_diff_sh16) {
1362  if (s->pitch_diff_sh16 > 0) {
1363  next_idx_sh16 = (idx_sh16) &~ 0xFFFF;
1364  } else
1365  next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF;
1366  len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8,
1367  1, size - n);
1368  } else
1369  len = size;
1370 
1371  ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch],
1372  wmavoice_ipol1_coeffs, 17,
1373  idx, 9, len);
1374  }
1375  } else /* ACB_TYPE_HAMMING */ {
1376  int block_pitch = block_pitch_sh2 >> 2;
1377  idx = block_pitch_sh2 & 3;
1378  if (idx) {
1379  ff_acelp_interpolatef(excitation, &excitation[-block_pitch],
1380  wmavoice_ipol2_coeffs, 4,
1381  idx, 8, size);
1382  } else
1383  av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch,
1384  sizeof(float) * size);
1385  }
1386 
1387  /* Interpolate ACB/FCB and use as excitation signal */
1388  ff_weighted_vector_sumf(excitation, excitation, pulses,
1389  acb_gain, fcb_gain, size);
1390 }
1391 
1392 /**
1393  * Parse data in a single block.
1394  * @note we assume enough bits are available, caller should check.
1395  *
1396  * @param s WMA Voice decoding context private data
1397  * @param gb bit I/O context
1398  * @param block_idx index of the to-be-read block
1399  * @param size amount of samples to be read in this block
1400  * @param block_pitch_sh2 pitch for this block << 2
1401  * @param lsps LSPs for (the end of) this frame
1402  * @param prev_lsps LSPs for the last frame
1403  * @param frame_desc frame type descriptor
1404  * @param excitation target memory for the ACB+FCB interpolated signal
1405  * @param synth target memory for the speech synthesis filter output
1406  * @return 0 on success, <0 on error.
1407  */
1408 static void synth_block(WMAVoiceContext *s, GetBitContext *gb,
1409  int block_idx, int size,
1410  int block_pitch_sh2,
1411  const double *lsps, const double *prev_lsps,
1412  const struct frame_type_desc *frame_desc,
1413  float *excitation, float *synth)
1414 {
1415  double i_lsps[MAX_LSPS];
1416  float lpcs[MAX_LSPS];
1417  float fac;
1418  int n;
1419 
1420  if (frame_desc->acb_type == ACB_TYPE_NONE)
1421  synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation);
1422  else
1423  synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2,
1424  frame_desc, excitation);
1425 
1426  /* convert interpolated LSPs to LPCs */
1427  fac = (block_idx + 0.5) / frame_desc->n_blocks;
1428  for (n = 0; n < s->lsps; n++) // LSF -> LSP
1429  i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n]));
1430  ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1431 
1432  /* Speech synthesis */
1433  ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps);
1434 }
1435 
1436 /**
1437  * Synthesize output samples for a single frame.
1438  * @note we assume enough bits are available, caller should check.
1439  *
1440  * @param ctx WMA Voice decoder context
1441  * @param gb bit I/O context (s->gb or one for cross-packet superframes)
1442  * @param frame_idx Frame number within superframe [0-2]
1443  * @param samples pointer to output sample buffer, has space for at least 160
1444  * samples
1445  * @param lsps LSP array
1446  * @param prev_lsps array of previous frame's LSPs
1447  * @param excitation target buffer for excitation signal
1448  * @param synth target buffer for synthesized speech data
1449  * @return 0 on success, <0 on error.
