6f3a6b2372a2dca7b67e7c3f4708092ef214545e
[libav.git] / libavcodec / wmavoice.c
1 /*
2 * Windows Media Audio Voice decoder.
3 * Copyright (c) 2009 Ronald S. Bultje
4 *
5 * This file is part of Libav.
6 *
7 * Libav 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 * Libav 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 Libav; 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 #include "avcodec.h"
30 #include "get_bits.h"
31 #include "put_bits.h"
32 #include "wmavoice_data.h"
33 #include "celp_math.h"
34 #include "celp_filters.h"
35 #include "acelp_vectors.h"
36 #include "acelp_filters.h"
37 #include "lsp.h"
38 #include "libavutil/lzo.h"
39 #include "dct.h"
40 #include "rdft.h"
41 #include "sinewin.h"
42
43 #define MAX_BLOCKS 8 ///< maximum number of blocks per frame
44 #define MAX_LSPS 16 ///< maximum filter order
45 #define MAX_LSPS_ALIGN16 16 ///< same as #MAX_LSPS; needs to be multiple
46 ///< of 16 for ASM input buffer alignment
47 #define MAX_FRAMES 3 ///< maximum number of frames per superframe
48 #define MAX_FRAMESIZE 160 ///< maximum number of samples per frame
49 #define MAX_SIGNAL_HISTORY 416 ///< maximum excitation signal history
50 #define MAX_SFRAMESIZE (MAX_FRAMESIZE * MAX_FRAMES)
51 ///< maximum number of samples per superframe
52 #define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that
53 ///< was split over two packets
54 #define VLC_NBITS 6 ///< number of bits to read per VLC iteration
55
56 /**
57 * Frame type VLC coding.
58 */
59 static VLC frame_type_vlc;
60
61 /**
62 * Adaptive codebook types.
63 */
64 enum {
65 ACB_TYPE_NONE = 0, ///< no adaptive codebook (only hardcoded fixed)
66 ACB_TYPE_ASYMMETRIC = 1, ///< adaptive codebook with per-frame pitch, which
67 ///< we interpolate to get a per-sample pitch.
68 ///< Signal is generated using an asymmetric sinc
69 ///< window function
70 ///< @note see #wmavoice_ipol1_coeffs
71 ACB_TYPE_HAMMING = 2 ///< Per-block pitch with signal generation using
72 ///< a Hamming sinc window function
73 ///< @note see #wmavoice_ipol2_coeffs
74 };
75
76 /**
77 * Fixed codebook types.
78 */
79 enum {
80 FCB_TYPE_SILENCE = 0, ///< comfort noise during silence
81 ///< generated from a hardcoded (fixed) codebook
82 ///< with per-frame (low) gain values
83 FCB_TYPE_HARDCODED = 1, ///< hardcoded (fixed) codebook with per-block
84 ///< gain values
85 FCB_TYPE_AW_PULSES = 2, ///< Pitch-adaptive window (AW) pulse signals,
86 ///< used in particular for low-bitrate streams
87 FCB_TYPE_EXC_PULSES = 3, ///< Innovation (fixed) codebook pulse sets in
88 ///< combinations of either single pulses or
89 ///< pulse pairs
90 };
91
92 /**
93 * Description of frame types.
94 */
95 static const struct frame_type_desc {
96 uint8_t n_blocks; ///< amount of blocks per frame (each block
97 ///< (contains 160/#n_blocks samples)
98 uint8_t log_n_blocks; ///< log2(#n_blocks)
99 uint8_t acb_type; ///< Adaptive codebook type (ACB_TYPE_*)
100 uint8_t fcb_type; ///< Fixed codebook type (FCB_TYPE_*)
101 uint8_t dbl_pulses; ///< how many pulse vectors have pulse pairs
102 ///< (rather than just one single pulse)
103 ///< only if #fcb_type == #FCB_TYPE_EXC_PULSES
104 uint16_t frame_size; ///< the amount of bits that make up the block
105 ///< data (per frame)
106 } frame_descs[17] = {
107 { 1, 0, ACB_TYPE_NONE, FCB_TYPE_SILENCE, 0, 0 },
108 { 2, 1, ACB_TYPE_NONE, FCB_TYPE_HARDCODED, 0, 28 },
109 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_AW_PULSES, 0, 46 },
110 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 80 },
111 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 104 },
112 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 0, 108 },
113 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 132 },
114 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 168 },
115 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 64 },
116 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 80 },
117 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 104 },
118 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 108 },
119 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 132 },
120 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 168 },
121 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 176 },
122 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 208 },
123 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 256 }
124 };
125
126 /**
127 * WMA Voice decoding context.
128 */
129 typedef struct {
130 /**
131 * @name Global values specified in the stream header / extradata or used all over.
132 * @{
133 */
134 AVFrame frame;
135 GetBitContext gb; ///< packet bitreader. During decoder init,
136 ///< it contains the extradata from the
137 ///< demuxer. During decoding, it contains
138 ///< packet data.
139 int8_t vbm_tree[25]; ///< converts VLC codes to frame type
140
141 int spillover_bitsize; ///< number of bits used to specify
142 ///< #spillover_nbits in the packet header
143 ///< = ceil(log2(ctx->block_align << 3))
144 int history_nsamples; ///< number of samples in history for signal
145 ///< prediction (through ACB)
146
147 /* postfilter specific values */
148 int do_apf; ///< whether to apply the averaged
149 ///< projection filter (APF)
150 int denoise_strength; ///< strength of denoising in Wiener filter
151 ///< [0-11]
152 int denoise_tilt_corr; ///< Whether to apply tilt correction to the
153 ///< Wiener filter coefficients (postfilter)
154 int dc_level; ///< Predicted amount of DC noise, based
155 ///< on which a DC removal filter is used
156
157 int lsps; ///< number of LSPs per frame [10 or 16]
158 int lsp_q_mode; ///< defines quantizer defaults [0, 1]
159 int lsp_def_mode; ///< defines different sets of LSP defaults
160 ///< [0, 1]
161 int frame_lsp_bitsize; ///< size (in bits) of LSPs, when encoded
162 ///< per-frame (independent coding)
163 int sframe_lsp_bitsize; ///< size (in bits) of LSPs, when encoded
164 ///< per superframe (residual coding)
165
166 int min_pitch_val; ///< base value for pitch parsing code
167 int max_pitch_val; ///< max value + 1 for pitch parsing
168 int pitch_nbits; ///< number of bits used to specify the
169 ///< pitch value in the frame header
170 int block_pitch_nbits; ///< number of bits used to specify the
171 ///< first block's pitch value
172 int block_pitch_range; ///< range of the block pitch
173 int block_delta_pitch_nbits; ///< number of bits used to specify the
174 ///< delta pitch between this and the last
175 ///< block's pitch value, used in all but
176 ///< first block
177 int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is
178 ///< from -this to +this-1)
179 uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale
180 ///< conversion
181
182 /**
183 * @}
184 *
185 * @name Packet values specified in the packet header or related to a packet.
186 *
187 * A packet is considered to be a single unit of data provided to this
188 * decoder by the demuxer.
189 * @{
190 */
191 int spillover_nbits; ///< number of bits of the previous packet's
192 ///< last superframe preceeding this
193 ///< packet's first full superframe (useful
194 ///< for re-synchronization also)
195 int has_residual_lsps; ///< if set, superframes contain one set of
196 ///< LSPs that cover all frames, encoded as
197 ///< independent and residual LSPs; if not
198 ///< set, each frame contains its own, fully
199 ///< independent, LSPs
200 int skip_bits_next; ///< number of bits to skip at the next call
201 ///< to #wmavoice_decode_packet() (since
202 ///< they're part of the previous superframe)
203
204 uint8_t sframe_cache[SFRAME_CACHE_MAXSIZE + FF_INPUT_BUFFER_PADDING_SIZE];
205 ///< cache for superframe data split over
206 ///< multiple packets
207 int sframe_cache_size; ///< set to >0 if we have data from an
208 ///< (incomplete) superframe from a previous
209 ///< packet that spilled over in the current
210 ///< packet; specifies the amount of bits in
211 ///< #sframe_cache
212 PutBitContext pb; ///< bitstream writer for #sframe_cache
213
214 /**
215 * @}
216 *
217 * @name Frame and superframe values
218 * Superframe and frame data - these can change from frame to frame,
219 * although some of them do in that case serve as a cache / history for
220 * the next frame or superframe.
221 * @{
222 */
223 double prev_lsps[MAX_LSPS]; ///< LSPs of the last frame of the previous
224 ///< superframe
225 int last_pitch_val; ///< pitch value of the previous frame
226 int last_acb_type; ///< frame type [0-2] of the previous frame
227 int pitch_diff_sh16; ///< ((cur_pitch_val - #last_pitch_val)
228 ///< << 16) / #MAX_FRAMESIZE
229 float silence_gain; ///< set for use in blocks if #ACB_TYPE_NONE
230
231 int aw_idx_is_ext; ///< whether the AW index was encoded in
232 ///< 8 bits (instead of 6)
233 int aw_pulse_range; ///< the range over which #aw_pulse_set1()
234 ///< can apply the pulse, relative to the
235 ///< value in aw_first_pulse_off. The exact
236 ///< position of the first AW-pulse is within
237 ///< [pulse_off, pulse_off + this], and
238 ///< depends on bitstream values; [16 or 24]
239 int aw_n_pulses[2]; ///< number of AW-pulses in each block; note
240 ///< that this number can be negative (in
241 ///< which case it basically means "zero")
242 int aw_first_pulse_off[2]; ///< index of first sample to which to
243 ///< apply AW-pulses, or -0xff if unset
244 int aw_next_pulse_off_cache; ///< the position (relative to start of the
245 ///< second block) at which pulses should
246 ///< start to be positioned, serves as a
247 ///< cache for pitch-adaptive window pulses
248 ///< between blocks
249
250 int frame_cntr; ///< current frame index [0 - 0xFFFE]; is
251 ///< only used for comfort noise in #pRNG()
252 float gain_pred_err[6]; ///< cache for gain prediction
253 float excitation_history[MAX_SIGNAL_HISTORY];
254 ///< cache of the signal of previous
255 ///< superframes, used as a history for
256 ///< signal generation
257 float synth_history[MAX_LSPS]; ///< see #excitation_history
258 /**
259 * @}
260 *
261 * @name Postfilter values
262 *
263 * Variables used for postfilter implementation, mostly history for
264 * smoothing and so on, and context variables for FFT/iFFT.
