1 /*
2 * Windows Media Audio Voice decoder.
3 * Copyright (c) 2009 Ronald S. Bultje
4 *
5 * This file is part of FFmpeg.
6 *
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
11 *
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
16 *
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20 */
21
22 /**
23 * @file
24 * @brief Windows Media Audio Voice compatible decoder
25 * @author Ronald S. Bultje <rsbultje@gmail.com>
26 */
27
28 #include <math.h>
29
30 #include "libavutil/channel_layout.h"
31 #include "libavutil/float_dsp.h"
32 #include "libavutil/mem_internal.h"
33 #include "libavutil/thread.h"
34 #include "avcodec.h"
35 #include "internal.h"
36 #include "get_bits.h"
37 #include "put_bits.h"
38 #include "wmavoice_data.h"
39 #include "celp_filters.h"
40 #include "acelp_vectors.h"
41 #include "acelp_filters.h"
42 #include "lsp.h"
43 #include "dct.h"
44 #include "rdft.h"
45 #include "sinewin.h"
46
47 #define MAX_BLOCKS 8 ///< maximum number of blocks per frame
48 #define MAX_LSPS 16 ///< maximum filter order
49 #define MAX_LSPS_ALIGN16 16 ///< same as #MAX_LSPS; needs to be multiple
50 ///< of 16 for ASM input buffer alignment
51 #define MAX_FRAMES 3 ///< maximum number of frames per superframe
52 #define MAX_FRAMESIZE 160 ///< maximum number of samples per frame
53 #define MAX_SIGNAL_HISTORY 416 ///< maximum excitation signal history
54 #define MAX_SFRAMESIZE (MAX_FRAMESIZE * MAX_FRAMES)
55 ///< maximum number of samples per superframe
56 #define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that
57 ///< was split over two packets
58 #define VLC_NBITS 6 ///< number of bits to read per VLC iteration
59
60 /**
61 * Frame type VLC coding.
62 */
63 static VLC frame_type_vlc;
64
65 /**
66 * Adaptive codebook types.
67 */
68 enum {
69 ACB_TYPE_NONE = 0, ///< no adaptive codebook (only hardcoded fixed)
70 ACB_TYPE_ASYMMETRIC = 1, ///< adaptive codebook with per-frame pitch, which
71 ///< we interpolate to get a per-sample pitch.
72 ///< Signal is generated using an asymmetric sinc
73 ///< window function
74 ///< @note see #wmavoice_ipol1_coeffs
75 ACB_TYPE_HAMMING = 2 ///< Per-block pitch with signal generation using
76 ///< a Hamming sinc window function
77 ///< @note see #wmavoice_ipol2_coeffs
78 };
79
80 /**
81 * Fixed codebook types.
82 */
83 enum {
84 FCB_TYPE_SILENCE = 0, ///< comfort noise during silence
85 ///< generated from a hardcoded (fixed) codebook
86 ///< with per-frame (low) gain values
87 FCB_TYPE_HARDCODED = 1, ///< hardcoded (fixed) codebook with per-block
88 ///< gain values
89 FCB_TYPE_AW_PULSES = 2, ///< Pitch-adaptive window (AW) pulse signals,
90 ///< used in particular for low-bitrate streams
91 FCB_TYPE_EXC_PULSES = 3, ///< Innovation (fixed) codebook pulse sets in
92 ///< combinations of either single pulses or
93 ///< pulse pairs
94 };
95
96 /**
97 * Description of frame types.
98 */
99 static const struct frame_type_desc {
100 uint8_t n_blocks; ///< amount of blocks per frame (each block
101 ///< (contains 160/#n_blocks samples)
102 uint8_t log_n_blocks; ///< log2(#n_blocks)
103 uint8_t acb_type; ///< Adaptive codebook type (ACB_TYPE_*)
104 uint8_t fcb_type; ///< Fixed codebook type (FCB_TYPE_*)
105 uint8_t dbl_pulses; ///< how many pulse vectors have pulse pairs
106 ///< (rather than just one single pulse)
107 ///< only if #fcb_type == #FCB_TYPE_EXC_PULSES
108 } frame_descs[17] = {
109 { 1, 0, ACB_TYPE_NONE, FCB_TYPE_SILENCE, 0 },
110 { 2, 1, ACB_TYPE_NONE, FCB_TYPE_HARDCODED, 0 },
111 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_AW_PULSES, 0 },
112 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2 },
113 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5 },
114 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 0 },
115 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2 },
116 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5 },
117 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0 },
118 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2 },
119 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5 },
120 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0 },
121 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2 },
122 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5 },
123 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0 },
124 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2 },
125 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5 }
126 };
127
128 /**
129 * WMA Voice decoding context.
130 */
131 typedef struct WMAVoiceContext {
132 /**
133 * @name Global values specified in the stream header / extradata or used all over.
134 * @{
135 */
136 GetBitContext gb; ///< packet bitreader. During decoder init,
137 ///< it contains the extradata from the
138 ///< demuxer. During decoding, it contains
139 ///< packet data.
140 int8_t vbm_tree[25]; ///< converts VLC codes to frame type
141
142 int spillover_bitsize; ///< number of bits used to specify
143 ///< #spillover_nbits in the packet header
144 ///< = ceil(log2(ctx->block_align << 3))
145 int history_nsamples; ///< number of samples in history for signal
146 ///< prediction (through ACB)
147
148 /* postfilter specific values */
149 int do_apf; ///< whether to apply the averaged
150 ///< projection filter (APF)
151 int denoise_strength; ///< strength of denoising in Wiener filter
152 ///< [0-11]
153 int denoise_tilt_corr; ///< Whether to apply tilt correction to the
154 ///< Wiener filter coefficients (postfilter)
155 int dc_level; ///< Predicted amount of DC noise, based
156 ///< on which a DC removal filter is used
157
158 int lsps; ///< number of LSPs per frame [10 or 16]
159 int lsp_q_mode; ///< defines quantizer defaults [0, 1]
160 int lsp_def_mode; ///< defines different sets of LSP defaults
161 ///< [0, 1]
162
163 int min_pitch_val; ///< base value for pitch parsing code
164 int max_pitch_val; ///< max value + 1 for pitch parsing
165 int pitch_nbits; ///< number of bits used to specify the
166 ///< pitch value in the frame header
167 int block_pitch_nbits; ///< number of bits used to specify the
168 ///< first block's pitch value
169 int block_pitch_range; ///< range of the block pitch
170 int block_delta_pitch_nbits; ///< number of bits used to specify the
171 ///< delta pitch between this and the last
172 ///< block's pitch value, used in all but
173 ///< first block
174 int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is
175 ///< from -this to +this-1)
176 uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale
177 ///< conversion
178
179 /**
180 * @}
181 *
182 * @name Packet values specified in the packet header or related to a packet.
183 *
184 * A packet is considered to be a single unit of data provided to this
185 * decoder by the demuxer.