1450  */
1451 static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx,
1452  float *samples,
1453  const double *lsps, const double *prev_lsps,
1454  float *excitation, float *synth)
1455 {
1456  WMAVoiceContext *s = ctx->priv_data;
1457  int n, n_blocks_x2, log_n_blocks_x2, av_uninit(cur_pitch_val);
1458  int pitch[MAX_BLOCKS], av_uninit(last_block_pitch);
1459 
1460  /* Parse frame type ("frame header"), see frame_descs */
1461  int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)], block_nsamples;
1462 
1463  if (bd_idx < 0) {
1464  av_log(ctx, AV_LOG_ERROR,
1465  "Invalid frame type VLC code, skipping\n");
1466  return AVERROR_INVALIDDATA;
1467  }
1468 
1469  block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks;
1470 
1471  /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */
1472  if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) {
1473  /* Pitch is provided per frame, which is interpreted as the pitch of
1474  * the last sample of the last block of this frame. We can interpolate
1475  * the pitch of other blocks (and even pitch-per-sample) by gradually
1476  * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */
1477  n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1;
1478  log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1;
1479  cur_pitch_val = s->min_pitch_val + get_bits(gb, s->pitch_nbits);
1480  cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1);
1481  if (s->last_acb_type == ACB_TYPE_NONE ||
1482  20 * abs(cur_pitch_val - s->last_pitch_val) >
1483  (cur_pitch_val + s->last_pitch_val))
1484  s->last_pitch_val = cur_pitch_val;
1485 
1486  /* pitch per block */
1487  for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1488  int fac = n * 2 + 1;
1489 
1490  pitch[n] = (MUL16(fac, cur_pitch_val) +
1491  MUL16((n_blocks_x2 - fac), s->last_pitch_val) +
1492  frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2;
1493  }
1494 
1495  /* "pitch-diff-per-sample" for calculation of pitch per sample */
1496  s->pitch_diff_sh16 =
1497  ((cur_pitch_val - s->last_pitch_val) << 16) / MAX_FRAMESIZE;
1498  }
1499 
1500  /* Global gain (if silence) and pitch-adaptive window coordinates */
1501  switch (frame_descs[bd_idx].fcb_type) {
1502  case FCB_TYPE_SILENCE:
1503  s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)];
1504  break;
1505  case FCB_TYPE_AW_PULSES:
1506  aw_parse_coords(s, gb, pitch);
1507  break;
1508  }
1509 
1510  for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1511  int bl_pitch_sh2;
1512 
1513  /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */
1514  switch (frame_descs[bd_idx].acb_type) {
1515  case ACB_TYPE_HAMMING: {
1516  /* Pitch is given per block. Per-block pitches are encoded as an
1517  * absolute value for the first block, and then delta values
1518  * relative to this value) for all subsequent blocks. The scale of
1519  * this pitch value is semi-logaritmic compared to its use in the
1520  * decoder, so we convert it to normal scale also. */
1521  int block_pitch,
1522  t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2,
1523  t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1,
1524  t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1;
1525 
1526  if (n == 0) {
1527  block_pitch = get_bits(gb, s->block_pitch_nbits);
1528  } else
1529  block_pitch = last_block_pitch - s->block_delta_pitch_hrange +
1530  get_bits(gb, s->block_delta_pitch_nbits);
1531  /* Convert last_ so that any next delta is within _range */
1532  last_block_pitch = av_clip(block_pitch,
1533  s->block_delta_pitch_hrange,
1534  s->block_pitch_range -
1535  s->block_delta_pitch_hrange);
1536 
1537  /* Convert semi-log-style scale back to normal scale */
1538  if (block_pitch < t1) {
1539  bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch;
1540  } else {
1541  block_pitch -= t1;
1542  if (block_pitch < t2) {
1543  bl_pitch_sh2 =
1544  (s->block_conv_table[1] << 2) + (block_pitch << 1);