265 * @{
266 */
267 RDFTContext rdft, irdft; ///< contexts for FFT-calculation in the
268 ///< postfilter (for denoise filter)
269 DCTContext dct, dst; ///< contexts for phase shift (in Hilbert
270 ///< transform, part of postfilter)
271 float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi]
272 ///< range
273 float postfilter_agc; ///< gain control memory, used in
274 ///< #adaptive_gain_control()
275 float dcf_mem[2]; ///< DC filter history
276 float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE];
277 ///< zero filter output (i.e. excitation)
278 ///< by postfilter
279 float denoise_filter_cache[MAX_FRAMESIZE];
280 int denoise_filter_cache_size; ///< samples in #denoise_filter_cache
281 DECLARE_ALIGNED(32, float, tilted_lpcs_pf)[0x80];
282 ///< aligned buffer for LPC tilting
283 DECLARE_ALIGNED(32, float, denoise_coeffs_pf)[0x80];
284 ///< aligned buffer for denoise coefficients
285 DECLARE_ALIGNED(32, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16];
286 ///< aligned buffer for postfilter speech
287 ///< synthesis
288 /**
289 * @}
290 */
291 } WMAVoiceContext;
292
293 /**
294 * Set up the variable bit mode (VBM) tree from container extradata.
295 * @param gb bit I/O context.
296 * The bit context (s->gb) should be loaded with byte 23-46 of the
297 * container extradata (i.e. the ones containing the VBM tree).
298 * @param vbm_tree pointer to array to which the decoded VBM tree will be
299 * written.
300 * @return 0 on success, <0 on error.
301 */
302 static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25])
303 {
304 static const uint8_t bits[] = {
305 2, 2, 2, 4, 4, 4,
306 6, 6, 6, 8, 8, 8,
307 10, 10, 10, 12, 12, 12,
308 14, 14, 14, 14
309 };
310 static const uint16_t codes[] = {
311 0x0000, 0x0001, 0x0002, // 00/01/10
312 0x000c, 0x000d, 0x000e, // 11+00/01/10
313 0x003c, 0x003d, 0x003e, // 1111+00/01/10
314 0x00fc, 0x00fd, 0x00fe, // 111111+00/01/10
315 0x03fc, 0x03fd, 0x03fe, // 11111111+00/01/10
316 0x0ffc, 0x0ffd, 0x0ffe, // 1111111111+00/01/10
317 0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx
318 };
319 int cntr[8], n, res;
320
321 memset(vbm_tree, 0xff, sizeof(vbm_tree[0]) * 25);
322 memset(cntr, 0, sizeof(cntr));
323 for (n = 0; n < 17; n++) {
324 res = get_bits(gb, 3);
325 if (cntr[res] > 3) // should be >= 3 + (res == 7))
326 return -1;
327 vbm_tree[res * 3 + cntr[res]++] = n;
328 }
329 INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits),
330 bits, 1, 1, codes, 2, 2, 132);
331 return 0;
332 }
333
334 /**
335 * Set up decoder with parameters from demuxer (extradata etc.).
336 */
337 static av_cold int wmavoice_decode_init(AVCodecContext *ctx)
338 {
339 int n, flags, pitch_range, lsp16_flag;
340 WMAVoiceContext *s = ctx->priv_data;
341
342 /**
343 * Extradata layout:
344 * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c),
345 * - byte 19-22: flags field (annoyingly in LE; see below for known
346 * values),
347 * - byte 23-46: variable bitmode tree (really just 17 * 3 bits,
348 * rest is 0).
349 */
350 if (ctx->extradata_size != 46) {
351 av_log(ctx, AV_LOG_ERROR,
352 "Invalid extradata size %d (should be 46)\n",
353 ctx->extradata_size);
354 return -1;
355 }
356 flags = AV_RL32(ctx->extradata + 18);
357 s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align);
358 s->do_apf = flags & 0x1;
359 if (s->do_apf) {
360 ff_rdft_init(&s->rdft, 7, DFT_R2C);
361 ff_rdft_init(&s->irdft, 7, IDFT_C2R);
362 ff_dct_init(&s->dct, 6, DCT_I);
363 ff_dct_init(&s->dst, 6, DST_I);
364
365 ff_sine_window_init(s->cos, 256);
366 memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0]));
367 for (n = 0; n < 255; n++) {
368 s->sin[n] = -s->sin[510 - n];
369 s->cos[510 - n] = s->cos[n];
370 }
371 }
372 s->denoise_strength = (flags >> 2) & 0xF;
373 if (s->denoise_strength >= 12) {
374 av_log(ctx, AV_LOG_ERROR,
375 "Invalid denoise filter strength %d (max=11)\n",
376 s->denoise_strength);
377 return -1;
378 }
379 s->denoise_tilt_corr = !!(flags & 0x40);
380 s->dc_level = (flags >> 7) & 0xF;
381 s->lsp_q_mode = !!(flags & 0x2000);
382 s->lsp_def_mode = !!(flags & 0x4000);
383 lsp16_flag = flags & 0x1000;
384 if (lsp16_flag) {
385 s->lsps = 16;
386 s->frame_lsp_bitsize = 34;
387 s->sframe_lsp_bitsize = 60;
388 } else {
389 s->lsps = 10;
390 s->frame_lsp_bitsize = 24;
391 s->sframe_lsp_bitsize = 48;
392 }
393 for (n = 0; n < s->lsps; n++)
394 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
395
396 init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3);
397 if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) {
398 av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n");
399 return -1;
400 }
401
402 s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8;
403 s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8;
404 pitch_range = s->max_pitch_val - s->min_pitch_val;
405 if (pitch_range <= 0) {
406 av_log(ctx, AV_LOG_ERROR, "Invalid pitch range; broken extradata?\n");
407 return -1;
408 }
409 s->pitch_nbits = av_ceil_log2(pitch_range);
410 s->last_pitch_val = 40;
411 s->last_acb_type = ACB_TYPE_NONE;
412 s->history_nsamples = s->max_pitch_val + 8;
413
414 if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) {
415 int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8,
416 max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8;
417
418 av_log(ctx, AV_LOG_ERROR,
419 "Unsupported samplerate %d (min=%d, max=%d)\n",
420 ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz
421
422 return -1;
423 }
424
425 s->block_conv_table[0] = s->min_pitch_val;
426 s->block_conv_table[1] = (pitch_range * 25) >> 6;
427 s->block_conv_table[2] = (pitch_range * 44) >> 6;
428 s->block_conv_table[3] = s->max_pitch_val - 1;
429 s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF;
430 if (s->block_delta_pitch_hrange <= 0) {
431 av_log(ctx, AV_LOG_ERROR, "Invalid delta pitch hrange; broken extradata?\n");
432 return -1;
433 }
434 s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange);
435 s->block_pitch_range = s->block_conv_table[2] +
436 s->block_conv_table[3] + 1 +
437 2 * (s->block_conv_table[1] - 2 * s->min_pitch_val);
438 s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range);
439
440 ctx->sample_fmt = AV_SAMPLE_FMT_FLT;
441
442 avcodec_get_frame_defaults(&s->frame);
443 ctx->coded_frame = &s->frame;
444
445 return 0;
446 }
447
448 /**
449 * @name Postfilter functions
450 * Postfilter functions (gain control, wiener denoise filter, DC filter,
451 * kalman smoothening, plus surrounding code to wrap it)
452 * @{
453 */
454 /**
455 * Adaptive gain control (as used in postfilter).
456 *
457 * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except
458 * that the energy here is calculated using sum(abs(...)), whereas the
459 * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)).
460 *
461 * @param out output buffer for filtered samples
462 * @param in input buffer containing the samples as they are after the
463 * postfilter steps so far
464 * @param speech_synth input buffer containing speech synth before postfilter
465 * @param size input buffer size
466 * @param alpha exponential filter factor
467 * @param gain_mem pointer to filter memory (single float)
468 */
469 static void adaptive_gain_control(float *out, const float *in,
470 const float *speech_synth,
471 int size, float alpha, float *gain_mem)
472 {
473 int i;
474 float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor;
475 float mem = *gain_mem;
476
477 for (i = 0; i < size; i++) {
478 speech_energy += fabsf(speech_synth[i]);
479 postfilter_energy += fabsf(in[i]);
480 }
481 gain_scale_factor = (1.0 - alpha) * speech_energy / postfilter_energy;
482
483 for (i = 0; i < size; i++) {
484 mem = alpha * mem + gain_scale_factor;
485 out[i] = in[i] * mem;
486 }
487
488 *gain_mem = mem;
489 }
490
491 /**
492 * Kalman smoothing function.
493 *
494 * This function looks back pitch +/- 3 samples back into history to find
495 * the best fitting curve (that one giving the optimal gain of the two
496 * signals, i.e. the highest dot product between the two), and then
497 * uses that signal history to smoothen the output of the speech synthesis
498 * filter.
499 *
500 * @param s WMA Voice decoding context
501 * @param pitch pitch of the speech signal
502 * @param in input speech signal
503 * @param out output pointer for smoothened signal
504 * @param size input/output buffer size
505 *
506 * @returns -1 if no smoothening took place, e.g. because no optimal
507 * fit could be found, or 0 on success.
