f3d0034ea883fad39144ad47c3a75787d0d441e2
[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 #define UNCHECKED_BITSTREAM_READER 1
29
30 #include <math.h>
31
32 #include "libavutil/channel_layout.h"
33 #include "libavutil/float_dsp.h"
34 #include "libavutil/mem.h"
35 #include "avcodec.h"
36 #include "internal.h"
37 #include "get_bits.h"
38 #include "put_bits.h"
39 #include "wmavoice_data.h"
40 #include "celp_filters.h"
41 #include "acelp_vectors.h"
42 #include "acelp_filters.h"
43 #include "lsp.h"
44 #include "dct.h"
45 #include "rdft.h"
46 #include "sinewin.h"
47
48 #define MAX_BLOCKS 8 ///< maximum number of blocks per frame
49 #define MAX_LSPS 16 ///< maximum filter order
50 #define MAX_LSPS_ALIGN16 16 ///< same as #MAX_LSPS; needs to be multiple
51 ///< of 16 for ASM input buffer alignment
52 #define MAX_FRAMES 3 ///< maximum number of frames per superframe
53 #define MAX_FRAMESIZE 160 ///< maximum number of samples per frame
54 #define MAX_SIGNAL_HISTORY 416 ///< maximum excitation signal history
55 #define MAX_SFRAMESIZE (MAX_FRAMESIZE * MAX_FRAMES)
56 ///< maximum number of samples per superframe
57 #define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that
58 ///< was split over two packets
59 #define VLC_NBITS 6 ///< number of bits to read per VLC iteration
60
61 /**
62 * Frame type VLC coding.
63 */
64 static VLC frame_type_vlc;
65
66 /**
67 * Adaptive codebook types.
68 */
69 enum {
70 ACB_TYPE_NONE = 0, ///< no adaptive codebook (only hardcoded fixed)
71 ACB_TYPE_ASYMMETRIC = 1, ///< adaptive codebook with per-frame pitch, which
72 ///< we interpolate to get a per-sample pitch.
73 ///< Signal is generated using an asymmetric sinc
74 ///< window function
75 ///< @note see #wmavoice_ipol1_coeffs
76 ACB_TYPE_HAMMING = 2 ///< Per-block pitch with signal generation using
77 ///< a Hamming sinc window function
78 ///< @note see #wmavoice_ipol2_coeffs
79 };
80
81 /**
82 * Fixed codebook types.
83 */
84 enum {
85 FCB_TYPE_SILENCE = 0, ///< comfort noise during silence
86 ///< generated from a hardcoded (fixed) codebook
87 ///< with per-frame (low) gain values
88 FCB_TYPE_HARDCODED = 1, ///< hardcoded (fixed) codebook with per-block
89 ///< gain values
90 FCB_TYPE_AW_PULSES = 2, ///< Pitch-adaptive window (AW) pulse signals,
91 ///< used in particular for low-bitrate streams
92 FCB_TYPE_EXC_PULSES = 3, ///< Innovation (fixed) codebook pulse sets in
93 ///< combinations of either single pulses or
94 ///< pulse pairs
95 };
96
97 /**
98 * Description of frame types.
99 */
100 static const struct frame_type_desc {
101 uint8_t n_blocks; ///< amount of blocks per frame (each block
102 ///< (contains 160/#n_blocks samples)
103 uint8_t log_n_blocks; ///< log2(#n_blocks)
104 uint8_t acb_type; ///< Adaptive codebook type (ACB_TYPE_*)
105 uint8_t fcb_type; ///< Fixed codebook type (FCB_TYPE_*)
106 uint8_t dbl_pulses; ///< how many pulse vectors have pulse pairs
107 ///< (rather than just one single pulse)
108 ///< only if #fcb_type == #FCB_TYPE_EXC_PULSES
109 uint16_t frame_size; ///< the amount of bits that make up the block
110 ///< data (per frame)
111 } frame_descs[17] = {
112 { 1, 0, ACB_TYPE_NONE, FCB_TYPE_SILENCE, 0, 0 },
113 { 2, 1, ACB_TYPE_NONE, FCB_TYPE_HARDCODED, 0, 28 },
114 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_AW_PULSES, 0, 46 },
115 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 80 },
116 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 104 },
117 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 0, 108 },
118 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 132 },
119 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 168 },
120 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 64 },
121 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 80 },
122 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 104 },
123 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 108 },
124 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 132 },
125 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 168 },
126 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 176 },
127 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 208 },
128 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 256 }
129 };
130
131 /**
132 * WMA Voice decoding context.
133 */
134 typedef struct {
135 /**
136 * @name Global values specified in the stream header / extradata or used all over.
137 * @{
138 */
139 GetBitContext gb; ///< packet bitreader. During decoder init,
140 ///< it contains the extradata from the
141 ///< demuxer. During decoding, it contains
142 ///< packet data.
143 int8_t vbm_tree[25]; ///< converts VLC codes to frame type
144
145 int spillover_bitsize; ///< number of bits used to specify
146 ///< #spillover_nbits in the packet header
147 ///< = ceil(log2(ctx->block_align << 3))
148 int history_nsamples; ///< number of samples in history for signal
149 ///< prediction (through ACB)
150
151 /* postfilter specific values */
152 int do_apf; ///< whether to apply the averaged
153 ///< projection filter (APF)
154 int denoise_strength; ///< strength of denoising in Wiener filter
155 ///< [0-11]
156 int denoise_tilt_corr; ///< Whether to apply tilt correction to the
157 ///< Wiener filter coefficients (postfilter)
158 int dc_level; ///< Predicted amount of DC noise, based
159 ///< on which a DC removal filter is used
160
161 int lsps; ///< number of LSPs per frame [10 or 16]
162 int lsp_q_mode; ///< defines quantizer defaults [0, 1]
163 int lsp_def_mode; ///< defines different sets of LSP defaults
164 ///< [0, 1]
165 int frame_lsp_bitsize; ///< size (in bits) of LSPs, when encoded
166 ///< per-frame (independent coding)
167 int sframe_lsp_bitsize; ///< size (in bits) of LSPs, when encoded
168 ///< per superframe (residual coding)
169
170 int min_pitch_val; ///< base value for pitch parsing code
171 int max_pitch_val; ///< max value + 1 for pitch parsing
172 int pitch_nbits; ///< number of bits used to specify the
173 ///< pitch value in the frame header
174 int block_pitch_nbits; ///< number of bits used to specify the
175 ///< first block's pitch value
176 int block_pitch_range; ///< range of the block pitch
177 int block_delta_pitch_nbits; ///< number of bits used to specify the
178 ///< delta pitch between this and the last
179 ///< block's pitch value, used in all but
180 ///< first block
181 int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is
182 ///< from -this to +this-1)
183 uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale
184 ///< conversion
185
186 /**
187 * @}
188 *
189 * @name Packet values specified in the packet header or related to a packet.
190 *
191 * A packet is considered to be a single unit of data provided to this
192 * decoder by the demuxer.
193 * @{
194 */
195 int spillover_nbits; ///< number of bits of the previous packet's
196 ///< last superframe preceding this
197 ///< packet's first full superframe (useful
198 ///< for re-synchronization also)
199 int has_residual_lsps; ///< if set, superframes contain one set of
200 ///< LSPs that cover all frames, encoded as
201 ///< independent and residual LSPs; if not
202 ///< set, each frame contains its own, fully
203 ///< independent, LSPs
204 int skip_bits_next; ///< number of bits to skip at the next call
205 ///< to #wmavoice_decode_packet() (since
206 ///< they're part of the previous superframe)
207
208 uint8_t sframe_cache[SFRAME_CACHE_MAXSIZE + FF_INPUT_BUFFER_PADDING_SIZE];
209 ///< cache for superframe data split over
210 ///< multiple packets
211 int sframe_cache_size; ///< set to >0 if we have data from an
212 ///< (incomplete) superframe from a previous
213 ///< packet that spilled over in the current
214 ///< packet; specifies the amount of bits in
215 ///< #sframe_cache
216 PutBitContext pb; ///< bitstream writer for #sframe_cache
217
218 /**
219 * @}
220 *
221 * @name Frame and superframe values
222 * Superframe and frame data - these can change from frame to frame,
223 * although some of them do in that case serve as a cache / history for
224 * the next frame or superframe.