1545  } else {
1546  block_pitch -= t2;
1547  if (block_pitch < t3) {
1548  bl_pitch_sh2 =
1549  (s->block_conv_table[2] + block_pitch) << 2;
1550  } else
1551  bl_pitch_sh2 = s->block_conv_table[3] << 2;
1552  }
1553  }
1554  pitch[n] = bl_pitch_sh2 >> 2;
1555  break;
1556  }
1557 
1558  case ACB_TYPE_ASYMMETRIC: {
1559  bl_pitch_sh2 = pitch[n] << 2;
1560  break;
1561  }
1562 
1563  default: // ACB_TYPE_NONE has no pitch
1564  bl_pitch_sh2 = 0;
1565  break;
1566  }
1567 
1568  synth_block(s, gb, n, block_nsamples, bl_pitch_sh2,
1569  lsps, prev_lsps, &frame_descs[bd_idx],
1570  &excitation[n * block_nsamples],
1571  &synth[n * block_nsamples]);
1572  }
1573 
1574  /* Averaging projection filter, if applicable. Else, just copy samples
1575  * from synthesis buffer */
1576  if (s->do_apf) {
1577  double i_lsps[MAX_LSPS];
1578  float lpcs[MAX_LSPS];
1579 
1580  for (n = 0; n < s->lsps; n++) // LSF -> LSP
1581  i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n]));
1582  ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1583  postfilter(s, synth, samples, 80, lpcs,
1584  &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx],
1585  frame_descs[bd_idx].fcb_type, pitch[0]);
1586 
1587  for (n = 0; n < s->lsps; n++) // LSF -> LSP
1588  i_lsps[n] = cos(lsps[n]);
1589  ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1590  postfilter(s, &synth[80], &samples[80], 80, lpcs,
1591  &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80],
1592  frame_descs[bd_idx].fcb_type, pitch[0]);
1593  } else
1594  memcpy(samples, synth, 160 * sizeof(synth[0]));
1595 
1596  /* Cache values for next frame */
1597  s->frame_cntr++;
1598  if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%)
1599  s->last_acb_type = frame_descs[bd_idx].acb_type;
1600  switch (frame_descs[bd_idx].acb_type) {
1601  case ACB_TYPE_NONE:
1602  s->last_pitch_val = 0;
1603  break;
1604  case ACB_TYPE_ASYMMETRIC:
1605  s->last_pitch_val = cur_pitch_val;
1606  break;
1607  case ACB_TYPE_HAMMING:
1608  s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1];
1609  break;
1610  }
1611 
1612  return 0;
1613 }
1614 
1615 /**
1616  * Ensure minimum value for first item, maximum value for last value,
1617  * proper spacing between each value and proper ordering.
1618  *
1619  * @param lsps array of LSPs
1620  * @param num size of LSP array
1621  *
1622  * @note basically a double version of #ff_acelp_reorder_lsf(), might be
1623  * useful to put in a generic location later on. Parts are also
1624  * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(),
1625  * which is in float.
1626  */
1627 static void stabilize_lsps(double *lsps, int num)
1628 {
1629  int n, m, l;
1630 
1631  /* set minimum value for first, maximum value for last and minimum
1632  * spacing between LSF values.
1633  * Very similar to ff_set_min_dist_lsf(), but in double. */
1634  lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI);
1635  for (n = 1; n < num; n++)
1636  lsps[n] = FFMAX(lsps[n], lsps[n - 1] + 0.0125 * M_PI);
1637  lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI);
1638 
1639  /* reorder (looks like one-time / non-recursed bubblesort).
1640  * Very similar to ff_sort_nearly_sorted_floats(), but in double. */
1641  for (n = 1; n < num; n++) {
1642  if (lsps[n] < lsps[n - 1]) {
1643  for (m = 1; m < num; m++) {
1644  double tmp = lsps[m];
1645  for (l = m - 1; l >= 0; l--) {
1646  if (lsps[l] <= tmp) break;
1647  lsps[l + 1] = lsps[l];
1648  }
1649  lsps[l + 1] = tmp;
1650  }
1651  break;
1652  }
1653  }
1654 }
1655 
1656 /**
1657  * Test if there's enough bits to read 1 superframe.
1658  *
1659  * @param orig_gb bit I/O context used for reading. This function
1660  * does not modify the state of the bitreader; it
1661  * only uses it to copy the current stream position
1662  * @param s WMA Voice decoding context private data
1663  * @return < 0 on error, 1 on not enough bits or 0 if OK.