508 */
509 static int kalman_smoothen(WMAVoiceContext *s, int pitch,
510 const float *in, float *out, int size)
511 {
512 int n;
513 float optimal_gain = 0, dot;
514 const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)],
515 *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)],
516 *best_hist_ptr;
517
518 /* find best fitting point in history */
519 do {
520 dot = ff_dot_productf(in, ptr, size);
521 if (dot > optimal_gain) {
522 optimal_gain = dot;
523 best_hist_ptr = ptr;
524 }
525 } while (--ptr >= end);
526
527 if (optimal_gain <= 0)
528 return -1;
529 dot = ff_dot_productf(best_hist_ptr, best_hist_ptr, size);
530 if (dot <= 0) // would be 1.0
531 return -1;
532
533 if (optimal_gain <= dot) {
534 dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000
535 } else
536 dot = 0.625;
537
538 /* actual smoothing */
539 for (n = 0; n < size; n++)
540 out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]);
541
542 return 0;
543 }
544
545 /**
546 * Get the tilt factor of a formant filter from its transfer function
547 * @see #tilt_factor() in amrnbdec.c, which does essentially the same,
548 * but somehow (??) it does a speech synthesis filter in the
549 * middle, which is missing here
550 *
551 * @param lpcs LPC coefficients
552 * @param n_lpcs Size of LPC buffer
553 * @returns the tilt factor
554 */
555 static float tilt_factor(const float *lpcs, int n_lpcs)
556 {
557 float rh0, rh1;
558
559 rh0 = 1.0 + ff_dot_productf(lpcs, lpcs, n_lpcs);
560 rh1 = lpcs[0] + ff_dot_productf(lpcs, &lpcs[1], n_lpcs - 1);
561
562 return rh1 / rh0;
563 }
564
565 /**
566 * Derive denoise filter coefficients (in real domain) from the LPCs.
567 */
568 static void calc_input_response(WMAVoiceContext *s, float *lpcs,
569 int fcb_type, float *coeffs, int remainder)
570 {
571 float last_coeff, min = 15.0, max = -15.0;
572 float irange, angle_mul, gain_mul, range, sq;
573 int n, idx;
574
575 /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */
576 s->rdft.rdft_calc(&s->rdft, lpcs);
577 #define log_range(var, assign) do { \
578 float tmp = log10f(assign); var = tmp; \
579 max = FFMAX(max, tmp); min = FFMIN(min, tmp); \
580 } while (0)
581 log_range(last_coeff, lpcs[1] * lpcs[1]);
582 for (n = 1; n < 64; n++)
583 log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] +
584 lpcs[n * 2 + 1] * lpcs[n * 2 + 1]);
585 log_range(lpcs[0], lpcs[0] * lpcs[0]);
586 #undef log_range
587 range = max - min;
588 lpcs[64] = last_coeff;
589
590 /* Now, use this spectrum to pick out these frequencies with higher
591 * (relative) power/energy (which we then take to be "not noise"),
592 * and set up a table (still in lpc[]) of (relative) gains per frequency.
593 * These frequencies will be maintained, while others ("noise") will be
594 * decreased in the filter output. */
595 irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63]
596 gain_mul = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) :
597 (5.0 / 14.7));
598 angle_mul = gain_mul * (8.0 * M_LN10 / M_PI);
599 for (n = 0; n <= 64; n++) {
600 float pwr;
601
602 idx = FFMAX(0, lrint((max - lpcs[n]) * irange) - 1);
603 pwr = wmavoice_denoise_power_table[s->denoise_strength][idx];
604 lpcs[n] = angle_mul * pwr;
605
606 /* 70.57 =~ 1/log10(1.0331663) */
607 idx = (pwr * gain_mul - 0.0295) * 70.570526123;
608 if (idx > 127) { // fallback if index falls outside table range
609 coeffs[n] = wmavoice_energy_table[127] *
610 powf(1.0331663, idx - 127);
611 } else
612 coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)];
613 }
614
615 /* calculate the Hilbert transform of the gains, which we do (since this
616 * is a sinus input) by doing a phase shift (in theory, H(sin())=cos()).
617 * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the
618 * "moment" of the LPCs in this filter. */
619 s->dct.dct_calc(&s->dct, lpcs);
620 s->dst.dct_calc(&s->dst, lpcs);
621
622 /* Split out the coefficient indexes into phase/magnitude pairs */
623 idx = 255 + av_clip(lpcs[64], -255, 255);
624 coeffs[0] = coeffs[0] * s->cos[idx];
625 idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255);
626 last_coeff = coeffs[64] * s->cos[idx];
627 for (n = 63;; n--) {
628 idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255);
629 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
630 coeffs[n * 2] = coeffs[n] * s->cos[idx];
631
632 if (!--n) break;
633
634 idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255);
635 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
636 coeffs[n * 2] = coeffs[n] * s->cos[idx];
637 }
638 coeffs[1] = last_coeff;
639
640 /* move into real domain */
641 s->irdft.rdft_calc(&s->irdft, coeffs);
642
643 /* tilt correction and normalize scale */
644 memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder));
645 if (s->denoise_tilt_corr) {
646 float tilt_mem = 0;
647
648 coeffs[remainder - 1] = 0;
649 ff_tilt_compensation(&tilt_mem,
650 -1.8 * tilt_factor(coeffs, remainder - 1),
651 coeffs, remainder);
652 }
653 sq = (1.0 / 64.0) * sqrtf(1 / ff_dot_productf(coeffs, coeffs, remainder));
654 for (n = 0; n < remainder; n++)
655 coeffs[n] *= sq;
656 }
657
658 /**
659 * This function applies a Wiener filter on the (noisy) speech signal as
660 * a means to denoise it.
661 *
662 * - take RDFT of LPCs to get the power spectrum of the noise + speech;
663 * - using this power spectrum, calculate (for each frequency) the Wiener
664 * filter gain, which depends on the frequency power and desired level
665 * of noise subtraction (when set too high, this leads to artifacts)
666 * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse
667 * of 4-8kHz);
668 * - by doing a phase shift, calculate the Hilbert transform of this array
669 * of per-frequency filter-gains to get the filtering coefficients;
670 * - smoothen/normalize/de-tilt these filter coefficients as desired;
671 * - take RDFT of noisy sound, apply the coefficients and take its IRDFT
672 * to get the denoised speech signal;
673 * - the leftover (i.e. output of the IRDFT on denoised speech data beyond
674 * the frame boundary) are saved and applied to subsequent frames by an
675 * overlap-add method (otherwise you get clicking-artifacts).
676 *
677 * @param s WMA Voice decoding context
678 * @param fcb_type Frame (codebook) type
679 * @param synth_pf input: the noisy speech signal, output: denoised speech
680 * data; should be 16-byte aligned (for ASM purposes)
681 * @param size size of the speech data
682 * @param lpcs LPCs used to synthesize this frame's speech data
683 */
684 static void wiener_denoise(WMAVoiceContext *s, int fcb_type,
685 float *synth_pf, int size,
686 const float *lpcs)
687 {
688 int remainder, lim, n;
689
690 if (fcb_type != FCB_TYPE_SILENCE) {
691 float *tilted_lpcs = s->tilted_lpcs_pf,
692 *coeffs = s->denoise_coeffs_pf, tilt_mem = 0;
693
694 tilted_lpcs[0] = 1.0;
695 memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps);
696 memset(&tilted_lpcs[s->lsps + 1], 0,
697 sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1));
698 ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps),
699 tilted_lpcs, s->lsps + 2);
700
701 /* The IRDFT output (127 samples for 7-bit filter) beyond the frame
702 * size is applied to the next frame. All input beyond this is zero,
703 * and thus all output beyond this will go towards zero, hence we can
704 * limit to min(size-1, 127-size) as a performance consideration. */
705 remainder = FFMIN(127 - size, size - 1);
706 calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder);
707
708 /* apply coefficients (in frequency spectrum domain), i.e. complex
709 * number multiplication */
710 memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size));
711 s->rdft.rdft_calc(&s->rdft, synth_pf);
712 s->rdft.rdft_calc(&s->rdft, coeffs);
713 synth_pf[0] *= coeffs[0];
714 synth_pf[1] *= coeffs[1];
715 for (n = 1; n < 64; n++) {
716 float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1];
717 synth_pf[n * 2] = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1];
718 synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1];
719 }
720 s->irdft.rdft_calc(&s->irdft, synth_pf);
721 }
722
723 /* merge filter output with the history of previous runs */
724 if (s->denoise_filter_cache_size) {
725 lim = FFMIN(s->denoise_filter_cache_size, size);
726 for (n = 0; n < lim; n++)
727 synth_pf[n] += s->denoise_filter_cache[n];
728 s->denoise_filter_cache_size -= lim;
729 memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size],
730 sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size);
731 }
732
733 /* move remainder of filter output into a cache for future runs */
734 if (fcb_type != FCB_TYPE_SILENCE) {
735 lim = FFMIN(remainder, s->denoise_filter_cache_size);
736 for (n = 0; n < lim; n++)
737 s->denoise_filter_cache[n] += synth_pf[size + n];
738 if (lim < remainder) {
739 memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim],
740 sizeof(s->denoise_filter_cache[0]) * (remainder - lim));
741 s->denoise_filter_cache_size = remainder;
742 }
743 }
744 }
745
746 /**
747 * Averaging projection filter, the postfilter used in WMAVoice.