186 * @{
187 */
188 int spillover_nbits; ///< number of bits of the previous packet's
189 ///< last superframe preceding this
190 ///< packet's first full superframe (useful
191 ///< for re-synchronization also)
192 int has_residual_lsps; ///< if set, superframes contain one set of
193 ///< LSPs that cover all frames, encoded as
194 ///< independent and residual LSPs; if not
195 ///< set, each frame contains its own, fully
196 ///< independent, LSPs
197 int skip_bits_next; ///< number of bits to skip at the next call
198 ///< to #wmavoice_decode_packet() (since
199 ///< they're part of the previous superframe)
200
201 uint8_t sframe_cache[SFRAME_CACHE_MAXSIZE + AV_INPUT_BUFFER_PADDING_SIZE];
202 ///< cache for superframe data split over
203 ///< multiple packets
204 int sframe_cache_size; ///< set to >0 if we have data from an
205 ///< (incomplete) superframe from a previous
206 ///< packet that spilled over in the current
207 ///< packet; specifies the amount of bits in
208 ///< #sframe_cache
209 PutBitContext pb; ///< bitstream writer for #sframe_cache
210
211 /**
212 * @}
213 *
214 * @name Frame and superframe values
215 * Superframe and frame data - these can change from frame to frame,
216 * although some of them do in that case serve as a cache / history for
217 * the next frame or superframe.
218 * @{
219 */
220 double prev_lsps[MAX_LSPS]; ///< LSPs of the last frame of the previous
221 ///< superframe
222 int last_pitch_val; ///< pitch value of the previous frame
223 int last_acb_type; ///< frame type [0-2] of the previous frame
224 int pitch_diff_sh16; ///< ((cur_pitch_val - #last_pitch_val)
225 ///< << 16) / #MAX_FRAMESIZE
226 float silence_gain; ///< set for use in blocks if #ACB_TYPE_NONE
227
228 int aw_idx_is_ext; ///< whether the AW index was encoded in
229 ///< 8 bits (instead of 6)
230 int aw_pulse_range; ///< the range over which #aw_pulse_set1()
231 ///< can apply the pulse, relative to the
232 ///< value in aw_first_pulse_off. The exact
233 ///< position of the first AW-pulse is within
234 ///< [pulse_off, pulse_off + this], and
235 ///< depends on bitstream values; [16 or 24]
236 int aw_n_pulses[2]; ///< number of AW-pulses in each block; note
237 ///< that this number can be negative (in
238 ///< which case it basically means "zero")
239 int aw_first_pulse_off[2]; ///< index of first sample to which to
240 ///< apply AW-pulses, or -0xff if unset
241 int aw_next_pulse_off_cache; ///< the position (relative to start of the
242 ///< second block) at which pulses should
243 ///< start to be positioned, serves as a
244 ///< cache for pitch-adaptive window pulses
245 ///< between blocks
246
247 int frame_cntr; ///< current frame index [0 - 0xFFFE]; is
248 ///< only used for comfort noise in #pRNG()
249 int nb_superframes; ///< number of superframes in current packet
250 float gain_pred_err[6]; ///< cache for gain prediction
251 float excitation_history[MAX_SIGNAL_HISTORY];
252 ///< cache of the signal of previous
253 ///< superframes, used as a history for
254 ///< signal generation
255 float synth_history[MAX_LSPS]; ///< see #excitation_history
256 /**
257 * @}
258 *
259 * @name Postfilter values
260 *
261 * Variables used for postfilter implementation, mostly history for
262 * smoothing and so on, and context variables for FFT/iFFT.
263 * @{
264 */
265 RDFTContext rdft, irdft; ///< contexts for FFT-calculation in the
266 ///< postfilter (for denoise filter)
267 DCTContext dct, dst; ///< contexts for phase shift (in Hilbert
268 ///< transform, part of postfilter)
269 float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi]
270 ///< range
271 float postfilter_agc; ///< gain control memory, used in
272 ///< #adaptive_gain_control()
273 float dcf_mem[2]; ///< DC filter history
274 float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE];
275 ///< zero filter output (i.e. excitation)
276 ///< by postfilter
277 float denoise_filter_cache[MAX_FRAMESIZE];
278 int denoise_filter_cache_size; ///< samples in #denoise_filter_cache
279 DECLARE_ALIGNED(32, float, tilted_lpcs_pf)[0x80];
280 ///< aligned buffer for LPC tilting
281 DECLARE_ALIGNED(32, float, denoise_coeffs_pf)[0x80];
282 ///< aligned buffer for denoise coefficients
283 DECLARE_ALIGNED(32, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16];
284 ///< aligned buffer for postfilter speech
285 ///< synthesis
286 /**
287 * @}
288 */
289 } WMAVoiceContext;
290
291 /**
292 * Set up the variable bit mode (VBM) tree from container extradata.
293 * @param gb bit I/O context.
294 * The bit context (s->gb) should be loaded with byte 23-46 of the
295 * container extradata (i.e. the ones containing the VBM tree).
296 * @param vbm_tree pointer to array to which the decoded VBM tree will be
297 * written.
298 * @return 0 on success, <0 on error.
299 */
decode_vbmtree(GetBitContext * gb,int8_t vbm_tree[25])300 static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25])
301 {
302 int cntr[8] = { 0 }, n, res;
303
304 memset(vbm_tree, 0xff, sizeof(vbm_tree[0]) * 25);
305 for (n = 0; n < 17; n++) {
306 res = get_bits(gb, 3);
307 if (cntr[res] > 3) // should be >= 3 + (res == 7))
308 return -1;
309 vbm_tree[res * 3 + cntr[res]++] = n;
310 }
311 return 0;
312 }
313
wmavoice_init_static_data(void)314 static av_cold void wmavoice_init_static_data(void)
315 {
316 static const uint8_t bits[] = {
317 2, 2, 2, 4, 4, 4,
318 6, 6, 6, 8, 8, 8,
319 10, 10, 10, 12, 12, 12,
320 14, 14, 14, 14
321 };
322 static const uint16_t codes[] = {
323 0x0000, 0x0001, 0x0002, // 00/01/10
324 0x000c, 0x000d, 0x000e, // 11+00/01/10
325 0x003c, 0x003d, 0x003e, // 1111+00/01/10
326 0x00fc, 0x00fd, 0x00fe, // 111111+00/01/10
327 0x03fc, 0x03fd, 0x03fe, // 11111111+00/01/10
328 0x0ffc, 0x0ffd, 0x0ffe, // 1111111111+00/01/10
329 0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx
330 };
331
332 INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits),
333 bits, 1, 1, codes, 2, 2, 132);
334 }
335
wmavoice_flush(AVCodecContext * ctx)336 static av_cold void wmavoice_flush(AVCodecContext *ctx)
337 {
338 WMAVoiceContext *s = ctx->priv_data;
339 int n;
340
341 s->postfilter_agc = 0;
342 s->sframe_cache_size = 0;
343 s->skip_bits_next = 0;
344 for (n = 0; n < s->lsps; n++)
345 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
346 memset(s->excitation_history, 0,
347 sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY);
348 memset(s->synth_history, 0,
349 sizeof(*s->synth_history) * MAX_LSPS);
350 memset(s->gain_pred_err, 0,
351 sizeof(s->gain_pred_err));
352
353 if (s->do_apf) {
354 memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0,
355 sizeof(*s->synth_filter_out_buf) * s->lsps);
356 memset(s->dcf_mem, 0,
357 sizeof(*s->dcf_mem) * 2);
358 memset(s->zero_exc_pf, 0,
359 sizeof(*s->zero_exc_pf) * s->history_nsamples);
360 memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache));
361 }
362 }
363
364 /**
365 * Set up decoder with parameters from demuxer (extradata etc.).