225 * @{
226 */
227 double prev_lsps[MAX_LSPS]; ///< LSPs of the last frame of the previous
228 ///< superframe
229 int last_pitch_val; ///< pitch value of the previous frame
230 int last_acb_type; ///< frame type [0-2] of the previous frame
231 int pitch_diff_sh16; ///< ((cur_pitch_val - #last_pitch_val)
232 ///< << 16) / #MAX_FRAMESIZE
233 float silence_gain; ///< set for use in blocks if #ACB_TYPE_NONE
234
235 int aw_idx_is_ext; ///< whether the AW index was encoded in
236 ///< 8 bits (instead of 6)
237 int aw_pulse_range; ///< the range over which #aw_pulse_set1()
238 ///< can apply the pulse, relative to the
239 ///< value in aw_first_pulse_off. The exact
240 ///< position of the first AW-pulse is within
241 ///< [pulse_off, pulse_off + this], and
242 ///< depends on bitstream values; [16 or 24]
243 int aw_n_pulses[2]; ///< number of AW-pulses in each block; note
244 ///< that this number can be negative (in
245 ///< which case it basically means "zero")
246 int aw_first_pulse_off[2]; ///< index of first sample to which to
247 ///< apply AW-pulses, or -0xff if unset
248 int aw_next_pulse_off_cache; ///< the position (relative to start of the
249 ///< second block) at which pulses should
250 ///< start to be positioned, serves as a
251 ///< cache for pitch-adaptive window pulses
252 ///< between blocks
253
254 int frame_cntr; ///< current frame index [0 - 0xFFFE]; is
255 ///< only used for comfort noise in #pRNG()
256 float gain_pred_err[6]; ///< cache for gain prediction
257 float excitation_history[MAX_SIGNAL_HISTORY];
258 ///< cache of the signal of previous
259 ///< superframes, used as a history for
260 ///< signal generation
261 float synth_history[MAX_LSPS]; ///< see #excitation_history
262 /**
263 * @}
264 *
265 * @name Postfilter values
266 *
267 * Variables used for postfilter implementation, mostly history for
268 * smoothing and so on, and context variables for FFT/iFFT.
269 * @{
270 */
271 RDFTContext rdft, irdft; ///< contexts for FFT-calculation in the
272 ///< postfilter (for denoise filter)
273 DCTContext dct, dst; ///< contexts for phase shift (in Hilbert
274 ///< transform, part of postfilter)
275 float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi]
276 ///< range
277 float postfilter_agc; ///< gain control memory, used in
278 ///< #adaptive_gain_control()
279 float dcf_mem[2]; ///< DC filter history
280 float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE];
281 ///< zero filter output (i.e. excitation)
282 ///< by postfilter
283 float denoise_filter_cache[MAX_FRAMESIZE];
284 int denoise_filter_cache_size; ///< samples in #denoise_filter_cache
285 DECLARE_ALIGNED(32, float, tilted_lpcs_pf)[0x80];
286 ///< aligned buffer for LPC tilting
287 DECLARE_ALIGNED(32, float, denoise_coeffs_pf)[0x80];
288 ///< aligned buffer for denoise coefficients
289 DECLARE_ALIGNED(32, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16];
290 ///< aligned buffer for postfilter speech
291 ///< synthesis
292 /**
293 * @}
294 */
295 } WMAVoiceContext;
296
297 /**
298 * Set up the variable bit mode (VBM) tree from container extradata.
299 * @param gb bit I/O context.
300 * The bit context (s->gb) should be loaded with byte 23-46 of the
301 * container extradata (i.e. the ones containing the VBM tree).
302 * @param vbm_tree pointer to array to which the decoded VBM tree will be
303 * written.
304 * @return 0 on success, <0 on error.
305 */
306 static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25])
307 {
308 int cntr[8] = { 0 }, n, res;
309
310 memset(vbm_tree, 0xff, sizeof(vbm_tree[0]) * 25);
311 for (n = 0; n < 17; n++) {
312 res = get_bits(gb, 3);
313 if (cntr[res] > 3) // should be >= 3 + (res == 7))
314 return -1;
315 vbm_tree[res * 3 + cntr[res]++] = n;
316 }
317 return 0;
318 }
319
320 static av_cold void wmavoice_init_static_data(AVCodec *codec)
321 {
322 static const uint8_t bits[] = {
323 2, 2, 2, 4, 4, 4,
324 6, 6, 6, 8, 8, 8,
325 10, 10, 10, 12, 12, 12,
326 14, 14, 14, 14
327 };
328 static const uint16_t codes[] = {
329 0x0000, 0x0001, 0x0002, // 00/01/10
330 0x000c, 0x000d, 0x000e, // 11+00/01/10
331 0x003c, 0x003d, 0x003e, // 1111+00/01/10
332 0x00fc, 0x00fd, 0x00fe, // 111111+00/01/10
333 0x03fc, 0x03fd, 0x03fe, // 11111111+00/01/10
334 0x0ffc, 0x0ffd, 0x0ffe, // 1111111111+00/01/10
335 0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx
336 };
337
338 INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits),
339 bits, 1, 1, codes, 2, 2, 132);
340 }
341
342 /**
343 * Set up decoder with parameters from demuxer (extradata etc.).
344 */
345 static av_cold int wmavoice_decode_init(AVCodecContext *ctx)
346 {
347 int n, flags, pitch_range, lsp16_flag;
348 WMAVoiceContext *s = ctx->priv_data;
349
350 /**
351 * Extradata layout:
352 * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c),
353 * - byte 19-22: flags field (annoyingly in LE; see below for known
354 * values),
355 * - byte 23-46: variable bitmode tree (really just 17 * 3 bits,
356 * rest is 0).
357 */
358 if (ctx->extradata_size != 46) {
359 av_log(ctx, AV_LOG_ERROR,
360 "Invalid extradata size %d (should be 46)\n",
361 ctx->extradata_size);
362 return AVERROR_INVALIDDATA;
363 }
364 flags = AV_RL32(ctx->extradata + 18);
365 s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align);
366 s->do_apf = flags & 0x1;
367 if (s->do_apf) {
368 ff_rdft_init(&s->rdft, 7, DFT_R2C);
369 ff_rdft_init(&s->irdft, 7, IDFT_C2R);
370 ff_dct_init(&s->dct, 6, DCT_I);
371 ff_dct_init(&s->dst, 6, DST_I);
372
373 ff_sine_window_init(s->cos, 256);
374 memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0]));
375 for (n = 0; n < 255; n++) {
376 s->sin[n] = -s->sin[510 - n];
377 s->cos[510 - n] = s->cos[n];
378 }
379 }
380 s->denoise_strength = (flags >> 2) & 0xF;
381 if (s->denoise_strength >= 12) {
382 av_log(ctx, AV_LOG_ERROR,
383 "Invalid denoise filter strength %d (max=11)\n",
384 s->denoise_strength);
385 return AVERROR_INVALIDDATA;
386 }
387 s->denoise_tilt_corr = !!(flags & 0x40);
388 s->dc_level = (flags >> 7) & 0xF;
389 s->lsp_q_mode = !!(flags & 0x2000);
390 s->lsp_def_mode = !!(flags & 0x4000);
391 lsp16_flag = flags & 0x1000;
392 if (lsp16_flag) {
393 s->lsps = 16;
394 s->frame_lsp_bitsize = 34;
395 s->sframe_lsp_bitsize = 60;
396 } else {
397 s->lsps = 10;
398 s->frame_lsp_bitsize = 24;
399 s->sframe_lsp_bitsize = 48;
400 }
401 for (n = 0; n < s->lsps; n++)
402 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
403
404 init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3);
405 if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) {
406 av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n");
407 return AVERROR_INVALIDDATA;
408 }
409
410 s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8;
411 s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8;
412 pitch_range = s->max_pitch_val - s->min_pitch_val;
413 if (pitch_range <= 0) {
414 av_log(ctx, AV_LOG_ERROR, "Invalid pitch range; broken extradata?\n");
415 return AVERROR_INVALIDDATA;
416 }
417 s->pitch_nbits = av_ceil_log2(pitch_range);
418 s->last_pitch_val = 40;
419 s->last_acb_type = ACB_TYPE_NONE;
420 s->history_nsamples = s->max_pitch_val + 8;
421
422 if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) {
423 int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8,
424 max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8;
425
426 av_log(ctx, AV_LOG_ERROR,
427 "Unsupported samplerate %d (min=%d, max=%d)\n",
428 ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz
429
430 return AVERROR(ENOSYS);
431 }
432
433 s->block_conv_table[0] = s->min_pitch_val;
434 s->block_conv_table[1] = (pitch_range * 25) >> 6;
435 s->block_conv_table[2] = (pitch_range * 44) >> 6;
436 s->block_conv_table[3] = s->max_pitch_val - 1;
437 s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF;
438 if (s->block_delta_pitch_hrange <= 0) {
439 av_log(ctx, AV_LOG_ERROR, "Invalid delta pitch hrange; broken extradata?\n");
440 return AVERROR_INVALIDDATA;
441 }
442 s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange);
443 s->block_pitch_range = s->block_conv_table[2] +
444 s->block_conv_table[3] + 1 +
445 2 * (s->block_conv_table[1] - 2 * s->min_pitch_val);
446 s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range);
447
448 ctx->channels = 1;
449 ctx->channel_layout = AV_CH_LAYOUT_MONO;
450 ctx->sample_fmt = AV_SAMPLE_FMT_FLT;
451
452 return 0;
453 }
454
455 /**
456 * @name Postfilter functions
457 * Postfilter functions (gain control, wiener denoise filter, DC filter,
458 * kalman smoothening, plus surrounding code to wrap it)
459 * @{
460 */
461 /**
462 * Adaptive gain control (as used in postfilter).
463 *
464 * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except
465 * that the energy here is calculated using sum(abs(...)), whereas the
466 * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)).