748 *
749 * This uses the following steps:
750 * - A zero-synthesis filter (generate excitation from synth signal)
751 * - Kalman smoothing on excitation, based on pitch
752 * - Re-synthesized smoothened output
753 * - Iterative Wiener denoise filter
754 * - Adaptive gain filter
755 * - DC filter
756 *
757 * @param s WMAVoice decoding context
758 * @param synth Speech synthesis output (before postfilter)
759 * @param samples Output buffer for filtered samples
760 * @param size Buffer size of synth & samples
761 * @param lpcs Generated LPCs used for speech synthesis
762 * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned)
763 * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses)
764 * @param pitch Pitch of the input signal
765 */
766 static void postfilter(WMAVoiceContext *s, const float *synth,
767 float *samples, int size,
768 const float *lpcs, float *zero_exc_pf,
769 int fcb_type, int pitch)
770 {
771 float synth_filter_in_buf[MAX_FRAMESIZE / 2],
772 *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16],
773 *synth_filter_in = zero_exc_pf;
774
775 assert(size <= MAX_FRAMESIZE / 2);
776
777 /* generate excitation from input signal */
778 ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps);
779
780 if (fcb_type >= FCB_TYPE_AW_PULSES &&
781 !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size))
782 synth_filter_in = synth_filter_in_buf;
783
784 /* re-synthesize speech after smoothening, and keep history */
785 ff_celp_lp_synthesis_filterf(synth_pf, lpcs,
786 synth_filter_in, size, s->lsps);
787 memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps],
788 sizeof(synth_pf[0]) * s->lsps);
789
790 wiener_denoise(s, fcb_type, synth_pf, size, lpcs);
791
792 adaptive_gain_control(samples, synth_pf, synth, size, 0.99,
793 &s->postfilter_agc);
794
795 if (s->dc_level > 8) {
796 /* remove ultra-low frequency DC noise / highpass filter;
797 * coefficients are identical to those used in SIPR decoding,
798 * and very closely resemble those used in AMR-NB decoding. */
799 ff_acelp_apply_order_2_transfer_function(samples, samples,
800 (const float[2]) { -1.99997, 1.0 },
801 (const float[2]) { -1.9330735188, 0.93589198496 },
802 0.93980580475, s->dcf_mem, size);
803 }
804 }
805 /**
806 * @}
807 */
808
809 /**
810 * Dequantize LSPs
811 * @param lsps output pointer to the array that will hold the LSPs
812 * @param num number of LSPs to be dequantized
813 * @param values quantized values, contains n_stages values
814 * @param sizes range (i.e. max value) of each quantized value
815 * @param n_stages number of dequantization runs
816 * @param table dequantization table to be used
817 * @param mul_q LSF multiplier
818 * @param base_q base (lowest) LSF values
819 */
820 static void dequant_lsps(double *lsps, int num,
821 const uint16_t *values,
822 const uint16_t *sizes,
823 int n_stages, const uint8_t *table,
824 const double *mul_q,
825 const double *base_q)
826 {
827 int n, m;
828
829 memset(lsps, 0, num * sizeof(*lsps));
830 for (n = 0; n < n_stages; n++) {
831 const uint8_t *t_off = &table[values[n] * num];
832 double base = base_q[n], mul = mul_q[n];
833
834 for (m = 0; m < num; m++)
835 lsps[m] += base + mul * t_off[m];
836
837 table += sizes[n] * num;
838 }
839 }
840
841 /**
842 * @name LSP dequantization routines
843 * LSP dequantization routines, for 10/16LSPs and independent/residual coding.
844 * @note we assume enough bits are available, caller should check.
845 * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;
846 * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits.
847 * @{
848 */
849 /**
850 * Parse 10 independently-coded LSPs.
851 */
852 static void dequant_lsp10i(GetBitContext *gb, double *lsps)
853 {
854 static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 };
855 static const double mul_lsf[4] = {
856 5.2187144800e-3, 1.4626986422e-3,
857 9.6179549166e-4, 1.1325736225e-3
858 };
859 static const double base_lsf[4] = {
860 M_PI * -2.15522e-1, M_PI * -6.1646e-2,
861 M_PI * -3.3486e-2, M_PI * -5.7408e-2
862 };
863 uint16_t v[4];
864
865 v[0] = get_bits(gb, 8);
866 v[1] = get_bits(gb, 6);
867 v[2] = get_bits(gb, 5);
868 v[3] = get_bits(gb, 5);
869
870 dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i,
871 mul_lsf, base_lsf);
872 }
873
874 /**
875 * Parse 10 independently-coded LSPs, and then derive the tables to
876 * generate LSPs for the other frames from them (residual coding).
877 */
878 static void dequant_lsp10r(GetBitContext *gb,
879 double *i_lsps, const double *old,
880 double *a1, double *a2, int q_mode)
881 {
882 static const uint16_t vec_sizes[3] = { 128, 64, 64 };
883 static const double mul_lsf[3] = {
884 2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3
885 };
886 static const double base_lsf[3] = {
887 M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2
888 };
889 const float (*ipol_tab)[2][10] = q_mode ?
890 wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a;
891 uint16_t interpol, v[3];
892 int n;
893
894 dequant_lsp10i(gb, i_lsps);
895
896 interpol = get_bits(gb, 5);
897 v[0] = get_bits(gb, 7);
898 v[1] = get_bits(gb, 6);
899 v[2] = get_bits(gb, 6);
900
901 for (n = 0; n < 10; n++) {
902 double delta = old[n] - i_lsps[n];
903 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
904 a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
905 }
906
907 dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r,
908 mul_lsf, base_lsf);
909 }
910
911 /**
912 * Parse 16 independently-coded LSPs.
913 */
914 static void dequant_lsp16i(GetBitContext *gb, double *lsps)
915 {
916 static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 };
917 static const double mul_lsf[5] = {
918 3.3439586280e-3, 6.9908173703e-4,
919 3.3216608306e-3, 1.0334960326e-3,
920 3.1899104283e-3
921 };
922 static const double base_lsf[5] = {
923 M_PI * -1.27576e-1, M_PI * -2.4292e-2,
924 M_PI * -1.28094e-1, M_PI * -3.2128e-2,
925 M_PI * -1.29816e-1
926 };
927 uint16_t v[5];
928
929 v[0] = get_bits(gb, 8);
930 v[1] = get_bits(gb, 6);
931 v[2] = get_bits(gb, 7);
932 v[3] = get_bits(gb, 6);
933 v[4] = get_bits(gb, 7);
934
935 dequant_lsps( lsps, 5, v, vec_sizes, 2,
936 wmavoice_dq_lsp16i1, mul_lsf, base_lsf);
937 dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2,
938 wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]);
939 dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1,
940 wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]);
941 }
942
943 /**
944 * Parse 16 independently-coded LSPs, and then derive the tables to
945 * generate LSPs for the other frames from them (residual coding).
946 */
947 static void dequant_lsp16r(GetBitContext *gb,
948 double *i_lsps, const double *old,
949 double *a1, double *a2, int q_mode)
950 {
951 static const uint16_t vec_sizes[3] = { 128, 128, 128 };
952 static const double mul_lsf[3] = {
953 1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3
954 };
955 static const double base_lsf[3] = {
956 M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2
957 };
958 const float (*ipol_tab)[2][16] = q_mode ?
959 wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a;
960 uint16_t interpol, v[3];
961 int n;
962
963 dequant_lsp16i(gb, i_lsps);
964
965 interpol = get_bits(gb, 5);
966 v[0] = get_bits(gb, 7);
967 v[1] = get_bits(gb, 7);
968 v[2] = get_bits(gb, 7);
969
970 for (n = 0; n < 16; n++) {
971 double delta = old[n] - i_lsps[n];
972 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
973 a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
974 }
975
976 dequant_lsps( a2, 10, v, vec_sizes, 1,
977 wmavoice_dq_lsp16r1, mul_lsf, base_lsf);
978 dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1,
979 wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]);
980 dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1,
981 wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]);
982 }
983
984 /**
985 * @}
986 * @name Pitch-adaptive window coding functions
987 * The next few functions are for pitch-adaptive window coding.
988 * @{
989 */
990 /**
991 * Parse the offset of the first pitch-adaptive window pulses, and
992 * the distribution of pulses between the two blocks in this frame.
993 * @param s WMA Voice decoding context private data
994 * @param gb bit I/O context
995 * @param pitch pitch for each block in this frame
996 */
997 static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb,
998 const int *pitch)
999 {
1000 static const int16_t start_offset[94] = {
1001 -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11,
1002 13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26,
1003 27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43,
1004 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67,
1005 69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91,
1006 93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115,
1007 117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139,
1008 141, 143, 145, 147, 149, 151, 153, 155, 157, 159
1009 };
1010 int bits, offset;
1011
1012 /* position of pulse */
1013 s->aw_idx_is_ext = 0;
1014 if ((bits = get_bits(gb, 6)) >= 54) {
1015 s->aw_idx_is_ext = 1;
1016 bits += (bits - 54) * 3 + get_bits(gb, 2);
1017 }
1018
1019 /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count
1020 * the distribution of the pulses in each block contained in this frame. */
1021 s->aw_pulse_range = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16;
1022 for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ;
1023 s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0];
1024 s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2;
1025 offset += s->aw_n_pulses[0] * pitch[0];
1026 s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1];
1027 s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2;
1028
1029 /* if continuing from a position before the block, reset position to
1030 * start of block (when corrected for the range over which it can be
1031 * spread in aw_pulse_set1()). */
1032 if (start_offset[bits] < MAX_FRAMESIZE / 2) {
1033 while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0)
1034 s->aw_first_pulse_off[1] -= pitch[1];
1035 if (start_offset[bits] < 0)
1036 while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0)
1037 s->aw_first_pulse_off[0] -= pitch[0];
1038 }
1039 }
1040
1041 /**
1042 * Apply second set of pitch-adaptive window pulses.