366 */
wmavoice_decode_init(AVCodecContext * ctx)367 static av_cold int wmavoice_decode_init(AVCodecContext *ctx)
368 {
369 static AVOnce init_static_once = AV_ONCE_INIT;
370 int n, flags, pitch_range, lsp16_flag, ret;
371 WMAVoiceContext *s = ctx->priv_data;
372
373 ff_thread_once(&init_static_once, wmavoice_init_static_data);
374
375 /**
376 * Extradata layout:
377 * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c),
378 * - byte 19-22: flags field (annoyingly in LE; see below for known
379 * values),
380 * - byte 23-46: variable bitmode tree (really just 17 * 3 bits,
381 * rest is 0).
382 */
383 if (ctx->extradata_size != 46) {
384 av_log(ctx, AV_LOG_ERROR,
385 "Invalid extradata size %d (should be 46)\n",
386 ctx->extradata_size);
387 return AVERROR_INVALIDDATA;
388 }
389 if (ctx->block_align <= 0 || ctx->block_align > (1<<22)) {
390 av_log(ctx, AV_LOG_ERROR, "Invalid block alignment %d.\n", ctx->block_align);
391 return AVERROR_INVALIDDATA;
392 }
393
394 flags = AV_RL32(ctx->extradata + 18);
395 s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align);
396 s->do_apf = flags & 0x1;
397 if (s->do_apf) {
398 if ((ret = ff_rdft_init(&s->rdft, 7, DFT_R2C)) < 0 ||
399 (ret = ff_rdft_init(&s->irdft, 7, IDFT_C2R)) < 0 ||
400 (ret = ff_dct_init (&s->dct, 6, DCT_I)) < 0 ||
401 (ret = ff_dct_init (&s->dst, 6, DST_I)) < 0)
402 return ret;
403
404 ff_sine_window_init(s->cos, 256);
405 memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0]));
406 for (n = 0; n < 255; n++) {
407 s->sin[n] = -s->sin[510 - n];
408 s->cos[510 - n] = s->cos[n];
409 }
410 }
411 s->denoise_strength = (flags >> 2) & 0xF;
412 if (s->denoise_strength >= 12) {
413 av_log(ctx, AV_LOG_ERROR,
414 "Invalid denoise filter strength %d (max=11)\n",
415 s->denoise_strength);
416 return AVERROR_INVALIDDATA;
417 }
418 s->denoise_tilt_corr = !!(flags & 0x40);
419 s->dc_level = (flags >> 7) & 0xF;
420 s->lsp_q_mode = !!(flags & 0x2000);
421 s->lsp_def_mode = !!(flags & 0x4000);
422 lsp16_flag = flags & 0x1000;
423 if (lsp16_flag) {
424 s->lsps = 16;
425 } else {
426 s->lsps = 10;
427 }
428 for (n = 0; n < s->lsps; n++)
429 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
430
431 init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3);
432 if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) {
433 av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n");
434 return AVERROR_INVALIDDATA;
435 }
436
437 if (ctx->sample_rate >= INT_MAX / (256 * 37))
438 return AVERROR_INVALIDDATA;
439
440 s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8;
441 s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8;
442 pitch_range = s->max_pitch_val - s->min_pitch_val;
443 if (pitch_range <= 0) {
444 av_log(ctx, AV_LOG_ERROR, "Invalid pitch range; broken extradata?\n");
445 return AVERROR_INVALIDDATA;
446 }
447 s->pitch_nbits = av_ceil_log2(pitch_range);
448 s->last_pitch_val = 40;
449 s->last_acb_type = ACB_TYPE_NONE;
450 s->history_nsamples = s->max_pitch_val + 8;
451
452 if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) {
453 int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8,
454 max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8;
455
456 av_log(ctx, AV_LOG_ERROR,
457 "Unsupported samplerate %d (min=%d, max=%d)\n",
458 ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz
459
460 return AVERROR(ENOSYS);
461 }
462
463 s->block_conv_table[0] = s->min_pitch_val;
464 s->block_conv_table[1] = (pitch_range * 25) >> 6;
465 s->block_conv_table[2] = (pitch_range * 44) >> 6;
466 s->block_conv_table[3] = s->max_pitch_val - 1;
467 s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF;
468 if (s->block_delta_pitch_hrange <= 0) {
469 av_log(ctx, AV_LOG_ERROR, "Invalid delta pitch hrange; broken extradata?\n");
470 return AVERROR_INVALIDDATA;
471 }
472 s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange);
473 s->block_pitch_range = s->block_conv_table[2] +
474 s->block_conv_table[3] + 1 +
475 2 * (s->block_conv_table[1] - 2 * s->min_pitch_val);
476 s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range);
477
478 ctx->channels = 1;
479 ctx->channel_layout = AV_CH_LAYOUT_MONO;
480 ctx->sample_fmt = AV_SAMPLE_FMT_FLT;
481
482 return 0;
483 }
484
485 /**
486 * @name Postfilter functions
487 * Postfilter functions (gain control, wiener denoise filter, DC filter,
488 * kalman smoothening, plus surrounding code to wrap it)
489 * @{
490 */
491 /**
492 * Adaptive gain control (as used in postfilter).
493 *
494 * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except
495 * that the energy here is calculated using sum(abs(...)), whereas the
496 * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)).
497 *
498 * @param out output buffer for filtered samples
499 * @param in input buffer containing the samples as they are after the
500 * postfilter steps so far
501 * @param speech_synth input buffer containing speech synth before postfilter
502 * @param size input buffer size
503 * @param alpha exponential filter factor
504 * @param gain_mem pointer to filter memory (single float)
505 */
adaptive_gain_control(float * out,const float * in,const float * speech_synth,int size,float alpha,float * gain_mem)506 static void adaptive_gain_control(float *out, const float *in,
507 const float *speech_synth,
508 int size, float alpha, float *gain_mem)
509 {
510 int i;
511 float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor;
512 float mem = *gain_mem;
513
514 for (i = 0; i < size; i++) {
515 speech_energy += fabsf(speech_synth[i]);
516 postfilter_energy += fabsf(in[i]);
517 }
518 gain_scale_factor = postfilter_energy == 0.0 ? 0.0 :
519 (1.0 - alpha) * speech_energy / postfilter_energy;
520
521 for (i = 0; i < size; i++) {
522 mem = alpha * mem + gain_scale_factor;
523 out[i] = in[i] * mem;
524 }
525
526 *gain_mem = mem;
527 }
528
529 /**
530 * Kalman smoothing function.
531 *
532 * This function looks back pitch +/- 3 samples back into history to find
533 * the best fitting curve (that one giving the optimal gain of the two
534 * signals, i.e. the highest dot product between the two), and then
535 * uses that signal history to smoothen the output of the speech synthesis
536 * filter.
537 *
538 * @param s WMA Voice decoding context
539 * @param pitch pitch of the speech signal
540 * @param in input speech signal
541 * @param out output pointer for smoothened signal
542 * @param size input/output buffer size
543 *
544 * @returns -1 if no smoothening took place, e.g. because no optimal
545 * fit could be found, or 0 on success.