467 *
468 * @param out output buffer for filtered samples
469 * @param in input buffer containing the samples as they are after the
470 * postfilter steps so far
471 * @param speech_synth input buffer containing speech synth before postfilter
472 * @param size input buffer size
473 * @param alpha exponential filter factor
474 * @param gain_mem pointer to filter memory (single float)
475 */
476 static void adaptive_gain_control(float *out, const float *in,
477 const float *speech_synth,
478 int size, float alpha, float *gain_mem)
479 {
480 int i;
481 float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor;
482 float mem = *gain_mem;
483
484 for (i = 0; i < size; i++) {
485 speech_energy += fabsf(speech_synth[i]);
486 postfilter_energy += fabsf(in[i]);
487 }
488 gain_scale_factor = (1.0 - alpha) * speech_energy / postfilter_energy;
489
490 for (i = 0; i < size; i++) {
491 mem = alpha * mem + gain_scale_factor;
492 out[i] = in[i] * mem;
493 }
494
495 *gain_mem = mem;
496 }
497
498 /**
499 * Kalman smoothing function.
500 *
501 * This function looks back pitch +/- 3 samples back into history to find
502 * the best fitting curve (that one giving the optimal gain of the two
503 * signals, i.e. the highest dot product between the two), and then
504 * uses that signal history to smoothen the output of the speech synthesis
505 * filter.
506 *
507 * @param s WMA Voice decoding context
508 * @param pitch pitch of the speech signal
509 * @param in input speech signal
510 * @param out output pointer for smoothened signal
511 * @param size input/output buffer size
512 *
513 * @returns -1 if no smoothening took place, e.g. because no optimal
514 * fit could be found, or 0 on success.
515 */
516 static int kalman_smoothen(WMAVoiceContext *s, int pitch,
517 const float *in, float *out, int size)
518 {
519 int n;
520 float optimal_gain = 0, dot;
521 const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)],
522 *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)],
523 *best_hist_ptr;
524
525 /* find best fitting point in history */
526 do {
527 dot = avpriv_scalarproduct_float_c(in, ptr, size);
528 if (dot > optimal_gain) {
529 optimal_gain = dot;
530 best_hist_ptr = ptr;
531 }
532 } while (--ptr >= end);
533
534 if (optimal_gain <= 0)
535 return -1;
536 dot = avpriv_scalarproduct_float_c(best_hist_ptr, best_hist_ptr, size);
537 if (dot <= 0) // would be 1.0
538 return -1;
539
540 if (optimal_gain <= dot) {
541 dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000
542 } else
543 dot = 0.625;
544
545 /* actual smoothing */
546 for (n = 0; n < size; n++)
547 out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]);
548
549 return 0;
550 }
551
552 /**
553 * Get the tilt factor of a formant filter from its transfer function
554 * @see #tilt_factor() in amrnbdec.c, which does essentially the same,
555 * but somehow (??) it does a speech synthesis filter in the
556 * middle, which is missing here
557 *
558 * @param lpcs LPC coefficients
559 * @param n_lpcs Size of LPC buffer
560 * @returns the tilt factor
561 */
562 static float tilt_factor(const float *lpcs, int n_lpcs)
563 {
564 float rh0, rh1;
565
566 rh0 = 1.0 + avpriv_scalarproduct_float_c(lpcs, lpcs, n_lpcs);
567 rh1 = lpcs[0] + avpriv_scalarproduct_float_c(lpcs, &lpcs[1], n_lpcs - 1);
568
569 return rh1 / rh0;
570 }
571
572 /**
573 * Derive denoise filter coefficients (in real domain) from the LPCs.
574 */
575 static void calc_input_response(WMAVoiceContext *s, float *lpcs,
576 int fcb_type, float *coeffs, int remainder)
577 {
578 float last_coeff, min = 15.0, max = -15.0;
579 float irange, angle_mul, gain_mul, range, sq;
580 int n, idx;
581
582 /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */
583 s->rdft.rdft_calc(&s->rdft, lpcs);
584 #define log_range(var, assign) do { \
585 float tmp = log10f(assign); var = tmp; \
586 max = FFMAX(max, tmp); min = FFMIN(min, tmp); \
587 } while (0)
588 log_range(last_coeff, lpcs[1] * lpcs[1]);
589 for (n = 1; n < 64; n++)
590 log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] +
591 lpcs[n * 2 + 1] * lpcs[n * 2 + 1]);
592 log_range(lpcs[0], lpcs[0] * lpcs[0]);
593 #undef log_range
594 range = max - min;
595 lpcs[64] = last_coeff;
596
597 /* Now, use this spectrum to pick out these frequencies with higher
598 * (relative) power/energy (which we then take to be "not noise"),
599 * and set up a table (still in lpc[]) of (relative) gains per frequency.
600 * These frequencies will be maintained, while others ("noise") will be
601 * decreased in the filter output. */
602 irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63]
603 gain_mul = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) :
604 (5.0 / 14.7));
605 angle_mul = gain_mul * (8.0 * M_LN10 / M_PI);
606 for (n = 0; n <= 64; n++) {
607 float pwr;
608
609 idx = FFMAX(0, lrint((max - lpcs[n]) * irange) - 1);
610 pwr = wmavoice_denoise_power_table[s->denoise_strength][idx];
611 lpcs[n] = angle_mul * pwr;
612
613 /* 70.57 =~ 1/log10(1.0331663) */
614 idx = (pwr * gain_mul - 0.0295) * 70.570526123;
615 if (idx > 127) { // fall back if index falls outside table range
616 coeffs[n] = wmavoice_energy_table[127] *
617 powf(1.0331663, idx - 127);
618 } else
619 coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)];
620 }
621
622 /* calculate the Hilbert transform of the gains, which we do (since this
623 * is a sinus input) by doing a phase shift (in theory, H(sin())=cos()).
624 * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the
625 * "moment" of the LPCs in this filter. */
626 s->dct.dct_calc(&s->dct, lpcs);
627 s->dst.dct_calc(&s->dst, lpcs);
628
629 /* Split out the coefficient indexes into phase/magnitude pairs */
630 idx = 255 + av_clip(lpcs[64], -255, 255);
631 coeffs[0] = coeffs[0] * s->cos[idx];
632 idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255);
633 last_coeff = coeffs[64] * s->cos[idx];
634 for (n = 63;; n--) {
635 idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255);
636 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
637 coeffs[n * 2] = coeffs[n] * s->cos[idx];
638
639 if (!--n) break;
640
641 idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255);
642 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
643 coeffs[n * 2] = coeffs[n] * s->cos[idx];
644 }
645 coeffs[1] = last_coeff;
646
647 /* move into real domain */
648 s->irdft.rdft_calc(&s->irdft, coeffs);
649
650 /* tilt correction and normalize scale */
651 memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder));
652 if (s->denoise_tilt_corr) {
653 float tilt_mem = 0;
654
655 coeffs[remainder - 1] = 0;
656 ff_tilt_compensation(&tilt_mem,
657 -1.8 * tilt_factor(coeffs, remainder - 1),
658 coeffs, remainder);
659 }
660 sq = (1.0 / 64.0) * sqrtf(1 / avpriv_scalarproduct_float_c(coeffs, coeffs,
661 remainder));
662 for (n = 0; n < remainder; n++)
663 coeffs[n] *= sq;
664 }
665
666 /**
667 * This function applies a Wiener filter on the (noisy) speech signal as
668 * a means to denoise it.
669 *
670 * - take RDFT of LPCs to get the power spectrum of the noise + speech;
671 * - using this power spectrum, calculate (for each frequency) the Wiener
672 * filter gain, which depends on the frequency power and desired level
673 * of noise subtraction (when set too high, this leads to artifacts)
674 * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse
675 * of 4-8kHz);
676 * - by doing a phase shift, calculate the Hilbert transform of this array
677 * of per-frequency filter-gains to get the filtering coefficients;
678 * - smoothen/normalize/de-tilt these filter coefficients as desired;
679 * - take RDFT of noisy sound, apply the coefficients and take its IRDFT
680 * to get the denoised speech signal;
681 * - the leftover (i.e. output of the IRDFT on denoised speech data beyond
682 * the frame boundary) are saved and applied to subsequent frames by an
683 * overlap-add method (otherwise you get clicking-artifacts).