1043 * @param s WMA Voice decoding context private data
1044 * @param gb bit I/O context
1045 * @param block_idx block index in frame [0, 1]
1046 * @param fcb structure containing fixed codebook vector info
1047 */
1048 static void aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb,
1049 int block_idx, AMRFixed *fcb)
1050 {
1051 uint16_t use_mask_mem[9]; // only 5 are used, rest is padding
1052 uint16_t *use_mask = use_mask_mem + 2;
1053 /* in this function, idx is the index in the 80-bit (+ padding) use_mask
1054 * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits
1055 * of idx are the position of the bit within a particular item in the
1056 * array (0 being the most significant bit, and 15 being the least
1057 * significant bit), and the remainder (>> 4) is the index in the
1058 * use_mask[]-array. This is faster and uses less memory than using a
1059 * 80-byte/80-int array. */
1060 int pulse_off = s->aw_first_pulse_off[block_idx],
1061 pulse_start, n, idx, range, aidx, start_off = 0;
1062
1063 /* set offset of first pulse to within this block */
1064 if (s->aw_n_pulses[block_idx] > 0)
1065 while (pulse_off + s->aw_pulse_range < 1)
1066 pulse_off += fcb->pitch_lag;
1067
1068 /* find range per pulse */
1069 if (s->aw_n_pulses[0] > 0) {
1070 if (block_idx == 0) {
1071 range = 32;
1072 } else /* block_idx = 1 */ {
1073 range = 8;
1074 if (s->aw_n_pulses[block_idx] > 0)
1075 pulse_off = s->aw_next_pulse_off_cache;
1076 }
1077 } else
1078 range = 16;
1079 pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0;
1080
1081 /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,
1082 * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus
1083 * we exclude that range from being pulsed again in this function. */
1084 memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0]));
1085 memset( use_mask, -1, 5 * sizeof(use_mask[0]));
1086 memset(&use_mask[5], 0, 2 * sizeof(use_mask[0]));
1087 if (s->aw_n_pulses[block_idx] > 0)
1088 for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) {
1089 int excl_range = s->aw_pulse_range; // always 16 or 24
1090 uint16_t *use_mask_ptr = &use_mask[idx >> 4];
1091 int first_sh = 16 - (idx & 15);
1092 *use_mask_ptr++ &= 0xFFFFu << first_sh;
1093 excl_range -= first_sh;
1094 if (excl_range >= 16) {
1095 *use_mask_ptr++ = 0;
1096 *use_mask_ptr &= 0xFFFF >> (excl_range - 16);
1097 } else
1098 *use_mask_ptr &= 0xFFFF >> excl_range;
1099 }
1100
1101 /* find the 'aidx'th offset that is not excluded */
1102 aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4);
1103 for (n = 0; n <= aidx; pulse_start++) {
1104 for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ;
1105 if (idx >= MAX_FRAMESIZE / 2) { // find from zero
1106 if (use_mask[0]) idx = 0x0F;
1107 else if (use_mask[1]) idx = 0x1F;
1108 else if (use_mask[2]) idx = 0x2F;
1109 else if (use_mask[3]) idx = 0x3F;
1110 else if (use_mask[4]) idx = 0x4F;
1111 else return;
1112 idx -= av_log2_16bit(use_mask[idx >> 4]);
1113 }
1114 if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) {
1115 use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15));
1116 n++;
1117 start_off = idx;
1118 }
1119 }
1120
1121 fcb->x[fcb->n] = start_off;
1122 fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0;
1123 fcb->n++;
1124
1125 /* set offset for next block, relative to start of that block */
1126 n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag;
1127 s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0;
1128 }
1129
1130 /**
1131 * Apply first set of pitch-adaptive window pulses.
1132 * @param s WMA Voice decoding context private data
1133 * @param gb bit I/O context
1134 * @param block_idx block index in frame [0, 1]
1135 * @param fcb storage location for fixed codebook pulse info
1136 */
1137 static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb,
1138 int block_idx, AMRFixed *fcb)
1139 {
1140 int val = get_bits(gb, 12 - 2 * (s->aw_idx_is_ext && !block_idx));
1141 float v;
1142
1143 if (s->aw_n_pulses[block_idx] > 0) {
1144 int n, v_mask, i_mask, sh, n_pulses;
1145
1146 if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each
1147 n_pulses = 3;
1148 v_mask = 8;
1149 i_mask = 7;
1150 sh = 4;
1151 } else { // 4 pulses, 1:sign + 2:index each
1152 n_pulses = 4;
1153 v_mask = 4;
1154 i_mask = 3;
1155 sh = 3;
1156 }
1157
1158 for (n = n_pulses - 1; n >= 0; n--, val >>= sh) {
1159 fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0;
1160 fcb->x[fcb->n] = (val & i_mask) * n_pulses + n +
1161 s->aw_first_pulse_off[block_idx];
1162 while (fcb->x[fcb->n] < 0)
1163 fcb->x[fcb->n] += fcb->pitch_lag;
1164 if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2)
1165 fcb->n++;
1166 }
1167 } else {
1168 int num2 = (val & 0x1FF) >> 1, delta, idx;
1169
1170 if (num2 < 1 * 79) { delta = 1; idx = num2 + 1; }
1171 else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; }
1172 else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; }
1173 else { delta = 7; idx = num2 + 1 - 3 * 75; }
1174 v = (val & 0x200) ? -1.0 : 1.0;
1175
1176 fcb->no_repeat_mask |= 3 << fcb->n;
1177 fcb->x[fcb->n] = idx - delta;
1178 fcb->y[fcb->n] = v;
1179 fcb->x[fcb->n + 1] = idx;
1180 fcb->y[fcb->n + 1] = (val & 1) ? -v : v;
1181 fcb->n += 2;
1182 }
1183 }
1184
1185 /**
1186 * @}
1187 *
1188 * Generate a random number from frame_cntr and block_idx, which will lief
1189 * in the range [0, 1000 - block_size] (so it can be used as an index in a
1190 * table of size 1000 of which you want to read block_size entries).
1191 *
1192 * @param frame_cntr current frame number
1193 * @param block_num current block index
1194 * @param block_size amount of entries we want to read from a table
1195 * that has 1000 entries
1196 * @return a (non-)random number in the [0, 1000 - block_size] range.
1197 */
1198 static int pRNG(int frame_cntr, int block_num, int block_size)
1199 {
1200 /* array to simplify the calculation of z:
1201 * y = (x % 9) * 5 + 6;
1202 * z = (49995 * x) / y;
1203 * Since y only has 9 values, we can remove the division by using a
1204 * LUT and using FASTDIV-style divisions. For each of the 9 values
1205 * of y, we can rewrite z as:
1206 * z = x * (49995 / y) + x * ((49995 % y) / y)
1207 * In this table, each col represents one possible value of y, the
1208 * first number is 49995 / y, and the second is the FASTDIV variant
1209 * of 49995 % y / y. */
1210 static const unsigned int div_tbl[9][2] = {
1211 { 8332, 3 * 715827883U }, // y = 6
1212 { 4545, 0 * 390451573U }, // y = 11
1213 { 3124, 11 * 268435456U }, // y = 16
1214 { 2380, 15 * 204522253U }, // y = 21
1215 { 1922, 23 * 165191050U }, // y = 26
1216 { 1612, 23 * 138547333U }, // y = 31
1217 { 1388, 27 * 119304648U }, // y = 36
1218 { 1219, 16 * 104755300U }, // y = 41
1219 { 1086, 39 * 93368855U } // y = 46
1220 };
1221 unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr;
1222 if (x >= 0xFFFF) x -= 0xFFFF; // max value of x is 8*1877+0xFFFE=0x13AA6,
1223 // so this is effectively a modulo (%)
1224 y = x - 9 * MULH(477218589, x); // x % 9
1225 z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1]));
1226 // z = x * 49995 / (y * 5 + 6)
1227 return z % (1000 - block_size);
1228 }
1229
1230 /**
1231 * Parse hardcoded signal for a single block.
1232 * @note see #synth_block().
1233 */
1234 static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb,
1235 int block_idx, int size,
1236 const struct frame_type_desc *frame_desc,
1237 float *excitation)
1238 {
1239 float gain;
1240 int n, r_idx;
1241
1242 assert(size <= MAX_FRAMESIZE);
1243
1244 /* Set the offset from which we start reading wmavoice_std_codebook */
1245 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1246 r_idx = pRNG(s->frame_cntr, block_idx, size);
1247 gain = s->silence_gain;
1248 } else /* FCB_TYPE_HARDCODED */ {
1249 r_idx = get_bits(gb, 8);
1250 gain = wmavoice_gain_universal[get_bits(gb, 6)];
1251 }
1252
1253 /* Clear gain prediction parameters */
1254 memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err));
1255
1256 /* Apply gain to hardcoded codebook and use that as excitation signal */
1257 for (n = 0; n < size; n++)
1258 excitation[n] = wmavoice_std_codebook[r_idx + n] * gain;
1259 }
1260
1261 /**
1262 * Parse FCB/ACB signal for a single block.
1263 * @note see #synth_block().
1264 */
1265 static void synth_block_fcb_acb(WMAVoiceContext *s, GetBitContext *gb,
1266 int block_idx, int size,
1267 int block_pitch_sh2,
1268 const struct frame_type_desc *frame_desc,
1269 float *excitation)
1270 {
1271 static const float gain_coeff[6] = {
1272 0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458
1273 };
1274 float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain;
1275 int n, idx, gain_weight;
1276 AMRFixed fcb;
1277
1278 assert(size <= MAX_FRAMESIZE / 2);
1279 memset(pulses, 0, sizeof(*pulses) * size);
1280
1281 fcb.pitch_lag = block_pitch_sh2 >> 2;
1282 fcb.pitch_fac = 1.0;
1283 fcb.no_repeat_mask = 0;
1284 fcb.n = 0;
1285
1286 /* For the other frame types, this is where we apply the innovation
1287 * (fixed) codebook pulses of the speech signal. */
1288 if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1289 aw_pulse_set1(s, gb, block_idx, &fcb);
1290 aw_pulse_set2(s, gb, block_idx, &fcb);
1291 } else /* FCB_TYPE_EXC_PULSES */ {
1292 int offset_nbits = 5 - frame_desc->log_n_blocks;
1293
1294 fcb.no_repeat_mask = -1;
1295 /* similar to ff_decode_10_pulses_35bits(), but with single pulses
1296 * (instead of double) for a subset of pulses */
1297 for (n = 0; n < 5; n++) {
1298 float sign;
1299 int pos1, pos2;
1300
1301 sign = get_bits1(gb) ? 1.0 : -1.0;
1302 pos1 = get_bits(gb, offset_nbits);
1303 fcb.x[fcb.n] = n + 5 * pos1;
1304 fcb.y[fcb.n++] = sign;
1305 if (n < frame_desc->dbl_pulses) {
1306 pos2 = get_bits(gb, offset_nbits);
1307 fcb.x[fcb.n] = n + 5 * pos2;
1308 fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign;
1309 }
1310 }
1311 }
1312 ff_set_fixed_vector(pulses, &fcb, 1.0, size);
1313
1314 /* Calculate gain for adaptive & fixed codebook signal.