546 */
kalman_smoothen(WMAVoiceContext * s,int pitch,const float * in,float * out,int size)547 static int kalman_smoothen(WMAVoiceContext *s, int pitch,
548 const float *in, float *out, int size)
549 {
550 int n;
551 float optimal_gain = 0, dot;
552 const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)],
553 *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)],
554 *best_hist_ptr = NULL;
555
556 /* find best fitting point in history */
557 do {
558 dot = avpriv_scalarproduct_float_c(in, ptr, size);
559 if (dot > optimal_gain) {
560 optimal_gain = dot;
561 best_hist_ptr = ptr;
562 }
563 } while (--ptr >= end);
564
565 if (optimal_gain <= 0)
566 return -1;
567 dot = avpriv_scalarproduct_float_c(best_hist_ptr, best_hist_ptr, size);
568 if (dot <= 0) // would be 1.0
569 return -1;
570
571 if (optimal_gain <= dot) {
572 dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000
573 } else
574 dot = 0.625;
575
576 /* actual smoothing */
577 for (n = 0; n < size; n++)
578 out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]);
579
580 return 0;
581 }
582
583 /**
584 * Get the tilt factor of a formant filter from its transfer function
585 * @see #tilt_factor() in amrnbdec.c, which does essentially the same,
586 * but somehow (??) it does a speech synthesis filter in the
587 * middle, which is missing here
588 *
589 * @param lpcs LPC coefficients
590 * @param n_lpcs Size of LPC buffer
591 * @returns the tilt factor
592 */
tilt_factor(const float * lpcs,int n_lpcs)593 static float tilt_factor(const float *lpcs, int n_lpcs)
594 {
595 float rh0, rh1;
596
597 rh0 = 1.0 + avpriv_scalarproduct_float_c(lpcs, lpcs, n_lpcs);
598 rh1 = lpcs[0] + avpriv_scalarproduct_float_c(lpcs, &lpcs[1], n_lpcs - 1);
599
600 return rh1 / rh0;
601 }
602
603 /**
604 * Derive denoise filter coefficients (in real domain) from the LPCs.
605 */
calc_input_response(WMAVoiceContext * s,float * lpcs,int fcb_type,float * coeffs,int remainder)606 static void calc_input_response(WMAVoiceContext *s, float *lpcs,
607 int fcb_type, float *coeffs, int remainder)
608 {
609 float last_coeff, min = 15.0, max = -15.0;
610 float irange, angle_mul, gain_mul, range, sq;
611 int n, idx;
612
613 /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */
614 s->rdft.rdft_calc(&s->rdft, lpcs);
615 #define log_range(var, assign) do { \
616 float tmp = log10f(assign); var = tmp; \
617 max = FFMAX(max, tmp); min = FFMIN(min, tmp); \
618 } while (0)
619 log_range(last_coeff, lpcs[1] * lpcs[1]);
620 for (n = 1; n < 64; n++)
621 log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] +
622 lpcs[n * 2 + 1] * lpcs[n * 2 + 1]);
623 log_range(lpcs[0], lpcs[0] * lpcs[0]);
624 #undef log_range
625 range = max - min;
626 lpcs[64] = last_coeff;
627
628 /* Now, use this spectrum to pick out these frequencies with higher
629 * (relative) power/energy (which we then take to be "not noise"),
630 * and set up a table (still in lpc[]) of (relative) gains per frequency.
631 * These frequencies will be maintained, while others ("noise") will be
632 * decreased in the filter output. */
633 irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63]
634 gain_mul = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) :
635 (5.0 / 14.7));
636 angle_mul = gain_mul * (8.0 * M_LN10 / M_PI);
637 for (n = 0; n <= 64; n++) {
638 float pwr;
639
640 idx = lrint((max - lpcs[n]) * irange - 1);
641 idx = FFMAX(0, idx);
642 pwr = wmavoice_denoise_power_table[s->denoise_strength][idx];
643 lpcs[n] = angle_mul * pwr;
644
645 /* 70.57 =~ 1/log10(1.0331663) */
646 idx = av_clipf((pwr * gain_mul - 0.0295) * 70.570526123, 0, INT_MAX / 2);
647
648 if (idx > 127) { // fall back if index falls outside table range
649 coeffs[n] = wmavoice_energy_table[127] *
650 powf(1.0331663, idx - 127);
651 } else
652 coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)];
653 }
654
655 /* calculate the Hilbert transform of the gains, which we do (since this
656 * is a sine input) by doing a phase shift (in theory, H(sin())=cos()).
657 * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the
658 * "moment" of the LPCs in this filter. */
659 s->dct.dct_calc(&s->dct, lpcs);
660 s->dst.dct_calc(&s->dst, lpcs);
661
662 /* Split out the coefficient indexes into phase/magnitude pairs */
663 idx = 255 + av_clip(lpcs[64], -255, 255);
664 coeffs[0] = coeffs[0] * s->cos[idx];
665 idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255);
666 last_coeff = coeffs[64] * s->cos[idx];
667 for (n = 63;; n--) {
668 idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255);
669 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
670 coeffs[n * 2] = coeffs[n] * s->cos[idx];
671
672 if (!--n) break;
673
674 idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255);
675 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
676 coeffs[n * 2] = coeffs[n] * s->cos[idx];
677 }
678 coeffs[1] = last_coeff;
679
680 /* move into real domain */
681 s->irdft.rdft_calc(&s->irdft, coeffs);
682
683 /* tilt correction and normalize scale */
684 memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder));
685 if (s->denoise_tilt_corr) {
686 float tilt_mem = 0;
687
688 coeffs[remainder - 1] = 0;
689 ff_tilt_compensation(&tilt_mem,
690 -1.8 * tilt_factor(coeffs, remainder - 1),
691 coeffs, remainder);
692 }
693 sq = (1.0 / 64.0) * sqrtf(1 / avpriv_scalarproduct_float_c(coeffs, coeffs,
694 remainder));
695 for (n = 0; n < remainder; n++)
696 coeffs[n] *= sq;
697 }
698
699 /**
700 * This function applies a Wiener filter on the (noisy) speech signal as
701 * a means to denoise it.
702 *
703 * - take RDFT of LPCs to get the power spectrum of the noise + speech;
704 * - using this power spectrum, calculate (for each frequency) the Wiener
705 * filter gain, which depends on the frequency power and desired level
706 * of noise subtraction (when set too high, this leads to artifacts)
707 * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse
708 * of 4-8kHz);
709 * - by doing a phase shift, calculate the Hilbert transform of this array
710 * of per-frequency filter-gains to get the filtering coefficients;
711 * - smoothen/normalize/de-tilt these filter coefficients as desired;
712 * - take RDFT of noisy sound, apply the coefficients and take its IRDFT
713 * to get the denoised speech signal;
714 * - the leftover (i.e. output of the IRDFT on denoised speech data beyond
715 * the frame boundary) are saved and applied to subsequent frames by an
716 * overlap-add method (otherwise you get clicking-artifacts).