684 *
685 * @param s WMA Voice decoding context
686 * @param fcb_type Frame (codebook) type
687 * @param synth_pf input: the noisy speech signal, output: denoised speech
688 * data; should be 16-byte aligned (for ASM purposes)
689 * @param size size of the speech data
690 * @param lpcs LPCs used to synthesize this frame's speech data
691 */
692 static void wiener_denoise(WMAVoiceContext *s, int fcb_type,
693 float *synth_pf, int size,
694 const float *lpcs)
695 {
696 int remainder, lim, n;
697
698 if (fcb_type != FCB_TYPE_SILENCE) {
699 float *tilted_lpcs = s->tilted_lpcs_pf,
700 *coeffs = s->denoise_coeffs_pf, tilt_mem = 0;
701
702 tilted_lpcs[0] = 1.0;
703 memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps);
704 memset(&tilted_lpcs[s->lsps + 1], 0,
705 sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1));
706 ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps),
707 tilted_lpcs, s->lsps + 2);
708
709 /* The IRDFT output (127 samples for 7-bit filter) beyond the frame
710 * size is applied to the next frame. All input beyond this is zero,
711 * and thus all output beyond this will go towards zero, hence we can
712 * limit to min(size-1, 127-size) as a performance consideration. */
713 remainder = FFMIN(127 - size, size - 1);
714 calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder);
715
716 /* apply coefficients (in frequency spectrum domain), i.e. complex
717 * number multiplication */
718 memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size));
719 s->rdft.rdft_calc(&s->rdft, synth_pf);
720 s->rdft.rdft_calc(&s->rdft, coeffs);
721 synth_pf[0] *= coeffs[0];
722 synth_pf[1] *= coeffs[1];
723 for (n = 1; n < 64; n++) {
724 float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1];
725 synth_pf[n * 2] = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1];
726 synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1];
727 }
728 s->irdft.rdft_calc(&s->irdft, synth_pf);
729 }
730
731 /* merge filter output with the history of previous runs */
732 if (s->denoise_filter_cache_size) {
733 lim = FFMIN(s->denoise_filter_cache_size, size);
734 for (n = 0; n < lim; n++)
735 synth_pf[n] += s->denoise_filter_cache[n];
736 s->denoise_filter_cache_size -= lim;
737 memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size],
738 sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size);
739 }
740
741 /* move remainder of filter output into a cache for future runs */
742 if (fcb_type != FCB_TYPE_SILENCE) {
743 lim = FFMIN(remainder, s->denoise_filter_cache_size);
744 for (n = 0; n < lim; n++)
745 s->denoise_filter_cache[n] += synth_pf[size + n];
746 if (lim < remainder) {
747 memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim],
748 sizeof(s->denoise_filter_cache[0]) * (remainder - lim));
749 s->denoise_filter_cache_size = remainder;
750 }
751 }
752 }
753
754 /**
755 * Averaging projection filter, the postfilter used in WMAVoice.
756 *
757 * This uses the following steps:
758 * - A zero-synthesis filter (generate excitation from synth signal)
759 * - Kalman smoothing on excitation, based on pitch
760 * - Re-synthesized smoothened output
761 * - Iterative Wiener denoise filter
762 * - Adaptive gain filter
763 * - DC filter
764 *
765 * @param s WMAVoice decoding context
766 * @param synth Speech synthesis output (before postfilter)
767 * @param samples Output buffer for filtered samples
768 * @param size Buffer size of synth & samples
769 * @param lpcs Generated LPCs used for speech synthesis
770 * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned)
771 * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses)
772 * @param pitch Pitch of the input signal
773 */
774 static void postfilter(WMAVoiceContext *s, const float *synth,
775 float *samples, int size,
776 const float *lpcs, float *zero_exc_pf,
777 int fcb_type, int pitch)
778 {
779 float synth_filter_in_buf[MAX_FRAMESIZE / 2],
780 *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16],
781 *synth_filter_in = zero_exc_pf;
782
783 assert(size <= MAX_FRAMESIZE / 2);
784
785 /* generate excitation from input signal */
786 ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps);
787
788 if (fcb_type >= FCB_TYPE_AW_PULSES &&
789 !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size))
790 synth_filter_in = synth_filter_in_buf;
791
792 /* re-synthesize speech after smoothening, and keep history */
793 ff_celp_lp_synthesis_filterf(synth_pf, lpcs,
794 synth_filter_in, size, s->lsps);
795 memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps],
796 sizeof(synth_pf[0]) * s->lsps);
797
798 wiener_denoise(s, fcb_type, synth_pf, size, lpcs);
799
800 adaptive_gain_control(samples, synth_pf, synth, size, 0.99,
801 &s->postfilter_agc);
802
803 if (s->dc_level > 8) {
804 /* remove ultra-low frequency DC noise / highpass filter;
805 * coefficients are identical to those used in SIPR decoding,
806 * and very closely resemble those used in AMR-NB decoding. */
807 ff_acelp_apply_order_2_transfer_function(samples, samples,
808 (const float[2]) { -1.99997, 1.0 },
809 (const float[2]) { -1.9330735188, 0.93589198496 },
810 0.93980580475, s->dcf_mem, size);
811 }
812 }
813 /**
814 * @}
815 */
816
817 /**
818 * Dequantize LSPs
819 * @param lsps output pointer to the array that will hold the LSPs
820 * @param num number of LSPs to be dequantized
821 * @param values quantized values, contains n_stages values
822 * @param sizes range (i.e. max value) of each quantized value
823 * @param n_stages number of dequantization runs
824 * @param table dequantization table to be used
825 * @param mul_q LSF multiplier
826 * @param base_q base (lowest) LSF values
827 */
828 static void dequant_lsps(double *lsps, int num,
829 const uint16_t *values,
830 const uint16_t *sizes,
831 int n_stages, const uint8_t *table,
832 const double *mul_q,
833 const double *base_q)
834 {
835 int n, m;
836
837 memset(lsps, 0, num * sizeof(*lsps));
838 for (n = 0; n < n_stages; n++) {
839 const uint8_t *t_off = &table[values[n] * num];
840 double base = base_q[n], mul = mul_q[n];
841
842 for (m = 0; m < num; m++)
843 lsps[m] += base + mul * t_off[m];
844
845 table += sizes[n] * num;
846 }
847 }
848
849 /**
850 * @name LSP dequantization routines
851 * LSP dequantization routines, for 10/16LSPs and independent/residual coding.
852 * @note we assume enough bits are available, caller should check.
853 * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;
854 * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits.
855 * @{
856 */
857 /**
858 * Parse 10 independently-coded LSPs.
859 */
860 static void dequant_lsp10i(GetBitContext *gb, double *lsps)
861 {
862 static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 };
863 static const double mul_lsf[4] = {
864 5.2187144800e-3, 1.4626986422e-3,
865 9.6179549166e-4, 1.1325736225e-3
866 };
867 static const double base_lsf[4] = {
868 M_PI * -2.15522e-1, M_PI * -6.1646e-2,
869 M_PI * -3.3486e-2, M_PI * -5.7408e-2
870 };
871 uint16_t v[4];
872
873 v[0] = get_bits(gb, 8);
874 v[1] = get_bits(gb, 6);
875 v[2] = get_bits(gb, 5);
876 v[3] = get_bits(gb, 5);
877
878 dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i,
879 mul_lsf, base_lsf);
880 }
881
882 /**
883 * Parse 10 independently-coded LSPs, and then derive the tables to
884 * generate LSPs for the other frames from them (residual coding).
885 */
886 static void dequant_lsp10r(GetBitContext *gb,
887 double *i_lsps, const double *old,
888 double *a1, double *a2, int q_mode)
889 {
890 static const uint16_t vec_sizes[3] = { 128, 64, 64 };
891 static const double mul_lsf[3] = {
892 2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3
893 };
894 static const double base_lsf[3] = {
895 M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2
896 };
897 const float (*ipol_tab)[2][10] = q_mode ?
898 wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a;
899 uint16_t interpol, v[3];
900 int n;
901
902 dequant_lsp10i(gb, i_lsps);
903
904 interpol = get_bits(gb, 5);
905 v[0] = get_bits(gb, 7);
906 v[1] = get_bits(gb, 6);
907 v[2] = get_bits(gb, 6);
908
909 for (n = 0; n < 10; n++) {
910 double delta = old[n] - i_lsps[n];
911 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
912 a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
913 }
914
915 dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r,
916 mul_lsf, base_lsf);
917 }
918
919 /**
920 * Parse 16 independently-coded LSPs.
921 */
922 static void dequant_lsp16i(GetBitContext *gb, double *lsps)
923 {
924 static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 };
925 static const double mul_lsf[5] = {
926 3.3439586280e-3, 6.9908173703e-4,
927 3.3216608306e-3, 1.0334960326e-3,
928 3.1899104283e-3
929 };
930 static const double base_lsf[5] = {
931 M_PI * -1.27576e-1, M_PI * -2.4292e-2,
932 M_PI * -1.28094e-1, M_PI * -3.2128e-2,
933 M_PI * -1.29816e-1
934 };
935 uint16_t v[5];
936
937 v[0] = get_bits(gb, 8);
938 v[1] = get_bits(gb, 6);
939 v[2] = get_bits(gb, 7);
940 v[3] = get_bits(gb, 6);
941 v[4] = get_bits(gb, 7);
942
943 dequant_lsps( lsps, 5, v, vec_sizes, 2,
944 wmavoice_dq_lsp16i1, mul_lsf, base_lsf);
945 dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2,
946 wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]);
947 dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1,
948 wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]);
949 }
950
951 /**
952 * Parse 16 independently-coded LSPs, and then derive the tables to
953 * generate LSPs for the other frames from them (residual coding).
954 */
955 static void dequant_lsp16r(GetBitContext *gb,
956 double *i_lsps, const double *old,
957 double *a1, double *a2, int q_mode)
958 {
959 static const uint16_t vec_sizes[3] = { 128, 128, 128 };
960 static const double mul_lsf[3] = {
961 1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3
962 };
963 static const double base_lsf[3] = {
964 M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2
965 };
966 const float (*ipol_tab)[2][16] = q_mode ?