1315 * see ff_amr_set_fixed_gain(). */
1316 idx = get_bits(gb, 7);
1317 fcb_gain = expf(ff_dot_productf(s->gain_pred_err, gain_coeff, 6) -
1318 5.2409161640 + wmavoice_gain_codebook_fcb[idx]);
1319 acb_gain = wmavoice_gain_codebook_acb[idx];
1320 pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx],
1321 -2.9957322736 /* log(0.05) */,
1322 1.6094379124 /* log(5.0) */);
1323
1324 gain_weight = 8 >> frame_desc->log_n_blocks;
1325 memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err,
1326 sizeof(*s->gain_pred_err) * (6 - gain_weight));
1327 for (n = 0; n < gain_weight; n++)
1328 s->gain_pred_err[n] = pred_err;
1329
1330 /* Calculation of adaptive codebook */
1331 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1332 int len;
1333 for (n = 0; n < size; n += len) {
1334 int next_idx_sh16;
1335 int abs_idx = block_idx * size + n;
1336 int pitch_sh16 = (s->last_pitch_val << 16) +
1337 s->pitch_diff_sh16 * abs_idx;
1338 int pitch = (pitch_sh16 + 0x6FFF) >> 16;
1339 int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000;
1340 idx = idx_sh16 >> 16;
1341 if (s->pitch_diff_sh16) {
1342 if (s->pitch_diff_sh16 > 0) {
1343 next_idx_sh16 = (idx_sh16) &~ 0xFFFF;
1344 } else
1345 next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF;
1346 len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8,
1347 1, size - n);
1348 } else
1349 len = size;
1350
1351 ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch],
1352 wmavoice_ipol1_coeffs, 17,
1353 idx, 9, len);
1354 }
1355 } else /* ACB_TYPE_HAMMING */ {
1356 int block_pitch = block_pitch_sh2 >> 2;
1357 idx = block_pitch_sh2 & 3;
1358 if (idx) {
1359 ff_acelp_interpolatef(excitation, &excitation[-block_pitch],
1360 wmavoice_ipol2_coeffs, 4,
1361 idx, 8, size);
1362 } else
1363 av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch,
1364 sizeof(float) * size);
1365 }
1366
1367 /* Interpolate ACB/FCB and use as excitation signal */
1368 ff_weighted_vector_sumf(excitation, excitation, pulses,
1369 acb_gain, fcb_gain, size);
1370 }
1371
1372 /**
1373 * Parse data in a single block.
1374 * @note we assume enough bits are available, caller should check.
1375 *
1376 * @param s WMA Voice decoding context private data
1377 * @param gb bit I/O context
1378 * @param block_idx index of the to-be-read block
1379 * @param size amount of samples to be read in this block
1380 * @param block_pitch_sh2 pitch for this block << 2
1381 * @param lsps LSPs for (the end of) this frame
1382 * @param prev_lsps LSPs for the last frame
1383 * @param frame_desc frame type descriptor
1384 * @param excitation target memory for the ACB+FCB interpolated signal
1385 * @param synth target memory for the speech synthesis filter output
1386 * @return 0 on success, <0 on error.
1387 */
1388 static void synth_block(WMAVoiceContext *s, GetBitContext *gb,
1389 int block_idx, int size,
1390 int block_pitch_sh2,
1391 const double *lsps, const double *prev_lsps,
1392 const struct frame_type_desc *frame_desc,
1393 float *excitation, float *synth)
1394 {
1395 double i_lsps[MAX_LSPS];
1396 float lpcs[MAX_LSPS];
1397 float fac;
1398 int n;
1399
1400 if (frame_desc->acb_type == ACB_TYPE_NONE)
1401 synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation);
1402 else
1403 synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2,
1404 frame_desc, excitation);
1405
1406 /* convert interpolated LSPs to LPCs */
1407 fac = (block_idx + 0.5) / frame_desc->n_blocks;
1408 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1409 i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n]));
1410 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1411
1412 /* Speech synthesis */
1413 ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps);
1414 }
1415
1416 /**
1417 * Synthesize output samples for a single frame.
1418 * @note we assume enough bits are available, caller should check.
1419 *
1420 * @param ctx WMA Voice decoder context
1421 * @param gb bit I/O context (s->gb or one for cross-packet superframes)
1422 * @param frame_idx Frame number within superframe [0-2]
1423 * @param samples pointer to output sample buffer, has space for at least 160
1424 * samples
1425 * @param lsps LSP array
1426 * @param prev_lsps array of previous frame's LSPs
1427 * @param excitation target buffer for excitation signal
1428 * @param synth target buffer for synthesized speech data
1429 * @return 0 on success, <0 on error.
1430 */
1431 static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx,
1432 float *samples,
1433 const double *lsps, const double *prev_lsps,
1434 float *excitation, float *synth)
1435 {
1436 WMAVoiceContext *s = ctx->priv_data;
1437 int n, n_blocks_x2, log_n_blocks_x2, cur_pitch_val;
1438 int pitch[MAX_BLOCKS], last_block_pitch;
1439
1440 /* Parse frame type ("frame header"), see frame_descs */
1441 int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)],
1442 block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks;
1443
1444 if (bd_idx < 0) {
1445 av_log(ctx, AV_LOG_ERROR,
1446 "Invalid frame type VLC code, skipping\n");
1447 return -1;
1448 }
1449
1450 /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */
1451 if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) {
1452 /* Pitch is provided per frame, which is interpreted as the pitch of
1453 * the last sample of the last block of this frame. We can interpolate
1454 * the pitch of other blocks (and even pitch-per-sample) by gradually
1455 * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */
1456 n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1;
1457 log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1;
1458 cur_pitch_val = s->min_pitch_val + get_bits(gb, s->pitch_nbits);
1459 cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1);
1460 if (s->last_acb_type == ACB_TYPE_NONE ||
1461 20 * abs(cur_pitch_val - s->last_pitch_val) >
1462 (cur_pitch_val + s->last_pitch_val))
1463 s->last_pitch_val = cur_pitch_val;
1464
1465 /* pitch per block */
1466 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1467 int fac = n * 2 + 1;
1468
1469 pitch[n] = (MUL16(fac, cur_pitch_val) +
1470 MUL16((n_blocks_x2 - fac), s->last_pitch_val) +
1471 frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2;
1472 }
1473
1474 /* "pitch-diff-per-sample" for calculation of pitch per sample */
1475 s->pitch_diff_sh16 =
1476 ((cur_pitch_val - s->last_pitch_val) << 16) / MAX_FRAMESIZE;
1477 }
1478
1479 /* Global gain (if silence) and pitch-adaptive window coordinates */
1480 switch (frame_descs[bd_idx].fcb_type) {
1481 case FCB_TYPE_SILENCE:
1482 s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)];
1483 break;
1484 case FCB_TYPE_AW_PULSES:
1485 aw_parse_coords(s, gb, pitch);
1486 break;
1487 }
1488
1489 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1490 int bl_pitch_sh2;
1491
1492 /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */
1493 switch (frame_descs[bd_idx].acb_type) {
1494 case ACB_TYPE_HAMMING: {
1495 /* Pitch is given per block. Per-block pitches are encoded as an
1496 * absolute value for the first block, and then delta values
1497 * relative to this value) for all subsequent blocks. The scale of
1498 * this pitch value is semi-logaritmic compared to its use in the
1499 * decoder, so we convert it to normal scale also. */
1500 int block_pitch,
1501 t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2,
1502 t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1,
1503 t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1;
1504
1505 if (n == 0) {
1506 block_pitch = get_bits(gb, s->block_pitch_nbits);
1507 } else
1508 block_pitch = last_block_pitch - s->block_delta_pitch_hrange +
1509 get_bits(gb, s->block_delta_pitch_nbits);
1510 /* Convert last_ so that any next delta is within _range */
1511 last_block_pitch = av_clip(block_pitch,
1512 s->block_delta_pitch_hrange,
1513 s->block_pitch_range -
1514 s->block_delta_pitch_hrange);
1515
1516 /* Convert semi-log-style scale back to normal scale */
1517 if (block_pitch < t1) {
1518 bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch;
1519 } else {
1520 block_pitch -= t1;
1521 if (block_pitch < t2) {
1522 bl_pitch_sh2 =
1523 (s->block_conv_table[1] << 2) + (block_pitch << 1);
1524 } else {
1525 block_pitch -= t2;
1526 if (block_pitch < t3) {
1527 bl_pitch_sh2 =
1528 (s->block_conv_table[2] + block_pitch) << 2;
1529 } else
1530 bl_pitch_sh2 = s->block_conv_table[3] << 2;
1531 }
1532 }
1533 pitch[n] = bl_pitch_sh2 >> 2;
1534 break;
1535 }
1536
1537 case ACB_TYPE_ASYMMETRIC: {
1538 bl_pitch_sh2 = pitch[n] << 2;
1539 break;
1540 }
1541
1542 default: // ACB_TYPE_NONE has no pitch
1543 bl_pitch_sh2 = 0;
1544 break;
1545 }
1546
1547 synth_block(s, gb, n, block_nsamples, bl_pitch_sh2,
1548 lsps, prev_lsps, &frame_descs[bd_idx],
1549 &excitation[n * block_nsamples],
1550 &synth[n * block_nsamples]);
1551 }
1552
1553 /* Averaging projection filter, if applicable. Else, just copy samples
1554 * from synthesis buffer */
1555 if (s->do_apf) {
1556 double i_lsps[MAX_LSPS];
1557 float lpcs[MAX_LSPS];
1558
1559 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1560 i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n]));
1561 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1562 postfilter(s, synth, samples, 80, lpcs,
1563 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx],
1564 frame_descs[bd_idx].fcb_type, pitch[0]);
1565
1566 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1567 i_lsps[n] = cos(lsps[n]);
1568 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1569 postfilter(s, &synth[80], &samples[80], 80, lpcs,
1570 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80],
1571 frame_descs[bd_idx].fcb_type, pitch[0]);
1572 } else
1573 memcpy(samples, synth, 160 * sizeof(synth[0]));
1574
1575 /* Cache values for next frame */
1576 s->frame_cntr++;
1577 if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%)
1578 s->last_acb_type = frame_descs[bd_idx].acb_type;
1579 switch (frame_descs[bd_idx].acb_type) {
1580 case ACB_TYPE_NONE:
1581 s->last_pitch_val = 0;
1582 break;
1583 case ACB_TYPE_ASYMMETRIC:
1584 s->last_pitch_val = cur_pitch_val;
1585 break;
1586 case ACB_TYPE_HAMMING:
1587 s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1];
1588 break;
1589 }
1590
1591 return 0;
1592 }
1593
1594 /**
1595 * Ensure minimum value for first item, maximum value for last value,
1596 * proper spacing between each value and proper ordering.