1664  */
1665 static int check_bits_for_superframe(GetBitContext *orig_gb,
1666  WMAVoiceContext *s)
1667 {
1668  GetBitContext s_gb, *gb = &s_gb;
1669  int n, need_bits, bd_idx;
1670  const struct frame_type_desc *frame_desc;
1671 
1672  /* initialize a copy */
1673  init_get_bits(gb, orig_gb->buffer, orig_gb->size_in_bits);
1674  skip_bits_long(gb, get_bits_count(orig_gb));
1675  av_assert1(get_bits_left(gb) == get_bits_left(orig_gb));
1676 
1677  /* superframe header */
1678  if (get_bits_left(gb) < 14)
1679  return 1;
1680  if (!get_bits1(gb))
1681  return AVERROR(ENOSYS); // WMAPro-in-WMAVoice superframe
1682  if (get_bits1(gb)) skip_bits(gb, 12); // number of samples in superframe
1683  if (s->has_residual_lsps) { // residual LSPs (for all frames)
1684  if (get_bits_left(gb) < s->sframe_lsp_bitsize)
1685  return 1;
1686  skip_bits_long(gb, s->sframe_lsp_bitsize);
1687  }
1688 
1689  /* frames */
1690  for (n = 0; n < MAX_FRAMES; n++) {
1691  int aw_idx_is_ext = 0;
1692 
1693  if (!s->has_residual_lsps) { // independent LSPs (per-frame)
1694  if (get_bits_left(gb) < s->frame_lsp_bitsize) return 1;
1695  skip_bits_long(gb, s->frame_lsp_bitsize);
1696  }
1697  bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)];
1698  if (bd_idx < 0)
1699  return AVERROR_INVALIDDATA; // invalid frame type VLC code
1700  frame_desc = &frame_descs[bd_idx];
1701  if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1702  if (get_bits_left(gb) < s->pitch_nbits)
1703  return 1;
1704  skip_bits_long(gb, s->pitch_nbits);
1705  }
1706  if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1707  skip_bits(gb, 8);
1708  } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1709  int tmp = get_bits(gb, 6);
1710  if (tmp >= 0x36) {
1711  skip_bits(gb, 2);
1712  aw_idx_is_ext = 1;
1713  }
1714  }
1715 
1716  /* blocks */
1717  if (frame_desc->acb_type == ACB_TYPE_HAMMING) {
1718  need_bits = s->block_pitch_nbits +
1719  (frame_desc->n_blocks - 1) * s->block_delta_pitch_nbits;
1720  } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1721  need_bits = 2 * !aw_idx_is_ext;
1722  } else
1723  need_bits = 0;
1724  need_bits += frame_desc->frame_size;
1725  if (get_bits_left(gb) < need_bits)
1726  return 1;
1727  skip_bits_long(gb, need_bits);
1728  }
1729 
1730  return 0;
1731 }
1732 
1733 /**
1734  * Synthesize output samples for a single superframe. If we have any data
1735  * cached in s->sframe_cache, that will be used instead of whatever is loaded
1736  * in s->gb.
1737  *
1738  * WMA Voice superframes contain 3 frames, each containing 160 audio samples,
1739  * to give a total of 480 samples per frame. See #synth_frame() for frame
1740  * parsing. In addition to 3 frames, superframes can also contain the LSPs
1741  * (if these are globally specified for all frames (residually); they can
1742  * also be specified individually per-frame. See the s->has_residual_lsps
1743  * option), and can specify the number of samples encoded in this superframe
1744  * (if less than 480), usually used to prevent blanks at track boundaries.
1745  *
1746  * @param ctx WMA Voice decoder context
1747  * @return 0 on success, <0 on error or 1 if there was not enough data to
1748  * fully parse the superframe
1749  */
1750 static int synth_superframe(AVCodecContext *ctx, AVFrame *frame,
1751  int *got_frame_ptr)
1752 {
1753  WMAVoiceContext *s = ctx->priv_data;
1754  GetBitContext *gb = &s->gb, s_gb;
1755  int n, res, n_samples = 480;
1756  double lsps[MAX_FRAMES][MAX_LSPS];
1757  const double *mean_lsf = s->lsps == 16 ?