967 wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a;
968 uint16_t interpol, v[3];
969 int n;
970
971 dequant_lsp16i(gb, i_lsps);
972
973 interpol = get_bits(gb, 5);
974 v[0] = get_bits(gb, 7);
975 v[1] = get_bits(gb, 7);
976 v[2] = get_bits(gb, 7);
977
978 for (n = 0; n < 16; n++) {
979 double delta = old[n] - i_lsps[n];
980 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
981 a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
982 }
983
984 dequant_lsps( a2, 10, v, vec_sizes, 1,
985 wmavoice_dq_lsp16r1, mul_lsf, base_lsf);
986 dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1,
987 wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]);
988 dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1,
989 wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]);
990 }
991
992 /**
993 * @}
994 * @name Pitch-adaptive window coding functions
995 * The next few functions are for pitch-adaptive window coding.
996 * @{
997 */
998 /**
999 * Parse the offset of the first pitch-adaptive window pulses, and
1000 * the distribution of pulses between the two blocks in this frame.
1001 * @param s WMA Voice decoding context private data
1002 * @param gb bit I/O context
1003 * @param pitch pitch for each block in this frame
1004 */
1005 static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb,
1006 const int *pitch)
1007 {
1008 static const int16_t start_offset[94] = {
1009 -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11,
1010 13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26,
1011 27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43,
1012 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67,
1013 69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91,
1014 93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115,
1015 117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139,
1016 141, 143, 145, 147, 149, 151, 153, 155, 157, 159
1017 };
1018 int bits, offset;
1019
1020 /* position of pulse */
1021 s->aw_idx_is_ext = 0;
1022 if ((bits = get_bits(gb, 6)) >= 54) {
1023 s->aw_idx_is_ext = 1;
1024 bits += (bits - 54) * 3 + get_bits(gb, 2);
1025 }
1026
1027 /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count
1028 * the distribution of the pulses in each block contained in this frame. */
1029 s->aw_pulse_range = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16;
1030 for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ;
1031 s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0];
1032 s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2;
1033 offset += s->aw_n_pulses[0] * pitch[0];
1034 s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1];
1035 s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2;
1036
1037 /* if continuing from a position before the block, reset position to
1038 * start of block (when corrected for the range over which it can be
1039 * spread in aw_pulse_set1()). */
1040 if (start_offset[bits] < MAX_FRAMESIZE / 2) {
1041 while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0)
1042 s->aw_first_pulse_off[1] -= pitch[1];
1043 if (start_offset[bits] < 0)
1044 while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0)
1045 s->aw_first_pulse_off[0] -= pitch[0];
1046 }
1047 }
1048
1049 /**
1050 * Apply second set of pitch-adaptive window pulses.
1051 * @param s WMA Voice decoding context private data
1052 * @param gb bit I/O context
1053 * @param block_idx block index in frame [0, 1]
1054 * @param fcb structure containing fixed codebook vector info
1055 */
1056 static void aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb,
1057 int block_idx, AMRFixed *fcb)
1058 {
1059 uint16_t use_mask_mem[9]; // only 5 are used, rest is padding
1060 uint16_t *use_mask = use_mask_mem + 2;
1061 /* in this function, idx is the index in the 80-bit (+ padding) use_mask
1062 * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits
1063 * of idx are the position of the bit within a particular item in the
1064 * array (0 being the most significant bit, and 15 being the least
1065 * significant bit), and the remainder (>> 4) is the index in the
1066 * use_mask[]-array. This is faster and uses less memory than using a
1067 * 80-byte/80-int array. */
1068 int pulse_off = s->aw_first_pulse_off[block_idx],
1069 pulse_start, n, idx, range, aidx, start_off = 0;
1070
1071 /* set offset of first pulse to within this block */
1072 if (s->aw_n_pulses[block_idx] > 0)
1073 while (pulse_off + s->aw_pulse_range < 1)
1074 pulse_off += fcb->pitch_lag;
1075
1076 /* find range per pulse */
1077 if (s->aw_n_pulses[0] > 0) {
1078 if (block_idx == 0) {
1079 range = 32;
1080 } else /* block_idx = 1 */ {
1081 range = 8;
1082 if (s->aw_n_pulses[block_idx] > 0)
1083 pulse_off = s->aw_next_pulse_off_cache;
1084 }
1085 } else
1086 range = 16;
1087 pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0;
1088
1089 /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,
1090 * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus
1091 * we exclude that range from being pulsed again in this function. */
1092 memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0]));
1093 memset( use_mask, -1, 5 * sizeof(use_mask[0]));
1094 memset(&use_mask[5], 0, 2 * sizeof(use_mask[0]));
1095 if (s->aw_n_pulses[block_idx] > 0)
1096 for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) {
1097 int excl_range = s->aw_pulse_range; // always 16 or 24
1098 uint16_t *use_mask_ptr = &use_mask[idx >> 4];
1099 int first_sh = 16 - (idx & 15);
1100 *use_mask_ptr++ &= 0xFFFFu << first_sh;
1101 excl_range -= first_sh;
1102 if (excl_range >= 16) {
1103 *use_mask_ptr++ = 0;
1104 *use_mask_ptr &= 0xFFFF >> (excl_range - 16);
1105 } else
1106 *use_mask_ptr &= 0xFFFF >> excl_range;
1107 }
1108
1109 /* find the 'aidx'th offset that is not excluded */
1110 aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4);
1111 for (n = 0; n <= aidx; pulse_start++) {
1112 for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ;
1113 if (idx >= MAX_FRAMESIZE / 2) { // find from zero
1114 if (use_mask[0]) idx = 0x0F;
1115 else if (use_mask[1]) idx = 0x1F;
1116 else if (use_mask[2]) idx = 0x2F;
1117 else if (use_mask[3]) idx = 0x3F;
1118 else if (use_mask[4]) idx = 0x4F;
1119 else return;
1120 idx -= av_log2_16bit(use_mask[idx >> 4]);
1121 }
1122 if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) {
1123 use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15));
1124 n++;
1125 start_off = idx;
1126 }
1127 }
1128
1129 fcb->x[fcb->n] = start_off;
1130 fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0;
1131 fcb->n++;
1132
1133 /* set offset for next block, relative to start of that block */
1134 n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag;
1135 s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 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 assert(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 assert(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 aw_pulse_set2(s, gb, block_idx, &fcb);
1299 } else /* FCB_TYPE_EXC_PULSES */ {
1300 int offset_nbits = 5 - frame_desc->log_n_blocks;
1301
1302 fcb.no_repeat_mask = -1;
1303 /* similar to ff_decode_10_pulses_35bits(), but with single pulses
1304 * (instead of double) for a subset of pulses */
1305 for (n = 0; n < 5; n++) {
1306 float sign;
1307 int pos1, pos2;
1308
1309 sign = get_bits1(gb) ? 1.0 : -1.0;
1310 pos1 = get_bits(gb, offset_nbits);
1311 fcb.x[fcb.n] = n + 5 * pos1;
1312 fcb.y[fcb.n++] = sign;
1313 if (n < frame_desc->dbl_pulses) {
1314 pos2 = get_bits(gb, offset_nbits);
1315 fcb.x[fcb.n] = n + 5 * pos2;
1316 fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign;
1317 }
1318 }
1319 }
1320 ff_set_fixed_vector(pulses, &fcb, 1.0, size);
1321
1322 /* Calculate gain for adaptive & fixed codebook signal.
1323 * see ff_amr_set_fixed_gain(). */
1324 idx = get_bits(gb, 7);
1325 fcb_gain = expf(avpriv_scalarproduct_float_c(s->gain_pred_err,
1326 gain_coeff, 6) -
1327 5.2409161640 + wmavoice_gain_codebook_fcb[idx]);
1328 acb_gain = wmavoice_gain_codebook_acb[idx];
1329 pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx],
1330 -2.9957322736 /* log(0.05) */,
1331 1.6094379124 /* log(5.0) */);
1332
1333 gain_weight = 8 >> frame_desc->log_n_blocks;
1334 memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err,
1335 sizeof(*s->gain_pred_err) * (6 - gain_weight));
1336 for (n = 0; n < gain_weight; n++)
1337 s->gain_pred_err[n] = pred_err;
1338
1339 /* Calculation of adaptive codebook */
1340 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1341 int len;
1342 for (n = 0; n < size; n += len) {
1343 int next_idx_sh16;
1344 int abs_idx = block_idx * size + n;
1345 int pitch_sh16 = (s->last_pitch_val << 16) +
1346 s->pitch_diff_sh16 * abs_idx;
1347 int pitch = (pitch_sh16 + 0x6FFF) >> 16;
1348 int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000;
1349 idx = idx_sh16 >> 16;
1350 if (s->pitch_diff_sh16) {
1351 if (s->pitch_diff_sh16 > 0) {
1352 next_idx_sh16 = (idx_sh16) &~ 0xFFFF;
1353 } else
1354 next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF;
1355 len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8,
1356 1, size - n);
1357 } else
1358 len = size;
1359
1360 ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch],
1361 wmavoice_ipol1_coeffs, 17,
1362 idx, 9, len);
1363 }
1364 } else /* ACB_TYPE_HAMMING */ {
1365 int block_pitch = block_pitch_sh2 >> 2;
1366 idx = block_pitch_sh2 & 3;
1367 if (idx) {
1368 ff_acelp_interpolatef(excitation, &excitation[-block_pitch],
1369 wmavoice_ipol2_coeffs, 4,
1370 idx, 8, size);
1371 } else
1372 av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch,
1373 sizeof(float) * size);
1374 }
1375
1376 /* Interpolate ACB/FCB and use as excitation signal */
1377 ff_weighted_vector_sumf(excitation, excitation, pulses,
1378 acb_gain, fcb_gain, size);
1379 }
1380
1381 /**
1382 * Parse data in a single block.