1597 *
1598 * @param lsps array of LSPs
1599 * @param num size of LSP array
1600 *
1601 * @note basically a double version of #ff_acelp_reorder_lsf(), might be
1602 * useful to put in a generic location later on. Parts are also
1603 * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(),
1604 * which is in float.
1605 */
1606 static void stabilize_lsps(double *lsps, int num)
1607 {
1608 int n, m, l;
1609
1610 /* set minimum value for first, maximum value for last and minimum
1611 * spacing between LSF values.
1612 * Very similar to ff_set_min_dist_lsf(), but in double. */
1613 lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI);
1614 for (n = 1; n < num; n++)
1615 lsps[n] = FFMAX(lsps[n], lsps[n - 1] + 0.0125 * M_PI);
1616 lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI);
1617
1618 /* reorder (looks like one-time / non-recursed bubblesort).
1619 * Very similar to ff_sort_nearly_sorted_floats(), but in double. */
1620 for (n = 1; n < num; n++) {
1621 if (lsps[n] < lsps[n - 1]) {
1622 for (m = 1; m < num; m++) {
1623 double tmp = lsps[m];
1624 for (l = m - 1; l >= 0; l--) {
1625 if (lsps[l] <= tmp) break;
1626 lsps[l + 1] = lsps[l];
1627 }
1628 lsps[l + 1] = tmp;
1629 }
1630 break;
1631 }
1632 }
1633 }
1634
1635 /**
1636 * Test if there's enough bits to read 1 superframe.
1637 *
1638 * @param orig_gb bit I/O context used for reading. This function
1639 * does not modify the state of the bitreader; it
1640 * only uses it to copy the current stream position
1641 * @param s WMA Voice decoding context private data
1642 * @return -1 if unsupported, 1 on not enough bits or 0 if OK.
1643 */
1644 static int check_bits_for_superframe(GetBitContext *orig_gb,
1645 WMAVoiceContext *s)
1646 {
1647 GetBitContext s_gb, *gb = &s_gb;
1648 int n, need_bits, bd_idx;
1649 const struct frame_type_desc *frame_desc;
1650
1651 /* initialize a copy */
1652 init_get_bits(gb, orig_gb->buffer, orig_gb->size_in_bits);
1653 skip_bits_long(gb, get_bits_count(orig_gb));
1654 assert(get_bits_left(gb) == get_bits_left(orig_gb));
1655
1656 /* superframe header */
1657 if (get_bits_left(gb) < 14)
1658 return 1;
1659 if (!get_bits1(gb))
1660 return -1; // WMAPro-in-WMAVoice superframe
1661 if (get_bits1(gb)) skip_bits(gb, 12); // number of samples in superframe
1662 if (s->has_residual_lsps) { // residual LSPs (for all frames)
1663 if (get_bits_left(gb) < s->sframe_lsp_bitsize)
1664 return 1;
1665 skip_bits_long(gb, s->sframe_lsp_bitsize);
1666 }
1667
1668 /* frames */
1669 for (n = 0; n < MAX_FRAMES; n++) {
1670 int aw_idx_is_ext = 0;
1671
1672 if (!s->has_residual_lsps) { // independent LSPs (per-frame)
1673 if (get_bits_left(gb) < s->frame_lsp_bitsize) return 1;
1674 skip_bits_long(gb, s->frame_lsp_bitsize);
1675 }
1676 bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)];
1677 if (bd_idx < 0)
1678 return -1; // invalid frame type VLC code
1679 frame_desc = &frame_descs[bd_idx];
1680 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1681 if (get_bits_left(gb) < s->pitch_nbits)
1682 return 1;
1683 skip_bits_long(gb, s->pitch_nbits);
1684 }
1685 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1686 skip_bits(gb, 8);
1687 } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1688 int tmp = get_bits(gb, 6);
1689 if (tmp >= 0x36) {
1690 skip_bits(gb, 2);
1691 aw_idx_is_ext = 1;
1692 }
1693 }
1694
1695 /* blocks */
1696 if (frame_desc->acb_type == ACB_TYPE_HAMMING) {
1697 need_bits = s->block_pitch_nbits +
1698 (frame_desc->n_blocks - 1) * s->block_delta_pitch_nbits;
1699 } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1700 need_bits = 2 * !aw_idx_is_ext;
1701 } else
1702 need_bits = 0;
1703 need_bits += frame_desc->frame_size;
1704 if (get_bits_left(gb) < need_bits)
1705 return 1;
1706 skip_bits_long(gb, need_bits);
1707 }
1708
1709 return 0;
1710 }
1711
1712 /**
1713 * Synthesize output samples for a single superframe. If we have any data
1714 * cached in s->sframe_cache, that will be used instead of whatever is loaded
1715 * in s->gb.
1716 *
1717 * WMA Voice superframes contain 3 frames, each containing 160 audio samples,
1718 * to give a total of 480 samples per frame. See #synth_frame() for frame
1719 * parsing. In addition to 3 frames, superframes can also contain the LSPs
1720 * (if these are globally specified for all frames (residually); they can
1721 * also be specified individually per-frame. See the s->has_residual_lsps
1722 * option), and can specify the number of samples encoded in this superframe
1723 * (if less than 480), usually used to prevent blanks at track boundaries.
1724 *
1725 * @param ctx WMA Voice decoder context
1726 * @param samples pointer to output buffer for voice samples
1727 * @param data_size pointer containing the size of #samples on input, and the
1728 * amount of #samples filled on output
1729 * @return 0 on success, <0 on error or 1 if there was not enough data to
1730 * fully parse the superframe
1731 */
1732 static int synth_superframe(AVCodecContext *ctx, int *got_frame_ptr)
1733 {
1734 WMAVoiceContext *s = ctx->priv_data;
1735 GetBitContext *gb = &s->gb, s_gb;
1736 int n, res, n_samples = 480;
1737 double lsps[MAX_FRAMES][MAX_LSPS];
1738 const double *mean_lsf = s->lsps == 16 ?
1739 wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode];
1740 float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12];
1741 float synth[MAX_LSPS + MAX_SFRAMESIZE];
1742 float *samples;
1743
1744 memcpy(synth, s->synth_history,
1745 s->lsps * sizeof(*synth));
1746 memcpy(excitation, s->excitation_history,
1747 s->history_nsamples * sizeof(*excitation));
1748
1749 if (s->sframe_cache_size > 0) {
1750 gb = &s_gb;
1751 init_get_bits(gb, s->sframe_cache, s->sframe_cache_size);
1752 s->sframe_cache_size = 0;
1753 }
1754
1755 if ((res = check_bits_for_superframe(gb, s)) == 1) {
1756 *got_frame_ptr = 0;
1757 return 1;
1758 }
1759
1760 /* First bit is speech/music bit, it differentiates between WMAVoice
1761 * speech samples (the actual codec) and WMAVoice music samples, which
1762 * are really WMAPro-in-WMAVoice-superframes. I've never seen those in
1763 * the wild yet. */
1764 if (!get_bits1(gb)) {
1765 av_log_missing_feature(ctx, "WMAPro-in-WMAVoice support", 1);
1766 return -1;
1767 }
1768
1769 /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */
1770 if (get_bits1(gb)) {
1771 if ((n_samples = get_bits(gb, 12)) > 480) {
1772 av_log(ctx, AV_LOG_ERROR,
1773 "Superframe encodes >480 samples (%d), not allowed\n",
1774 n_samples);
1775 return -1;
1776 }
1777 }
1778 /* Parse LSPs, if global for the superframe (can also be per-frame). */
1779 if (s->has_residual_lsps) {
1780 double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2];
1781
1782 for (n = 0; n < s->lsps; n++)
1783 prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n];
1784
1785 if (s->lsps == 10) {
1786 dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1787 } else /* s->lsps == 16 */
1788 dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1789
1790 for (n = 0; n < s->lsps; n++) {
1791 lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]);
1792 lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]);
1793 lsps[2][n] += mean_lsf[n];
1794 }
1795 for (n = 0; n < 3; n++)
1796 stabilize_lsps(lsps[n], s->lsps);
1797 }
1798
1799 /* get output buffer */
1800 s->frame.nb_samples = 480;
1801 if ((res = ctx->get_buffer(ctx, &s->frame)) < 0) {
1802 av_log(ctx, AV_LOG_ERROR, "get_buffer() failed\n");
1803 return res;
1804 }
1805 s->frame.nb_samples = n_samples;
1806 samples = (float *)s->frame.data[0];
1807
1808 /* Parse frames, optionally preceeded by per-frame (independent) LSPs. */
1809 for (n = 0; n < 3; n++) {
1810 if (!s->has_residual_lsps) {
1811 int m;
1812
1813 if (s->lsps == 10) {
1814 dequant_lsp10i(gb, lsps[n]);
1815 } else /* s->lsps == 16 */
1816 dequant_lsp16i(gb, lsps[n]);
1817
1818 for (m = 0; m < s->lsps; m++)
1819 lsps[n][m] += mean_lsf[m];
1820 stabilize_lsps(lsps[n], s->lsps);
1821 }
1822
1823 if ((res = synth_frame(ctx, gb, n,
1824 &samples[n * MAX_FRAMESIZE],
1825 lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1],
1826 &excitation[s->history_nsamples + n * MAX_FRAMESIZE],
1827 &synth[s->lsps + n * MAX_FRAMESIZE]))) {
1828 *got_frame_ptr = 0;
1829 return res;
1830 }
1831 }
1832
1833 /* Statistics? FIXME - we don't check for length, a slight overrun
1834 * will be caught by internal buffer padding, and anything else
1835 * will be skipped, not read. */
1836 if (get_bits1(gb)) {
1837 res = get_bits(gb, 4);
1838 skip_bits(gb, 10 * (res + 1));
1839 }
1840
1841 *got_frame_ptr = 1;
1842
1843 /* Update history */
1844 memcpy(s->prev_lsps, lsps[2],
1845 s->lsps * sizeof(*s->prev_lsps));
1846 memcpy(s->synth_history, &synth[MAX_SFRAMESIZE],
1847 s->lsps * sizeof(*synth));
1848 memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE],
1849 s->history_nsamples * sizeof(*excitation));
1850 if (s->do_apf)
1851 memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE],
1852 s->history_nsamples * sizeof(*s->zero_exc_pf));
1853
1854 return 0;
1855 }
1856
1857 /**
1858 * Parse the packet header at the start of each packet (input data to this
1859 * decoder).