717 *
718 * @param s WMA Voice decoding context
719 * @param fcb_type Frame (codebook) type
720 * @param synth_pf input: the noisy speech signal, output: denoised speech
721 * data; should be 16-byte aligned (for ASM purposes)
722 * @param size size of the speech data
723 * @param lpcs LPCs used to synthesize this frame's speech data
724 */
wiener_denoise(WMAVoiceContext * s,int fcb_type,float * synth_pf,int size,const float * lpcs)725 static void wiener_denoise(WMAVoiceContext *s, int fcb_type,
726 float *synth_pf, int size,
727 const float *lpcs)
728 {
729 int remainder, lim, n;
730
731 if (fcb_type != FCB_TYPE_SILENCE) {
732 float *tilted_lpcs = s->tilted_lpcs_pf,
733 *coeffs = s->denoise_coeffs_pf, tilt_mem = 0;
734
735 tilted_lpcs[0] = 1.0;
736 memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps);
737 memset(&tilted_lpcs[s->lsps + 1], 0,
738 sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1));
739 ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps),
740 tilted_lpcs, s->lsps + 2);
741
742 /* The IRDFT output (127 samples for 7-bit filter) beyond the frame
743 * size is applied to the next frame. All input beyond this is zero,
744 * and thus all output beyond this will go towards zero, hence we can
745 * limit to min(size-1, 127-size) as a performance consideration. */
746 remainder = FFMIN(127 - size, size - 1);
747 calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder);
748
749 /* apply coefficients (in frequency spectrum domain), i.e. complex
750 * number multiplication */
751 memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size));
752 s->rdft.rdft_calc(&s->rdft, synth_pf);
753 s->rdft.rdft_calc(&s->rdft, coeffs);
754 synth_pf[0] *= coeffs[0];
755 synth_pf[1] *= coeffs[1];
756 for (n = 1; n < 64; n++) {
757 float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1];
758 synth_pf[n * 2] = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1];
759 synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1];
760 }
761 s->irdft.rdft_calc(&s->irdft, synth_pf);
762 }
763
764 /* merge filter output with the history of previous runs */
765 if (s->denoise_filter_cache_size) {
766 lim = FFMIN(s->denoise_filter_cache_size, size);
767 for (n = 0; n < lim; n++)
768 synth_pf[n] += s->denoise_filter_cache[n];
769 s->denoise_filter_cache_size -= lim;
770 memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size],
771 sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size);
772 }
773
774 /* move remainder of filter output into a cache for future runs */
775 if (fcb_type != FCB_TYPE_SILENCE) {
776 lim = FFMIN(remainder, s->denoise_filter_cache_size);
777 for (n = 0; n < lim; n++)
778 s->denoise_filter_cache[n] += synth_pf[size + n];
779 if (lim < remainder) {
780 memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim],
781 sizeof(s->denoise_filter_cache[0]) * (remainder - lim));
782 s->denoise_filter_cache_size = remainder;
783 }
784 }
785 }
786
787 /**
788 * Averaging projection filter, the postfilter used in WMAVoice.
789 *
790 * This uses the following steps:
791 * - A zero-synthesis filter (generate excitation from synth signal)
792 * - Kalman smoothing on excitation, based on pitch
793 * - Re-synthesized smoothened output
794 * - Iterative Wiener denoise filter
795 * - Adaptive gain filter
796 * - DC filter
797 *
798 * @param s WMAVoice decoding context
799 * @param synth Speech synthesis output (before postfilter)
800 * @param samples Output buffer for filtered samples
801 * @param size Buffer size of synth & samples
802 * @param lpcs Generated LPCs used for speech synthesis
803 * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned)
804 * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses)
805 * @param pitch Pitch of the input signal
806 */
postfilter(WMAVoiceContext * s,const float * synth,float * samples,int size,const float * lpcs,float * zero_exc_pf,int fcb_type,int pitch)807 static void postfilter(WMAVoiceContext *s, const float *synth,
808 float *samples, int size,
809 const float *lpcs, float *zero_exc_pf,
810 int fcb_type, int pitch)
811 {
812 float synth_filter_in_buf[MAX_FRAMESIZE / 2],
813 *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16],
814 *synth_filter_in = zero_exc_pf;
815
816 av_assert0(size <= MAX_FRAMESIZE / 2);
817
818 /* generate excitation from input signal */
819 ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps);
820
821 if (fcb_type >= FCB_TYPE_AW_PULSES &&
822 !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size))
823 synth_filter_in = synth_filter_in_buf;
824
825 /* re-synthesize speech after smoothening, and keep history */
826 ff_celp_lp_synthesis_filterf(synth_pf, lpcs,
827 synth_filter_in, size, s->lsps);
828 memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps],
829 sizeof(synth_pf[0]) * s->lsps);
830
831 wiener_denoise(s, fcb_type, synth_pf, size, lpcs);
832
833 adaptive_gain_control(samples, synth_pf, synth, size, 0.99,
834 &s->postfilter_agc);
835
836 if (s->dc_level > 8) {
837 /* remove ultra-low frequency DC noise / highpass filter;
838 * coefficients are identical to those used in SIPR decoding,
839 * and very closely resemble those used in AMR-NB decoding. */
840 ff_acelp_apply_order_2_transfer_function(samples, samples,
841 (const float[2]) { -1.99997, 1.0 },
842 (const float[2]) { -1.9330735188, 0.93589198496 },
843 0.93980580475, s->dcf_mem, size);
844 }
845 }
846 /**
847 * @}
848 */
849
850 /**
851 * Dequantize LSPs
852 * @param lsps output pointer to the array that will hold the LSPs
853 * @param num number of LSPs to be dequantized
854 * @param values quantized values, contains n_stages values
855 * @param sizes range (i.e. max value) of each quantized value
856 * @param n_stages number of dequantization runs
857 * @param table dequantization table to be used
858 * @param mul_q LSF multiplier
859 * @param base_q base (lowest) LSF values
860 */
dequant_lsps(double * lsps,int num,const uint16_t * values,const uint16_t * sizes,int n_stages,const uint8_t * table,const double * mul_q,const double * base_q)861 static void dequant_lsps(double *lsps, int num,
862 const uint16_t *values,
863 const uint16_t *sizes,
864 int n_stages, const uint8_t *table,
865 const double *mul_q,
866 const double *base_q)
867 {
868 int n, m;
869
870 memset(lsps, 0, num * sizeof(*lsps));
871 for (n = 0; n < n_stages; n++) {
872 const uint8_t *t_off = &table[values[n] * num];
873 double base = base_q[n], mul = mul_q[n];
874
875 for (m = 0; m < num; m++)
876 lsps[m] += base + mul * t_off[m];
877
878 table += sizes[n] * num;
879 }
880 }
881
882 /**
883 * @name LSP dequantization routines
884 * LSP dequantization routines, for 10/16LSPs and independent/residual coding.
885 * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;
886 * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits.
887 * @{
888 */
889 /**
890 * Parse 10 independently-coded LSPs.
891 */
dequant_lsp10i(GetBitContext * gb,double * lsps)892 static void dequant_lsp10i(GetBitContext *gb, double *lsps)
893 {
894 static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 };
895 static const double mul_lsf[4] = {
896 5.2187144800e-3, 1.4626986422e-3,
897 9.6179549166e-4, 1.1325736225e-3
898 };
899 static const double base_lsf[4] = {
900 M_PI * -2.15522e-1, M_PI * -6.1646e-2,
901 M_PI * -3.3486e-2, M_PI * -5.7408e-2
902 };
903 uint16_t v[4];
904
905 v[0] = get_bits(gb, 8);
906 v[1] = get_bits(gb, 6);
907 v[2] = get_bits(gb, 5);
908 v[3] = get_bits(gb, 5);
909
910 dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i,
911 mul_lsf, base_lsf);
912 }
913
914 /**
915 * Parse 10 independently-coded LSPs, and then derive the tables to
916 * generate LSPs for the other frames from them (residual coding).
917 */
dequant_lsp10r(GetBitContext * gb,double * i_lsps,const double * old,double * a1,double * a2,int q_mode)918 static void dequant_lsp10r(GetBitContext *gb,
919 double *i_lsps, const double *old,
920 double *a1, double *a2, int q_mode)
921 {
922 static const uint16_t vec_sizes[3] = { 128, 64, 64 };
923 static const double mul_lsf[3] = {
924 2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3
925 };
926 static const double base_lsf[3] = {
927 M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2
928 };
929 const float (*ipol_tab)[2][10] = q_mode ?