1383 * @note we assume enough bits are available, caller should check.
1384 *
1385 * @param s WMA Voice decoding context private data
1386 * @param gb bit I/O context
1387 * @param block_idx index of the to-be-read block
1388 * @param size amount of samples to be read in this block
1389 * @param block_pitch_sh2 pitch for this block << 2
1390 * @param lsps LSPs for (the end of) this frame
1391 * @param prev_lsps LSPs for the last frame
1392 * @param frame_desc frame type descriptor
1393 * @param excitation target memory for the ACB+FCB interpolated signal
1394 * @param synth target memory for the speech synthesis filter output
1395 * @return 0 on success, <0 on error.
1396 */
1397 static void synth_block(WMAVoiceContext *s, GetBitContext *gb,
1398 int block_idx, int size,
1399 int block_pitch_sh2,
1400 const double *lsps, const double *prev_lsps,
1401 const struct frame_type_desc *frame_desc,
1402 float *excitation, float *synth)
1403 {
1404 double i_lsps[MAX_LSPS];
1405 float lpcs[MAX_LSPS];
1406 float fac;
1407 int n;
1408
1409 if (frame_desc->acb_type == ACB_TYPE_NONE)
1410 synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation);
1411 else
1412 synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2,
1413 frame_desc, excitation);
1414
1415 /* convert interpolated LSPs to LPCs */
1416 fac = (block_idx + 0.5) / frame_desc->n_blocks;
1417 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1418 i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n]));
1419 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1420
1421 /* Speech synthesis */
1422 ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps);
1423 }
1424
1425 /**
1426 * Synthesize output samples for a single frame.
1427 * @note we assume enough bits are available, caller should check.
1428 *
1429 * @param ctx WMA Voice decoder context
1430 * @param gb bit I/O context (s->gb or one for cross-packet superframes)
1431 * @param frame_idx Frame number within superframe [0-2]
1432 * @param samples pointer to output sample buffer, has space for at least 160
1433 * samples
1434 * @param lsps LSP array
1435 * @param prev_lsps array of previous frame's LSPs
1436 * @param excitation target buffer for excitation signal
1437 * @param synth target buffer for synthesized speech data
1438 * @return 0 on success, <0 on error.
1439 */
1440 static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx,
1441 float *samples,
1442 const double *lsps, const double *prev_lsps,
1443 float *excitation, float *synth)
1444 {
1445 WMAVoiceContext *s = ctx->priv_data;
1446 int n, n_blocks_x2, log_n_blocks_x2, cur_pitch_val;
1447 int pitch[MAX_BLOCKS], last_block_pitch;
1448
1449 /* Parse frame type ("frame header"), see frame_descs */
1450 int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)], block_nsamples;
1451
1452 if (bd_idx < 0) {
1453 av_log(ctx, AV_LOG_ERROR,
1454 "Invalid frame type VLC code, skipping\n");
1455 return AVERROR_INVALIDDATA;
1456 }
1457
1458 block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks;
1459
1460 /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */
1461 if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) {
1462 /* Pitch is provided per frame, which is interpreted as the pitch of
1463 * the last sample of the last block of this frame. We can interpolate
1464 * the pitch of other blocks (and even pitch-per-sample) by gradually
1465 * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */
1466 n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1;
1467 log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1;
1468 cur_pitch_val = s->min_pitch_val + get_bits(gb, s->pitch_nbits);
1469 cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1);
1470 if (s->last_acb_type == ACB_TYPE_NONE ||
1471 20 * abs(cur_pitch_val - s->last_pitch_val) >
1472 (cur_pitch_val + s->last_pitch_val))
1473 s->last_pitch_val = cur_pitch_val;
1474
1475 /* pitch per block */
1476 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1477 int fac = n * 2 + 1;
1478
1479 pitch[n] = (MUL16(fac, cur_pitch_val) +
1480 MUL16((n_blocks_x2 - fac), s->last_pitch_val) +
1481 frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2;
1482 }
1483
1484 /* "pitch-diff-per-sample" for calculation of pitch per sample */
1485 s->pitch_diff_sh16 =
1486 ((cur_pitch_val - s->last_pitch_val) << 16) / MAX_FRAMESIZE;
1487 }
1488
1489 /* Global gain (if silence) and pitch-adaptive window coordinates */
1490 switch (frame_descs[bd_idx].fcb_type) {
1491 case FCB_TYPE_SILENCE:
1492 s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)];
1493 break;
1494 case FCB_TYPE_AW_PULSES:
1495 aw_parse_coords(s, gb, pitch);
1496 break;
1497 }
1498
1499 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1500 int bl_pitch_sh2;
1501
1502 /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */
1503 switch (frame_descs[bd_idx].acb_type) {
1504 case ACB_TYPE_HAMMING: {
1505 /* Pitch is given per block. Per-block pitches are encoded as an
1506 * absolute value for the first block, and then delta values
1507 * relative to this value) for all subsequent blocks. The scale of
1508 * this pitch value is semi-logaritmic compared to its use in the
1509 * decoder, so we convert it to normal scale also. */
1510 int block_pitch,
1511 t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2,
1512 t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1,
1513 t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1;
1514
1515 if (n == 0) {
1516 block_pitch = get_bits(gb, s->block_pitch_nbits);
1517 } else
1518 block_pitch = last_block_pitch - s->block_delta_pitch_hrange +
1519 get_bits(gb, s->block_delta_pitch_nbits);
1520 /* Convert last_ so that any next delta is within _range */
1521 last_block_pitch = av_clip(block_pitch,
1522 s->block_delta_pitch_hrange,
1523 s->block_pitch_range -
1524 s->block_delta_pitch_hrange);
1525
1526 /* Convert semi-log-style scale back to normal scale */
1527 if (block_pitch < t1) {
1528 bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch;
1529 } else {
1530 block_pitch -= t1;
1531 if (block_pitch < t2) {
1532 bl_pitch_sh2 =
1533 (s->block_conv_table[1] << 2) + (block_pitch << 1);
1534 } else {
1535 block_pitch -= t2;
1536 if (block_pitch < t3) {
1537 bl_pitch_sh2 =
1538 (s->block_conv_table[2] + block_pitch) << 2;
1539 } else
1540 bl_pitch_sh2 = s->block_conv_table[3] << 2;
1541 }
1542 }
1543 pitch[n] = bl_pitch_sh2 >> 2;
1544 break;
1545 }
1546
1547 case ACB_TYPE_ASYMMETRIC: {
1548 bl_pitch_sh2 = pitch[n] << 2;
1549 break;
1550 }
1551
1552 default: // ACB_TYPE_NONE has no pitch
1553 bl_pitch_sh2 = 0;
1554 break;
1555 }
1556
1557 synth_block(s, gb, n, block_nsamples, bl_pitch_sh2,
1558 lsps, prev_lsps, &frame_descs[bd_idx],
1559 &excitation[n * block_nsamples],
1560 &synth[n * block_nsamples]);
1561 }
1562
1563 /* Averaging projection filter, if applicable. Else, just copy samples
1564 * from synthesis buffer */
1565 if (s->do_apf) {
1566 double i_lsps[MAX_LSPS];
1567 float lpcs[MAX_LSPS];
1568
1569 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1570 i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n]));
1571 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1572 postfilter(s, synth, samples, 80, lpcs,
1573 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx],
1574 frame_descs[bd_idx].fcb_type, pitch[0]);
1575
1576 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1577 i_lsps[n] = cos(lsps[n]);
1578 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1579 postfilter(s, &synth[80], &samples[80], 80, lpcs,
1580 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80],
1581 frame_descs[bd_idx].fcb_type, pitch[0]);
1582 } else
1583 memcpy(samples, synth, 160 * sizeof(synth[0]));
1584
1585 /* Cache values for next frame */
1586 s->frame_cntr++;
1587 if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%)
1588 s->last_acb_type = frame_descs[bd_idx].acb_type;
1589 switch (frame_descs[bd_idx].acb_type) {
1590 case ACB_TYPE_NONE:
1591 s->last_pitch_val = 0;
1592 break;
1593 case ACB_TYPE_ASYMMETRIC:
1594 s->last_pitch_val = cur_pitch_val;
1595 break;
1596 case ACB_TYPE_HAMMING:
1597 s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1];
1598 break;
1599 }
1600
1601 return 0;
1602 }
1603
1604 /**
1605 * Ensure minimum value for first item, maximum value for last value,
1606 * proper spacing between each value and proper ordering.
1607 *
1608 * @param lsps array of LSPs
1609 * @param num size of LSP array
1610 *
1611 * @note basically a double version of #ff_acelp_reorder_lsf(), might be
1612 * useful to put in a generic location later on. Parts are also
1613 * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(),
1614 * which is in float.
1615 */
1616 static void stabilize_lsps(double *lsps, int num)
1617 {
1618 int n, m, l;
1619
1620 /* set minimum value for first, maximum value for last and minimum
1621 * spacing between LSF values.