1860 *
1861 * @param s WMA Voice decoding context private data
1862 * @return 1 if not enough bits were available, or 0 on success.
1863 */
1864 static int parse_packet_header(WMAVoiceContext *s)
1865 {
1866 GetBitContext *gb = &s->gb;
1867 unsigned int res;
1868
1869 if (get_bits_left(gb) < 11)
1870 return 1;
1871 skip_bits(gb, 4); // packet sequence number
1872 s->has_residual_lsps = get_bits1(gb);
1873 do {
1874 res = get_bits(gb, 6); // number of superframes per packet
1875 // (minus first one if there is spillover)
1876 if (get_bits_left(gb) < 6 * (res == 0x3F) + s->spillover_bitsize)
1877 return 1;
1878 } while (res == 0x3F);
1879 s->spillover_nbits = get_bits(gb, s->spillover_bitsize);
1880
1881 return 0;
1882 }
1883
1884 /**
1885 * Copy (unaligned) bits from gb/data/size to pb.
1886 *
1887 * @param pb target buffer to copy bits into
1888 * @param data source buffer to copy bits from
1889 * @param size size of the source data, in bytes
1890 * @param gb bit I/O context specifying the current position in the source.
1891 * data. This function might use this to align the bit position to
1892 * a whole-byte boundary before calling #avpriv_copy_bits() on aligned
1893 * source data
1894 * @param nbits the amount of bits to copy from source to target
1895 *
1896 * @note after calling this function, the current position in the input bit
1897 * I/O context is undefined.
1898 */
1899 static void copy_bits(PutBitContext *pb,
1900 const uint8_t *data, int size,
1901 GetBitContext *gb, int nbits)
1902 {
1903 int rmn_bytes, rmn_bits;
1904
1905 rmn_bits = rmn_bytes = get_bits_left(gb);
1906 if (rmn_bits < nbits)
1907 return;
1908 if (nbits > pb->size_in_bits - put_bits_count(pb))
1909 return;
1910 rmn_bits &= 7; rmn_bytes >>= 3;
1911 if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0)
1912 put_bits(pb, rmn_bits, get_bits(gb, rmn_bits));
1913 avpriv_copy_bits(pb, data + size - rmn_bytes,
1914 FFMIN(nbits - rmn_bits, rmn_bytes << 3));
1915 }
1916
1917 /**
1918 * Packet decoding: a packet is anything that the (ASF) demuxer contains,
1919 * and we expect that the demuxer / application provides it to us as such
1920 * (else you'll probably get garbage as output). Every packet has a size of
1921 * ctx->block_align bytes, starts with a packet header (see
1922 * #parse_packet_header()), and then a series of superframes. Superframe
1923 * boundaries may exceed packets, i.e. superframes can split data over
1924 * multiple (two) packets.
1925 *
1926 * For more information about frames, see #synth_superframe().
1927 */
1928 static int wmavoice_decode_packet(AVCodecContext *ctx, void *data,
1929 int *got_frame_ptr, AVPacket *avpkt)
1930 {
1931 WMAVoiceContext *s = ctx->priv_data;
1932 GetBitContext *gb = &s->gb;
1933 int size, res, pos;
1934
1935 /* Packets are sometimes a multiple of ctx->block_align, with a packet
1936 * header at each ctx->block_align bytes. However, Libav's ASF demuxer
1937 * feeds us ASF packets, which may concatenate multiple "codec" packets
1938 * in a single "muxer" packet, so we artificially emulate that by
1939 * capping the packet size at ctx->block_align. */
1940 for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align);
1941 if (!size) {
1942 *got_frame_ptr = 0;
1943 return 0;
1944 }
1945 init_get_bits(&s->gb, avpkt->data, size << 3);
1946
1947 /* size == ctx->block_align is used to indicate whether we are dealing with
1948 * a new packet or a packet of which we already read the packet header
1949 * previously. */
1950 if (size == ctx->block_align) { // new packet header
1951 if ((res = parse_packet_header(s)) < 0)
1952 return res;
1953
1954 /* If the packet header specifies a s->spillover_nbits, then we want
1955 * to push out all data of the previous packet (+ spillover) before
1956 * continuing to parse new superframes in the current packet. */
1957 if (s->spillover_nbits > 0) {
1958 if (s->sframe_cache_size > 0) {
1959 int cnt = get_bits_count(gb);
1960 copy_bits(&s->pb, avpkt->data, size, gb, s->spillover_nbits);
1961 flush_put_bits(&s->pb);
1962 s->sframe_cache_size += s->spillover_nbits;
1963 if ((res = synth_superframe(ctx, got_frame_ptr)) == 0 &&
1964 *got_frame_ptr) {
1965 cnt += s->spillover_nbits;
1966 s->skip_bits_next = cnt & 7;
1967 *(AVFrame *)data = s->frame;
1968 return cnt >> 3;
1969 } else
1970 skip_bits_long (gb, s->spillover_nbits - cnt +
1971 get_bits_count(gb)); // resync
1972 } else
1973 skip_bits_long(gb, s->spillover_nbits); // resync
1974 }
1975 } else if (s->skip_bits_next)
1976 skip_bits(gb, s->skip_bits_next);
1977
1978 /* Try parsing superframes in current packet */
1979 s->sframe_cache_size = 0;
1980 s->skip_bits_next = 0;
1981 pos = get_bits_left(gb);
1982 if ((res = synth_superframe(ctx, got_frame_ptr)) < 0) {
1983 return res;
1984 } else if (*got_frame_ptr) {
1985 int cnt = get_bits_count(gb);
1986 s->skip_bits_next = cnt & 7;
1987 *(AVFrame *)data = s->frame;
1988 return cnt >> 3;
1989 } else if ((s->sframe_cache_size = pos) > 0) {
1990 /* rewind bit reader to start of last (incomplete) superframe... */
1991 init_get_bits(gb, avpkt->data, size << 3);
1992 skip_bits_long(gb, (size << 3) - pos);
1993 assert(get_bits_left(gb) == pos);
1994
1995 /* ...and cache it for spillover in next packet */
1996 init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE);
1997 copy_bits(&s->pb, avpkt->data, size, gb, s->sframe_cache_size);
1998 // FIXME bad - just copy bytes as whole and add use the
1999 // skip_bits_next field
2000 }
2001
2002 return size;
2003 }
2004
2005 static av_cold int wmavoice_decode_end(AVCodecContext *ctx)
2006 {
2007 WMAVoiceContext *s = ctx->priv_data;
2008
2009 if (s->do_apf) {
2010 ff_rdft_end(&s->rdft);
2011 ff_rdft_end(&s->irdft);
2012 ff_dct_end(&s->dct);
2013 ff_dct_end(&s->dst);
2014 }
2015
2016 return 0;
2017 }
2018
2019 static av_cold void wmavoice_flush(AVCodecContext *ctx)
2020 {
2021 WMAVoiceContext *s = ctx->priv_data;
2022 int n;
2023
2024 s->postfilter_agc = 0;
2025 s->sframe_cache_size = 0;
2026 s->skip_bits_next = 0;
2027 for (n = 0; n < s->lsps; n++)
2028 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
2029 memset(s->excitation_history, 0,
2030 sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY);
2031 memset(s->synth_history, 0,
2032 sizeof(*s->synth_history) * MAX_LSPS);
2033 memset(s->gain_pred_err, 0,
2034 sizeof(s->gain_pred_err));
2035
2036 if (s->do_apf) {
2037 memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0,
2038 sizeof(*s->synth_filter_out_buf) * s->lsps);
2039 memset(s->dcf_mem, 0,
2040 sizeof(*s->dcf_mem) * 2);
2041 memset(s->zero_exc_pf, 0,
2042 sizeof(*s->zero_exc_pf) * s->history_nsamples);
2043 memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache));
2044 }
2045 }
2046
2047 AVCodec ff_wmavoice_decoder = {
2048 .name = "wmavoice",
2049 .type = AVMEDIA_TYPE_AUDIO,
2050 .id = CODEC_ID_WMAVOICE,
2051 .priv_data_size = sizeof(WMAVoiceContext),
2052 .init = wmavoice_decode_init,
2053 .close = wmavoice_decode_end,
2054 .decode = wmavoice_decode_packet,
2055 .capabilities = CODEC_CAP_SUBFRAMES | CODEC_CAP_DR1,
2056 .flush = wmavoice_flush,
2057 .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"),
2058 };