1758  wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode];
1759  float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12];
1760  float synth[MAX_LSPS + MAX_SFRAMESIZE];
1761  float *samples;
1762 
1763  memcpy(synth, s->synth_history,
1764  s->lsps * sizeof(*synth));
1765  memcpy(excitation, s->excitation_history,
1766  s->history_nsamples * sizeof(*excitation));
1767 
1768  if (s->sframe_cache_size > 0) {
1769  gb = &s_gb;
1770  init_get_bits(gb, s->sframe_cache, s->sframe_cache_size);
1771  s->sframe_cache_size = 0;
1772  }
1773 
1774  if ((res = check_bits_for_superframe(gb, s)) == 1) {
1775  *got_frame_ptr = 0;
1776  return 1;
1777  } else if (res < 0)
1778  return res;
1779 
1780  /* First bit is speech/music bit, it differentiates between WMAVoice
1781  * speech samples (the actual codec) and WMAVoice music samples, which
1782  * are really WMAPro-in-WMAVoice-superframes. I've never seen those in
1783  * the wild yet. */
1784  if (!get_bits1(gb)) {
1785  avpriv_request_sample(ctx, "WMAPro-in-WMAVoice");
1786  return AVERROR_PATCHWELCOME;
1787  }
1788 
1789  /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */
1790  if (get_bits1(gb)) {
1791  if ((n_samples = get_bits(gb, 12)) > 480) {
1792  av_log(ctx, AV_LOG_ERROR,
1793  "Superframe encodes >480 samples (%d), not allowed\n",
1794  n_samples);
1795  return AVERROR_INVALIDDATA;
1796  }
1797  }
1798  /* Parse LSPs, if global for the superframe (can also be per-frame). */
1799  if (s->has_residual_lsps) {
1800  double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2];
1801 
1802  for (n = 0; n < s->lsps; n++)
1803  prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n];
1804 
1805  if (s->lsps == 10) {
1806  dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1807  } else /* s->lsps == 16 */
1808  dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1809 
1810  for (n = 0; n < s->lsps; n++) {
1811  lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]);
1812  lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]);
1813  lsps[2][n] += mean_lsf[n];
1814  }
1815  for (n = 0; n < 3; n++)
1816  stabilize_lsps(lsps[n], s->lsps);
1817  }
1818 
1819  /* get output buffer */
1820  frame->nb_samples = 480;
1821  if ((res = ff_get_buffer(ctx, frame, 0)) < 0)
1822  return res;
1823  frame->nb_samples = n_samples;
1824  samples = (float *)frame->data[0];
1825 
1826  /* Parse frames, optionally preceded by per-frame (independent) LSPs. */
1827  for (n = 0; n < 3; n++) {
1828  if (!s->has_residual_lsps) {
1829  int m;
1830 
1831  if (s->lsps == 10) {
1832  dequant_lsp10i(gb, lsps[n]);
1833  } else /* s->lsps == 16 */
1834  dequant_lsp16i(gb, lsps[n]);
1835 
1836  for (m = 0; m < s->lsps; m++)
1837  lsps[n][m] += mean_lsf[m];
1838  stabilize_lsps(lsps[n], s->lsps);
1839  }
1840 
1841  if ((res = synth_frame(ctx, gb, n,
1842  &samples[n * MAX_FRAMESIZE],
1843  lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1],
1844  &excitation[s->history_nsamples + n * MAX_FRAMESIZE],
1845  &synth[s->lsps + n * MAX_FRAMESIZE]))) {
1846  *got_frame_ptr = 0;
1847  return res;
1848  }
1849  }
1850 
1851  /* Statistics? FIXME - we don't check for length, a slight overrun
1852  * will be caught by internal buffer padding, and anything else
1853  * will be skipped, not read. */
1854  if (get_bits1(gb)) {
1855  res = get_bits(gb, 4);
1856  skip_bits(gb, 10 * (res + 1));
1857  }
1858 
1859  *got_frame_ptr = 1;
1860 
1861  /* Update history */
1862  memcpy(s->prev_lsps, lsps[2],
1863  s->lsps * sizeof(*s->prev_lsps));
1864  memcpy(s->synth_history, &synth[MAX_SFRAMESIZE],
1865  s->lsps * sizeof(*synth));
1866  memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE],
1867  s->history_nsamples * sizeof(*excitation));
1868  if (s->do_apf)
1869  memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE],
1870  s->history_nsamples * sizeof(*s->zero_exc_pf));
1871 
1872  return 0;
1873 }
1874 
1875 /**
1876  * Parse the packet header at the start of each packet (input data to this
1877  * decoder).
1878  *
1879  * @param s WMA Voice decoding context private data
1880  * @return 1 if not enough bits were available, or 0 on success.