930 wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a;
931 uint16_t interpol, v[3];
932 int n;
933
934 dequant_lsp10i(gb, i_lsps);
935
936 interpol = get_bits(gb, 5);
937 v[0] = get_bits(gb, 7);
938 v[1] = get_bits(gb, 6);
939 v[2] = get_bits(gb, 6);
940
941 for (n = 0; n < 10; n++) {
942 double delta = old[n] - i_lsps[n];
943 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
944 a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
945 }
946
947 dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r,
948 mul_lsf, base_lsf);
949 }
950
951 /**
952 * Parse 16 independently-coded LSPs.
953 */
dequant_lsp16i(GetBitContext * gb,double * lsps)954 static void dequant_lsp16i(GetBitContext *gb, double *lsps)
955 {
956 static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 };
957 static const double mul_lsf[5] = {
958 3.3439586280e-3, 6.9908173703e-4,
959 3.3216608306e-3, 1.0334960326e-3,
960 3.1899104283e-3
961 };
962 static const double base_lsf[5] = {
963 M_PI * -1.27576e-1, M_PI * -2.4292e-2,
964 M_PI * -1.28094e-1, M_PI * -3.2128e-2,
965 M_PI * -1.29816e-1
966 };
967 uint16_t v[5];
968
969 v[0] = get_bits(gb, 8);
970 v[1] = get_bits(gb, 6);
971 v[2] = get_bits(gb, 7);
972 v[3] = get_bits(gb, 6);
973 v[4] = get_bits(gb, 7);
974
975 dequant_lsps( lsps, 5, v, vec_sizes, 2,
976 wmavoice_dq_lsp16i1, mul_lsf, base_lsf);
977 dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2,
978 wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]);
979 dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1,
980 wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]);
981 }
982
983 /**
984 * Parse 16 independently-coded LSPs, and then derive the tables to
985 * generate LSPs for the other frames from them (residual coding).
986 */
dequant_lsp16r(GetBitContext * gb,double * i_lsps,const double * old,double * a1,double * a2,int q_mode)987 static void dequant_lsp16r(GetBitContext *gb,
988 double *i_lsps, const double *old,
989 double *a1, double *a2, int q_mode)
990 {
991 static const uint16_t vec_sizes[3] = { 128, 128, 128 };
992 static const double mul_lsf[3] = {
993 1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3
994 };
995 static const double base_lsf[3] = {
996 M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2
997 };
998 const float (*ipol_tab)[2][16] = q_mode ?
999 wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a;
1000 uint16_t interpol, v[3];
1001 int n;
1002
1003 dequant_lsp16i(gb, i_lsps);
1004
1005 interpol = get_bits(gb, 5);
1006 v[0] = get_bits(gb, 7);
1007 v[1] = get_bits(gb, 7);
1008 v[2] = get_bits(gb, 7);
1009
1010 for (n = 0; n < 16; n++) {
1011 double delta = old[n] - i_lsps[n];
1012 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
1013 a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
1014 }
1015
1016 dequant_lsps( a2, 10, v, vec_sizes, 1,
1017 wmavoice_dq_lsp16r1, mul_lsf, base_lsf);
1018 dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1,
1019 wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]);
1020 dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1,
1021 wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]);
1022 }
1023
1024 /**
1025 * @}
1026 * @name Pitch-adaptive window coding functions
1027 * The next few functions are for pitch-adaptive window coding.
1028 * @{
1029 */
1030 /**
1031 * Parse the offset of the first pitch-adaptive window pulses, and
1032 * the distribution of pulses between the two blocks in this frame.
1033 * @param s WMA Voice decoding context private data
1034 * @param gb bit I/O context
1035 * @param pitch pitch for each block in this frame
1036 */
aw_parse_coords(WMAVoiceContext * s,GetBitContext * gb,const int * pitch)1037 static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb,
1038 const int *pitch)
1039 {
1040 static const int16_t start_offset[94] = {
1041 -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11,
1042 13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26,
1043 27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43,
1044 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67,
1045 69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91,
1046 93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115,
1047 117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139,
1048 141, 143, 145, 147, 149, 151, 153, 155, 157, 159
1049 };
1050 int bits, offset;
1051
1052 /* position of pulse */
1053 s->aw_idx_is_ext = 0;
1054 if ((bits = get_bits(gb, 6)) >= 54) {
1055 s->aw_idx_is_ext = 1;
1056 bits += (bits - 54) * 3 + get_bits(gb, 2);
1057 }
1058
1059 /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count
1060 * the distribution of the pulses in each block contained in this frame. */
1061 s->aw_pulse_range = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16;
1062 for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ;
1063 s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0];
1064 s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2;
1065 offset += s->aw_n_pulses[0] * pitch[0];
1066 s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1];
1067 s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2;
1068
1069 /* if continuing from a position before the block, reset position to
1070 * start of block (when corrected for the range over which it can be
1071 * spread in aw_pulse_set1()). */
1072 if (start_offset[bits] < MAX_FRAMESIZE / 2) {
1073 while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0)
1074 s->aw_first_pulse_off[1] -= pitch[1];
1075 if (start_offset[bits] < 0)
1076 while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0)
1077 s->aw_first_pulse_off[0] -= pitch[0];
1078 }
1079 }
1080
1081 /**
1082 * Apply second set of pitch-adaptive window pulses.
1083 * @param s WMA Voice decoding context private data
1084 * @param gb bit I/O context
1085 * @param block_idx block index in frame [0, 1]
1086 * @param fcb structure containing fixed codebook vector info
1087 * @return -1 on error, 0 otherwise
1088 */
aw_pulse_set2(WMAVoiceContext * s,GetBitContext * gb,int block_idx,AMRFixed * fcb)1089 static int aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb,
1090 int block_idx, AMRFixed *fcb)
1091 {
1092 uint16_t use_mask_mem[9]; // only 5 are used, rest is padding
1093 uint16_t *use_mask = use_mask_mem + 2;
1094 /* in this function, idx is the index in the 80-bit (+ padding) use_mask
1095 * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits
1096 * of idx are the position of the bit within a particular item in the
1097 * array (0 being the most significant bit, and 15 being the least
1098 * significant bit), and the remainder (>> 4) is the index in the
1099 * use_mask[]-array. This is faster and uses less memory than using a
1100 * 80-byte/80-int array. */
1101 int pulse_off = s->aw_first_pulse_off[block_idx],
1102 pulse_start, n, idx, range, aidx, start_off = 0;
1103
1104 /* set offset of first pulse to within this block */
1105 if (s->aw_n_pulses[block_idx] > 0)
1106 while (pulse_off + s->aw_pulse_range < 1)
1107 pulse_off += fcb->pitch_lag;
1108
1109 /* find range per pulse */
1110 if (s->aw_n_pulses[0] > 0) {
1111 if (block_idx == 0) {
1112 range = 32;
1113 } else /* block_idx = 1 */ {
1114 range = 8;
1115 if (s->aw_n_pulses[block_idx] > 0)
1116 pulse_off = s->aw_next_pulse_off_cache;
1117 }
1118 } else
1119 range = 16;
1120 pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0;
1121
1122 /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,
1123 * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus
1124 * we exclude that range from being pulsed again in this function. */
1125 memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0]));
1126 memset( use_mask, -1, 5 * sizeof(use_mask[0]));
1127 memset(&use_mask[5], 0, 2 * sizeof(use_mask[0]));
1128 if (s->aw_n_pulses[block_idx] > 0)
1129 for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) {
1130 int excl_range = s->aw_pulse_range; // always 16 or 24
1131 uint16_t *use_mask_ptr = &use_mask[idx >> 4];
1132 int first_sh = 16 - (idx & 15);
1133 *use_mask_ptr++ &= 0xFFFFu << first_sh;
1134 excl_range -= first_sh;
1135 if (excl_range >= 16) {
1136 *use_mask_ptr++ = 0;
1137 *use_mask_ptr &= 0xFFFF >> (excl_range - 16);
1138 } else
1139 *use_mask_ptr &= 0xFFFF >> excl_range;
1140 }
1141
1142 /* find the 'aidx'th offset that is not excluded */
1143 aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4);
1144 for (n = 0; n <= aidx; pulse_start++) {
1145 for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ;
1146 if (idx >= MAX_FRAMESIZE / 2) { // find from zero
1147 if (use_mask[0]) idx = 0x0F;
1148 else if (use_mask[1]) idx = 0x1F;
1149 else if (use_mask[2]) idx = 0x2F;
1150 else if (use_mask[3]) idx = 0x3F;
1151 else if (use_mask[4]) idx = 0x4F;
1152 else return -1;
1153 idx -= av_log2_16bit(use_mask[idx >> 4]);
1154 }
1155 if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) {
1156 use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15));
1157 n++;
1158 start_off = idx;
1159 }
1160 }
1161
1162 fcb->x[fcb->n] = start_off;
1163 fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0;
1164 fcb->n++;
1165
1166 /* set offset for next block, relative to start of that block */
1167 n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag;
1168 s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0;
1169 return 0;
1170 }
1171
1172 /**
1173 * Apply first set of pitch-adaptive window pulses.