1622 * Very similar to ff_set_min_dist_lsf(), but in double. */
1623 lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI);
1624 for (n = 1; n < num; n++)
1625 lsps[n] = FFMAX(lsps[n], lsps[n - 1] + 0.0125 * M_PI);
1626 lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI);
1627
1628 /* reorder (looks like one-time / non-recursed bubblesort).
1629 * Very similar to ff_sort_nearly_sorted_floats(), but in double. */
1630 for (n = 1; n < num; n++) {
1631 if (lsps[n] < lsps[n - 1]) {
1632 for (m = 1; m < num; m++) {
1633 double tmp = lsps[m];
1634 for (l = m - 1; l >= 0; l--) {
1635 if (lsps[l] <= tmp) break;
1636 lsps[l + 1] = lsps[l];
1637 }
1638 lsps[l + 1] = tmp;
1639 }
1640 break;
1641 }
1642 }
1643 }
1644
1645 /**
1646 * Test if there's enough bits to read 1 superframe.
1647 *
1648 * @param orig_gb bit I/O context used for reading. This function
1649 * does not modify the state of the bitreader; it
1650 * only uses it to copy the current stream position
1651 * @param s WMA Voice decoding context private data
1652 * @return < 0 on error, 1 on not enough bits or 0 if OK.
1653 */
1654 static int check_bits_for_superframe(GetBitContext *orig_gb,
1655 WMAVoiceContext *s)
1656 {
1657 GetBitContext s_gb, *gb = &s_gb;
1658 int n, need_bits, bd_idx;
1659 const struct frame_type_desc *frame_desc;
1660
1661 /* initialize a copy */
1662 init_get_bits(gb, orig_gb->buffer, orig_gb->size_in_bits);
1663 skip_bits_long(gb, get_bits_count(orig_gb));
1664 assert(get_bits_left(gb) == get_bits_left(orig_gb));
1665
1666 /* superframe header */
1667 if (get_bits_left(gb) < 14)
1668 return 1;
1669 if (!get_bits1(gb))
1670 return AVERROR(ENOSYS); // WMAPro-in-WMAVoice superframe
1671 if (get_bits1(gb)) skip_bits(gb, 12); // number of samples in superframe
1672 if (s->has_residual_lsps) { // residual LSPs (for all frames)
1673 if (get_bits_left(gb) < s->sframe_lsp_bitsize)
1674 return 1;
1675 skip_bits_long(gb, s->sframe_lsp_bitsize);
1676 }
1677
1678 /* frames */
1679 for (n = 0; n < MAX_FRAMES; n++) {
1680 int aw_idx_is_ext = 0;
1681
1682 if (!s->has_residual_lsps) { // independent LSPs (per-frame)
1683 if (get_bits_left(gb) < s->frame_lsp_bitsize) return 1;
1684 skip_bits_long(gb, s->frame_lsp_bitsize);
1685 }
1686 bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)];
1687 if (bd_idx < 0)
1688 return AVERROR_INVALIDDATA; // invalid frame type VLC code
1689 frame_desc = &frame_descs[bd_idx];
1690 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1691 if (get_bits_left(gb) < s->pitch_nbits)
1692 return 1;
1693 skip_bits_long(gb, s->pitch_nbits);
1694 }
1695 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1696 skip_bits(gb, 8);
1697 } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1698 int tmp = get_bits(gb, 6);
1699 if (tmp >= 0x36) {
1700 skip_bits(gb, 2);
1701 aw_idx_is_ext = 1;
1702 }
1703 }
1704
1705 /* blocks */
1706 if (frame_desc->acb_type == ACB_TYPE_HAMMING) {
1707 need_bits = s->block_pitch_nbits +
1708 (frame_desc->n_blocks - 1) * s->block_delta_pitch_nbits;
1709 } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1710 need_bits = 2 * !aw_idx_is_ext;
1711 } else
1712 need_bits = 0;
1713 need_bits += frame_desc->frame_size;
1714 if (get_bits_left(gb) < need_bits)
1715 return 1;
1716 skip_bits_long(gb, need_bits);
1717 }
1718
1719 return 0;
1720 }
1721
1722 /**
1723 * Synthesize output samples for a single superframe. If we have any data
1724 * cached in s->sframe_cache, that will be used instead of whatever is loaded
1725 * in s->gb.
1726 *
1727 * WMA Voice superframes contain 3 frames, each containing 160 audio samples,
1728 * to give a total of 480 samples per frame. See #synth_frame() for frame
1729 * parsing. In addition to 3 frames, superframes can also contain the LSPs
1730 * (if these are globally specified for all frames (residually); they can
1731 * also be specified individually per-frame. See the s->has_residual_lsps
1732 * option), and can specify the number of samples encoded in this superframe
1733 * (if less than 480), usually used to prevent blanks at track boundaries.
1734 *
1735 * @param ctx WMA Voice decoder context
1736 * @return 0 on success, <0 on error or 1 if there was not enough data to
1737 * fully parse the superframe
1738 */
1739 static int synth_superframe(AVCodecContext *ctx, AVFrame *frame,
1740 int *got_frame_ptr)
1741 {
1742 WMAVoiceContext *s = ctx->priv_data;
1743 GetBitContext *gb = &s->gb, s_gb;
1744 int n, res, n_samples = 480;
1745 double lsps[MAX_FRAMES][MAX_LSPS];
1746 const double *mean_lsf = s->lsps == 16 ?
1747 wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode];
1748 float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12];
1749 float synth[MAX_LSPS + MAX_SFRAMESIZE];
1750 float *samples;
1751
1752 memcpy(synth, s->synth_history,
1753 s->lsps * sizeof(*synth));
1754 memcpy(excitation, s->excitation_history,
1755 s->history_nsamples * sizeof(*excitation));
1756
1757 if (s->sframe_cache_size > 0) {
1758 gb = &s_gb;
1759 init_get_bits(gb, s->sframe_cache, s->sframe_cache_size);
1760 s->sframe_cache_size = 0;
1761 }
1762
1763 if ((res = check_bits_for_superframe(gb, s)) == 1) {
1764 *got_frame_ptr = 0;
1765 return 1;
1766 } else if (res < 0)
1767 return res;
1768
1769 /* First bit is speech/music bit, it differentiates between WMAVoice
1770 * speech samples (the actual codec) and WMAVoice music samples, which
1771 * are really WMAPro-in-WMAVoice-superframes. I've never seen those in
1772 * the wild yet. */
1773 if (!get_bits1(gb)) {
1774 avpriv_request_sample(ctx, "WMAPro-in-WMAVoice");
1775 return AVERROR_PATCHWELCOME;
1776 }
1777
1778 /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */
1779 if (get_bits1(gb)) {
1780 if ((n_samples = get_bits(gb, 12)) > 480) {
1781 av_log(ctx, AV_LOG_ERROR,
1782 "Superframe encodes >480 samples (%d), not allowed\n",
1783 n_samples);
1784 return AVERROR_INVALIDDATA;
1785 }
1786 }
1787 /* Parse LSPs, if global for the superframe (can also be per-frame). */
1788 if (s->has_residual_lsps) {
1789 double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2];
1790
1791 for (n = 0; n < s->lsps; n++)
1792 prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n];
1793
1794 if (s->lsps == 10) {
1795 dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1796 } else /* s->lsps == 16 */
1797 dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1798
1799 for (n = 0; n < s->lsps; n++) {
1800 lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]);
1801 lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]);
1802 lsps[2][n] += mean_lsf[n];
1803 }
1804 for (n = 0; n < 3; n++)
1805 stabilize_lsps(lsps[n], s->lsps);
1806 }
1807
1808 /* get output buffer */
1809 frame->nb_samples = 480;
1810 if ((res = ff_get_buffer(ctx, frame, 0)) < 0) {
1811 av_log(ctx, AV_LOG_ERROR, "get_buffer() failed\n");
1812 return res;
1813 }
1814 frame->nb_samples = n_samples;
1815 samples = (float *)frame->data[0];
1816
1817 /* Parse frames, optionally preceded by per-frame (independent) LSPs. */
1818 for (n = 0; n < 3; n++) {
1819 if (!s->has_residual_lsps) {
1820 int m;
1821
1822 if (s->lsps == 10) {
1823 dequant_lsp10i(gb, lsps[n]);
1824 } else /* s->lsps == 16 */
1825 dequant_lsp16i(gb, lsps[n]);
1826
1827 for (m = 0; m < s->lsps; m++)
1828 lsps[n][m] += mean_lsf[m];
1829 stabilize_lsps(lsps[n], s->lsps);
1830 }
1831
1832 if ((res = synth_frame(ctx, gb, n,
1833 &samples[n * MAX_FRAMESIZE],
1834 lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1],
1835 &excitation[s->history_nsamples + n * MAX_FRAMESIZE],
1836 &synth[s->lsps + n * MAX_FRAMESIZE]))) {
1837 *got_frame_ptr = 0;
1838 return res;
1839 }
1840 }
1841
1842 /* Statistics? FIXME - we don't check for length, a slight overrun
1843 * will be caught by internal buffer padding, and anything else
1844 * will be skipped, not read. */
1845 if (get_bits1(gb)) {
1846 res = get_bits(gb, 4);
1847 skip_bits(gb, 10 * (res + 1));
1848 }
1849
1850 *got_frame_ptr = 1;
1851
1852 /* Update history */
1853 memcpy(s->prev_lsps, lsps[2],
1854 s->lsps * sizeof(*s->prev_lsps));
1855 memcpy(s->synth_history, &synth[MAX_SFRAMESIZE],
1856 s->lsps * sizeof(*synth));
1857 memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE],
1858 s->history_nsamples * sizeof(*excitation));
1859 if (s->do_apf)
1860 memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE],
1861 s->history_nsamples * sizeof(*s->zero_exc_pf));
1862
1863 return 0;
1864 }
1865
1866 /**
1867 * Parse the packet header at the start of each packet (input data to this
1868 * decoder).