1881  */
1882 static int parse_packet_header(WMAVoiceContext *s)
1883 {
1884  GetBitContext *gb = &s->gb;
1885  unsigned int res;
1886 
1887  if (get_bits_left(gb) < 11)
1888  return 1;
1889  skip_bits(gb, 4); // packet sequence number
1890  s->has_residual_lsps = get_bits1(gb);
1891  do {
1892  res = get_bits(gb, 6); // number of superframes per packet
1893  // (minus first one if there is spillover)
1894  if (get_bits_left(gb) < 6 * (res == 0x3F) + s->spillover_bitsize)
1895  return 1;
1896  } while (res == 0x3F);
1897  s->spillover_nbits = get_bits(gb, s->spillover_bitsize);
1898 
1899  return 0;
1900 }
1901 
1902 /**
1903  * Copy (unaligned) bits from gb/data/size to pb.
1904  *
1905  * @param pb target buffer to copy bits into
1906  * @param data source buffer to copy bits from
1907  * @param size size of the source data, in bytes
1908  * @param gb bit I/O context specifying the current position in the source.
1909  * data. This function might use this to align the bit position to
1910  * a whole-byte boundary before calling #avpriv_copy_bits() on aligned
1911  * source data
1912  * @param nbits the amount of bits to copy from source to target
1913  *
1914  * @note after calling this function, the current position in the input bit
1915  * I/O context is undefined.
1916  */
1917 static void copy_bits(PutBitContext *pb,
1918  const uint8_t *data, int size,
1919  GetBitContext *gb, int nbits)
1920 {
1921  int rmn_bytes, rmn_bits;
1922 
1923  rmn_bits = rmn_bytes = get_bits_left(gb);
1924  if (rmn_bits < nbits)
1925  return;
1926  if (nbits > pb->size_in_bits - put_bits_count(pb))
1927  return;
1928  rmn_bits &= 7; rmn_bytes >>= 3;
1929  if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0)
1930  put_bits(pb, rmn_bits, get_bits(gb, rmn_bits));
1931  avpriv_copy_bits(pb, data + size - rmn_bytes,
1932  FFMIN(nbits - rmn_bits, rmn_bytes << 3));
1933 }
1934 
1935 /**
1936  * Packet decoding: a packet is anything that the (ASF) demuxer contains,
1937  * and we expect that the demuxer / application provides it to us as such
1938  * (else you'll probably get garbage as output). Every packet has a size of
1939  * ctx->block_align bytes, starts with a packet header (see
1940  * #parse_packet_header()), and then a series of superframes. Superframe
1941  * boundaries may exceed packets, i.e. superframes can split data over
1942  * multiple (two) packets.
1943  *
1944  * For more information about frames, see #synth_superframe().
1945  */
1946 static int wmavoice_decode_packet(AVCodecContext *ctx, void *data,
1947  int *got_frame_ptr, AVPacket *avpkt)
1948 {
1949  WMAVoiceContext *s = ctx->priv_data;
1950  GetBitContext *gb = &s->gb;
1951  int size, res, pos;
1952 
1953  /* Packets are sometimes a multiple of ctx->block_align, with a packet
1954  * header at each ctx->block_align bytes. However, FFmpeg's ASF demuxer
1955  * feeds us ASF packets, which may concatenate multiple "codec" packets
1956  * in a single "muxer" packet, so we artificially emulate that by
1957  * capping the packet size at ctx->block_align. */
1958  for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align);
1959  if (!size) {
1960  *got_frame_ptr = 0;
1961  return 0;
1962  }
1963  init_get_bits(&s->gb, avpkt->data, size << 3);
1964 
1965  /* size == ctx->block_align is used to indicate whether we are dealing with
1966  * a new packet or a packet of which we already read the packet header
1967  * previously. */
1968  if (size == ctx->block_align) { // new packet header
1969  if ((res = parse_packet_header(s)) < 0)
1970  return res;
1971 
1972  /* If the packet header specifies a s->spillover_nbits, then we want