1174 * @param s WMA Voice decoding context private data
1175 * @param gb bit I/O context
1176 * @param block_idx block index in frame [0, 1]
1177 * @param fcb storage location for fixed codebook pulse info
1178 */
aw_pulse_set1(WMAVoiceContext * s,GetBitContext * gb,int block_idx,AMRFixed * fcb)1179 static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb,
1180 int block_idx, AMRFixed *fcb)
1181 {
1182 int val = get_bits(gb, 12 - 2 * (s->aw_idx_is_ext && !block_idx));
1183 float v;
1184
1185 if (s->aw_n_pulses[block_idx] > 0) {
1186 int n, v_mask, i_mask, sh, n_pulses;
1187
1188 if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each
1189 n_pulses = 3;
1190 v_mask = 8;
1191 i_mask = 7;
1192 sh = 4;
1193 } else { // 4 pulses, 1:sign + 2:index each
1194 n_pulses = 4;
1195 v_mask = 4;
1196 i_mask = 3;
1197 sh = 3;
1198 }
1199
1200 for (n = n_pulses - 1; n >= 0; n--, val >>= sh) {
1201 fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0;
1202 fcb->x[fcb->n] = (val & i_mask) * n_pulses + n +
1203 s->aw_first_pulse_off[block_idx];
1204 while (fcb->x[fcb->n] < 0)
1205 fcb->x[fcb->n] += fcb->pitch_lag;
1206 if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2)
1207 fcb->n++;
1208 }
1209 } else {
1210 int num2 = (val & 0x1FF) >> 1, delta, idx;
1211
1212 if (num2 < 1 * 79) { delta = 1; idx = num2 + 1; }
1213 else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; }
1214 else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; }
1215 else { delta = 7; idx = num2 + 1 - 3 * 75; }
1216 v = (val & 0x200) ? -1.0 : 1.0;
1217
1218 fcb->no_repeat_mask |= 3 << fcb->n;
1219 fcb->x[fcb->n] = idx - delta;
1220 fcb->y[fcb->n] = v;
1221 fcb->x[fcb->n + 1] = idx;
1222 fcb->y[fcb->n + 1] = (val & 1) ? -v : v;
1223 fcb->n += 2;
1224 }
1225 }
1226
1227 /**
1228 * @}
1229 *
1230 * Generate a random number from frame_cntr and block_idx, which will live
1231 * in the range [0, 1000 - block_size] (so it can be used as an index in a
1232 * table of size 1000 of which you want to read block_size entries).
1233 *
1234 * @param frame_cntr current frame number
1235 * @param block_num current block index
1236 * @param block_size amount of entries we want to read from a table
1237 * that has 1000 entries
1238 * @return a (non-)random number in the [0, 1000 - block_size] range.
1239 */
pRNG(int frame_cntr,int block_num,int block_size)1240 static int pRNG(int frame_cntr, int block_num, int block_size)
1241 {
1242 /* array to simplify the calculation of z:
1243 * y = (x % 9) * 5 + 6;
1244 * z = (49995 * x) / y;
1245 * Since y only has 9 values, we can remove the division by using a
1246 * LUT and using FASTDIV-style divisions. For each of the 9 values
1247 * of y, we can rewrite z as:
1248 * z = x * (49995 / y) + x * ((49995 % y) / y)
1249 * In this table, each col represents one possible value of y, the
1250 * first number is 49995 / y, and the second is the FASTDIV variant
1251 * of 49995 % y / y. */
1252 static const unsigned int div_tbl[9][2] = {
1253 { 8332, 3 * 715827883U }, // y = 6
1254 { 4545, 0 * 390451573U }, // y = 11
1255 { 3124, 11 * 268435456U }, // y = 16
1256 { 2380, 15 * 204522253U }, // y = 21
1257 { 1922, 23 * 165191050U }, // y = 26
1258 { 1612, 23 * 138547333U }, // y = 31
1259 { 1388, 27 * 119304648U }, // y = 36
1260 { 1219, 16 * 104755300U }, // y = 41
1261 { 1086, 39 * 93368855U } // y = 46
1262 };
1263 unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr;
1264 if (x >= 0xFFFF) x -= 0xFFFF; // max value of x is 8*1877+0xFFFE=0x13AA6,
1265 // so this is effectively a modulo (%)
1266 y = x - 9 * MULH(477218589, x); // x % 9
1267 z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1]));
1268 // z = x * 49995 / (y * 5 + 6)
1269 return z % (1000 - block_size);
1270 }
1271
1272 /**
1273 * Parse hardcoded signal for a single block.
1274 * @note see #synth_block().
1275 */
synth_block_hardcoded(WMAVoiceContext * s,GetBitContext * gb,int block_idx,int size,const struct frame_type_desc * frame_desc,float * excitation)1276 static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb,
1277 int block_idx, int size,
1278 const struct frame_type_desc *frame_desc,
1279 float *excitation)
1280 {
1281 float gain;
1282 int n, r_idx;
1283
1284 av_assert0(size <= MAX_FRAMESIZE);
1285
1286 /* Set the offset from which we start reading wmavoice_std_codebook */
1287 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1288 r_idx = pRNG(s->frame_cntr, block_idx, size);
1289 gain = s->silence_gain;
1290 } else /* FCB_TYPE_HARDCODED */ {
1291 r_idx = get_bits(gb, 8);
1292 gain = wmavoice_gain_universal[get_bits(gb, 6)];
1293 }
1294
1295 /* Clear gain prediction parameters */
1296 memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err));
1297
1298 /* Apply gain to hardcoded codebook and use that as excitation signal */
1299 for (n = 0; n < size; n++)
1300 excitation[n] = wmavoice_std_codebook[r_idx + n] * gain;
1301 }
1302
1303 /**
1304 * Parse FCB/ACB signal for a single block.
1305 * @note see #synth_block().