1869 *
1870 * @param s WMA Voice decoding context private data
1871 * @return 1 if not enough bits were available, or 0 on success.
1872 */
1873 static int parse_packet_header(WMAVoiceContext *s)
1874 {
1875 GetBitContext *gb = &s->gb;
1876 unsigned int res;
1877
1878 if (get_bits_left(gb) < 11)
1879 return 1;
1880 skip_bits(gb, 4); // packet sequence number
1881 s->has_residual_lsps = get_bits1(gb);
1882 do {
1883 res = get_bits(gb, 6); // number of superframes per packet
1884 // (minus first one if there is spillover)
1885 if (get_bits_left(gb) < 6 * (res == 0x3F) + s->spillover_bitsize)
1886 return 1;
1887 } while (res == 0x3F);
1888 s->spillover_nbits = get_bits(gb, s->spillover_bitsize);
1889
1890 return 0;
1891 }
1892
1893 /**
1894 * Copy (unaligned) bits from gb/data/size to pb.
1895 *
1896 * @param pb target buffer to copy bits into
1897 * @param data source buffer to copy bits from
1898 * @param size size of the source data, in bytes
1899 * @param gb bit I/O context specifying the current position in the source.
1900 * data. This function might use this to align the bit position to
1901 * a whole-byte boundary before calling #avpriv_copy_bits() on aligned
1902 * source data
1903 * @param nbits the amount of bits to copy from source to target
1904 *
1905 * @note after calling this function, the current position in the input bit
1906 * I/O context is undefined.
1907 */
1908 static void copy_bits(PutBitContext *pb,
1909 const uint8_t *data, int size,
1910 GetBitContext *gb, int nbits)
1911 {
1912 int rmn_bytes, rmn_bits;
1913
1914 rmn_bits = rmn_bytes = get_bits_left(gb);
1915 if (rmn_bits < nbits)
1916 return;
1917 if (nbits > pb->size_in_bits - put_bits_count(pb))
1918 return;
1919 rmn_bits &= 7; rmn_bytes >>= 3;
1920 if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0)
1921 put_bits(pb, rmn_bits, get_bits(gb, rmn_bits));
1922 avpriv_copy_bits(pb, data + size - rmn_bytes,
1923 FFMIN(nbits - rmn_bits, rmn_bytes << 3));
1924 }
1925
1926 /**
1927 * Packet decoding: a packet is anything that the (ASF) demuxer contains,
1928 * and we expect that the demuxer / application provides it to us as such
1929 * (else you'll probably get garbage as output). Every packet has a size of
1930 * ctx->block_align bytes, starts with a packet header (see
1931 * #parse_packet_header()), and then a series of superframes. Superframe
1932 * boundaries may exceed packets, i.e. superframes can split data over
1933 * multiple (two) packets.
1934 *
1935 * For more information about frames, see #synth_superframe().
1936 */
1937 static int wmavoice_decode_packet(AVCodecContext *ctx, void *data,
1938 int *got_frame_ptr, AVPacket *avpkt)
1939 {
1940 WMAVoiceContext *s = ctx->priv_data;
1941 GetBitContext *gb = &s->gb;
1942 int size, res, pos;
1943
1944 /* Packets are sometimes a multiple of ctx->block_align, with a packet
1945 * header at each ctx->block_align bytes. However, Libav's ASF demuxer
1946 * feeds us ASF packets, which may concatenate multiple "codec" packets
1947 * in a single "muxer" packet, so we artificially emulate that by
1948 * capping the packet size at ctx->block_align. */
1949 for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align);
1950 if (!size) {
1951 *got_frame_ptr = 0;
1952 return 0;
1953 }
1954 init_get_bits(&s->gb, avpkt->data, size << 3);
1955
1956 /* size == ctx->block_align is used to indicate whether we are dealing with
1957 * a new packet or a packet of which we already read the packet header
1958 * previously. */
1959 if (size == ctx->block_align) { // new packet header
1960 if ((res = parse_packet_header(s)) < 0)
1961 return res;
1962
1963 /* If the packet header specifies a s->spillover_nbits, then we want
1964 * to push out all data of the previous packet (+ spillover) before
1965 * continuing to parse new superframes in the current packet. */
1966 if (s->spillover_nbits > 0) {
1967 if (s->sframe_cache_size > 0) {
1968 int cnt = get_bits_count(gb);
1969 copy_bits(&s->pb, avpkt->data, size, gb, s->spillover_nbits);
1970 flush_put_bits(&s->pb);
1971 s->sframe_cache_size += s->spillover_nbits;
1972 if ((res = synth_superframe(ctx, data, got_frame_ptr)) == 0 &&
1973 *got_frame_ptr) {
1974 cnt += s->spillover_nbits;
1975 s->skip_bits_next = cnt & 7;
1976 return cnt >> 3;
1977 } else
1978 skip_bits_long (gb, s->spillover_nbits - cnt +
1979 get_bits_count(gb)); // resync
1980 } else
1981 skip_bits_long(gb, s->spillover_nbits); // resync
1982 }
1983 } else if (s->skip_bits_next)
1984 skip_bits(gb, s->skip_bits_next);
1985
1986 /* Try parsing superframes in current packet */
1987 s->sframe_cache_size = 0;
1988 s->skip_bits_next = 0;
1989 pos = get_bits_left(gb);
1990 if ((res = synth_superframe(ctx, data, got_frame_ptr)) < 0) {
1991 return res;
1992 } else if (*got_frame_ptr) {
1993 int cnt = get_bits_count(gb);
1994 s->skip_bits_next = cnt & 7;
1995 return cnt >> 3;
1996 } else if ((s->sframe_cache_size = pos) > 0) {
1997 /* rewind bit reader to start of last (incomplete) superframe... */
1998 init_get_bits(gb, avpkt->data, size << 3);
1999 skip_bits_long(gb, (size << 3) - pos);
2000 assert(get_bits_left(gb) == pos);
2001
2002 /* ...and cache it for spillover in next packet */
2003 init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE);
2004 copy_bits(&s->pb, avpkt->data, size, gb, s->sframe_cache_size);
2005 // FIXME bad - just copy bytes as whole and add use the
2006 // skip_bits_next field
2007 }
2008
2009 return size;
2010 }
2011
2012 static av_cold int wmavoice_decode_end(AVCodecContext *ctx)
2013 {
2014 WMAVoiceContext *s = ctx->priv_data;
2015
2016 if (s->do_apf) {
2017 ff_rdft_end(&s->rdft);
2018 ff_rdft_end(&s->irdft);
2019 ff_dct_end(&s->dct);
2020 ff_dct_end(&s->dst);
2021 }
2022
2023 return 0;
2024 }
2025
2026 static av_cold void wmavoice_flush(AVCodecContext *ctx)
2027 {
2028 WMAVoiceContext *s = ctx->priv_data;
2029 int n;
2030
2031 s->postfilter_agc = 0;
2032 s->sframe_cache_size = 0;
2033 s->skip_bits_next = 0;
2034 for (n = 0; n < s->lsps; n++)
2035 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
2036 memset(s->excitation_history, 0,
2037 sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY);
2038 memset(s->synth_history, 0,
2039 sizeof(*s->synth_history) * MAX_LSPS);
2040 memset(s->gain_pred_err, 0,
2041 sizeof(s->gain_pred_err));
2042
2043 if (s->do_apf) {
2044 memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0,
2045 sizeof(*s->synth_filter_out_buf) * s->lsps);
2046 memset(s->dcf_mem, 0,
2047 sizeof(*s->dcf_mem) * 2);
2048 memset(s->zero_exc_pf, 0,
2049 sizeof(*s->zero_exc_pf) * s->history_nsamples);
2050 memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache));
2051 }
2052 }
2053
2054 AVCodec ff_wmavoice_decoder = {
2055 .name = "wmavoice",
2056 .type = AVMEDIA_TYPE_AUDIO,
2057 .id = AV_CODEC_ID_WMAVOICE,
2058 .priv_data_size = sizeof(WMAVoiceContext),
2059 .init = wmavoice_decode_init,
2060 .init_static_data = wmavoice_init_static_data,
2061 .close = wmavoice_decode_end,
2062 .decode = wmavoice_decode_packet,
2063 .capabilities = CODEC_CAP_SUBFRAMES | CODEC_CAP_DR1,
2064 .flush = wmavoice_flush,
2065 .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"),
2066 };