1973  * to push out all data of the previous packet (+ spillover) before
1974  * continuing to parse new superframes in the current packet. */
1975  if (s->spillover_nbits > 0) {
1976  if (s->sframe_cache_size > 0) {
1977  int cnt = get_bits_count(gb);
1978  copy_bits(&s->pb, avpkt->data, size, gb, s->spillover_nbits);
1979  flush_put_bits(&s->pb);
1980  s->sframe_cache_size += s->spillover_nbits;
1981  if ((res = synth_superframe(ctx, data, got_frame_ptr)) == 0 &&
1982  *got_frame_ptr) {
1983  cnt += s->spillover_nbits;
1984  s->skip_bits_next = cnt & 7;
1985  return cnt >> 3;
1986  } else
1987  skip_bits_long (gb, s->spillover_nbits - cnt +
1988  get_bits_count(gb)); // resync
1989  } else
1990  skip_bits_long(gb, s->spillover_nbits); // resync
1991  }
1992  } else if (s->skip_bits_next)
1993  skip_bits(gb, s->skip_bits_next);
1994 
1995  /* Try parsing superframes in current packet */
1996  s->sframe_cache_size = 0;
1997  s->skip_bits_next = 0;
1998  pos = get_bits_left(gb);
1999  if ((res = synth_superframe(ctx, data, got_frame_ptr)) < 0) {
2000  return res;
2001  } else if (*got_frame_ptr) {
2002  int cnt = get_bits_count(gb);
2003  s->skip_bits_next = cnt & 7;
2004  return cnt >> 3;
2005  } else if ((s->sframe_cache_size = pos) > 0) {
2006  /* rewind bit reader to start of last (incomplete) superframe... */
2007  init_get_bits(gb, avpkt->data, size << 3);
2008  skip_bits_long(gb, (size << 3) - pos);
2009  av_assert1(get_bits_left(gb) == pos);
2010 
2011  /* ...and cache it for spillover in next packet */
2012  init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE);
2013  copy_bits(&s->pb, avpkt->data, size, gb, s->sframe_cache_size);
2014  // FIXME bad - just copy bytes as whole and add use the
2015  // skip_bits_next field
2016  }
2017 
2018  return size;
2019 }
2020 
2021 static av_cold int wmavoice_decode_end(AVCodecContext *ctx)
2022 {
2023  WMAVoiceContext *s = ctx->priv_data;
2024 
2025  if (s->do_apf) {
2026  ff_rdft_end(&s->rdft);
2027  ff_rdft_end(&s->irdft);
2028  ff_dct_end(&s->dct);
2029  ff_dct_end(&s->dst);
2030  }
2031 
2032  return 0;
2033 }
2034 
2035 static av_cold void wmavoice_flush(AVCodecContext *ctx)
2036 {
2037  WMAVoiceContext *s = ctx->priv_data;
2038  int n;
2039 
2040  s->postfilter_agc = 0;
2041  s->sframe_cache_size = 0;
2042  s->skip_bits_next = 0;
2043  for (n = 0; n < s->lsps; n++)
2044  s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
2045  memset(s->excitation_history, 0,
2046  sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY);
2047  memset(s->synth_history, 0,
2048  sizeof(*s->synth_history) * MAX_LSPS);
2049  memset(s->gain_pred_err, 0,
2050  sizeof(s->gain_pred_err));
2051 
2052  if (s->do_apf) {
2053  memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0,
2054  sizeof(*s->synth_filter_out_buf) * s->lsps);
2055  memset(s->dcf_mem, 0,
2056  sizeof(*s->dcf_mem) * 2);
2057  memset(s->zero_exc_pf, 0,
2058  sizeof(*s->zero_exc_pf) * s->history_nsamples);
2059  memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache));
2060  }
2061 }
2062 
2063 AVCodec ff_wmavoice_decoder = {
2064  .name = "wmavoice",
2065  .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"),
2066  .type = AVMEDIA_TYPE_AUDIO,
2067  .id = AV_CODEC_ID_WMAVOICE,
2068  .priv_data_size = sizeof(WMAVoiceContext),
2069  .init = wmavoice_decode_init,
2070  .init_static_data = wmavoice_init_static_data,
2071  .close = wmavoice_decode_end,
2072  .decode = wmavoice_decode_packet,
2073  .capabilities = CODEC_CAP_SUBFRAMES | CODEC_CAP_DR1,
2074  .flush = wmavoice_flush,
2075 };
Generated on Thu Jul 31 2014 19:21:17 for FFmpeg by   doxygen 1.8.2