1306 */
synth_block_fcb_acb(WMAVoiceContext * s,GetBitContext * gb,int block_idx,int size,int block_pitch_sh2,const struct frame_type_desc * frame_desc,float * excitation)1307 static void synth_block_fcb_acb(WMAVoiceContext *s, GetBitContext *gb,
1308 int block_idx, int size,
1309 int block_pitch_sh2,
1310 const struct frame_type_desc *frame_desc,
1311 float *excitation)
1312 {
1313 static const float gain_coeff[6] = {
1314 0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458
1315 };
1316 float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain;
1317 int n, idx, gain_weight;
1318 AMRFixed fcb;
1319
1320 av_assert0(size <= MAX_FRAMESIZE / 2);
1321 memset(pulses, 0, sizeof(*pulses) * size);
1322
1323 fcb.pitch_lag = block_pitch_sh2 >> 2;
1324 fcb.pitch_fac = 1.0;
1325 fcb.no_repeat_mask = 0;
1326 fcb.n = 0;
1327
1328 /* For the other frame types, this is where we apply the innovation
1329 * (fixed) codebook pulses of the speech signal. */
1330 if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1331 aw_pulse_set1(s, gb, block_idx, &fcb);
1332 if (aw_pulse_set2(s, gb, block_idx, &fcb)) {
1333 /* Conceal the block with silence and return.
1334 * Skip the correct amount of bits to read the next
1335 * block from the correct offset. */
1336 int r_idx = pRNG(s->frame_cntr, block_idx, size);
1337
1338 for (n = 0; n < size; n++)
1339 excitation[n] =
1340 wmavoice_std_codebook[r_idx + n] * s->silence_gain;
1341 skip_bits(gb, 7 + 1);
1342 return;
1343 }
1344 } else /* FCB_TYPE_EXC_PULSES */ {
1345 int offset_nbits = 5 - frame_desc->log_n_blocks;
1346
1347 fcb.no_repeat_mask = -1;
1348 /* similar to ff_decode_10_pulses_35bits(), but with single pulses
1349 * (instead of double) for a subset of pulses */
1350 for (n = 0; n < 5; n++) {
1351 float sign;
1352 int pos1, pos2;
1353
1354 sign = get_bits1(gb) ? 1.0 : -1.0;
1355 pos1 = get_bits(gb, offset_nbits);
1356 fcb.x[fcb.n] = n + 5 * pos1;
1357 fcb.y[fcb.n++] = sign;
1358 if (n < frame_desc->dbl_pulses) {
1359 pos2 = get_bits(gb, offset_nbits);
1360 fcb.x[fcb.n] = n + 5 * pos2;
1361 fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign;
1362 }
1363 }
1364 }
1365 ff_set_fixed_vector(pulses, &fcb, 1.0, size);
1366
1367 /* Calculate gain for adaptive & fixed codebook signal.
1368 * see ff_amr_set_fixed_gain(). */
1369 idx = get_bits(gb, 7);
1370 fcb_gain = expf(avpriv_scalarproduct_float_c(s->gain_pred_err,
1371 gain_coeff, 6) -
1372 5.2409161640 + wmavoice_gain_codebook_fcb[idx]);
1373 acb_gain = wmavoice_gain_codebook_acb[idx];
1374 pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx],
1375 -2.9957322736 /* log(0.05) */,
1376 1.6094379124 /* log(5.0) */);
1377
1378 gain_weight = 8 >> frame_desc->log_n_blocks;
1379 memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err,
1380 sizeof(*s->gain_pred_err) * (6 - gain_weight));
1381 for (n = 0; n < gain_weight; n++)
1382 s->gain_pred_err[n] = pred_err;
1383
1384 /* Calculation of adaptive codebook */
1385 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1386 int len;
1387 for (n = 0; n < size; n += len) {
1388 int next_idx_sh16;
1389 int abs_idx = block_idx * size + n;
1390 int pitch_sh16 = (s->last_pitch_val << 16) +
1391 s->pitch_diff_sh16 * abs_idx;
1392 int pitch = (pitch_sh16 + 0x6FFF) >> 16;
1393 int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000;
1394 idx = idx_sh16 >> 16;
1395 if (s->pitch_diff_sh16) {
1396 if (s->pitch_diff_sh16 > 0) {
1397 next_idx_sh16 = (idx_sh16) &~ 0xFFFF;
1398 } else
1399 next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF;
1400 len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8,
1401 1, size - n);
1402 } else
1403 len = size;
1404
1405 ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch],
1406 wmavoice_ipol1_coeffs, 17,
1407 idx, 9, len);
1408 }
1409 } else /* ACB_TYPE_HAMMING */ {
1410 int block_pitch = block_pitch_sh2 >> 2;
1411 idx = block_pitch_sh2 & 3;
1412 if (idx) {
1413 ff_acelp_interpolatef(excitation, &excitation[-block_pitch],
1414 wmavoice_ipol2_coeffs, 4,
1415 idx, 8, size);
1416 } else
1417 av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch,
1418 sizeof(float) * size);
1419 }
1420
1421 /* Interpolate ACB/FCB and use as excitation signal */
1422 ff_weighted_vector_sumf(excitation, excitation, pulses,
1423 acb_gain, fcb_gain, size);
1424 }
1425
1426 /**
1427 * Parse data in a single block.
1428 *
1429 * @param s WMA Voice decoding context private data
1430 * @param gb bit I/O context
1431 * @param block_idx index of the to-be-read block
1432 * @param size amount of samples to be read in this block
1433 * @param block_pitch_sh2 pitch for this block << 2
1434 * @param lsps LSPs for (the end of) this frame
1435 * @param prev_lsps LSPs for the last frame
1436 * @param frame_desc frame type descriptor
1437 * @param excitation target memory for the ACB+FCB interpolated signal
1438 * @param synth target memory for the speech synthesis filter output
1439 * @return 0 on success, <0 on error.
1440 */
synth_block(WMAVoiceContext * s,GetBitContext * gb,int block_idx,int size,int block_pitch_sh2,const double * lsps,const double * prev_lsps,const struct frame_type_desc * frame_desc,float * excitation,float * synth)1441 static void synth_block(WMAVoiceContext *s, GetBitContext *gb,
1442 int block_idx, int size,
1443 int block_pitch_sh2,
1444 const double *lsps, const double *prev_lsps,
1445 const struct frame_type_desc *frame_desc,
1446 float *excitation, float *synth)
1447 {
1448 double i_lsps[MAX_LSPS];
1449 float lpcs[MAX_LSPS];
1450 float fac;
1451 int n;
1452
1453 if (frame_desc->acb_type == ACB_TYPE_NONE)
1454 synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation);
1455 else
1456 synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2,
1457 frame_desc, excitation);
1458
1459 /* convert interpolated LSPs to LPCs */
1460 fac = (block_idx + 0.5) / frame_desc->n_blocks;
1461 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1462 i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n]));
1463 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1464
1465 /* Speech synthesis */
1466 ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps);
1467 }
1468
1469 /**
1470 * Synthesize output samples for a single frame.
1471 *
1472 * @param ctx WMA Voice decoder context
1473 * @param gb bit I/O context (s->gb or one for cross-packet superframes)
1474 * @param frame_idx Frame number within superframe [0-2]
1475 * @param samples pointer to output sample buffer, has space for at least 160
1476 * samples
1477 * @param lsps LSP array
1478 * @param prev_lsps array of previous frame's LSPs
1479 * @param excitation target buffer for excitation signal
1480 * @param synth target buffer for synthesized speech data
1481 * @return 0 on success, <0 on error.
1482 */
synth_frame(AVCodecContext * ctx,GetBitContext * gb,int frame_idx,float * samples,const double * lsps,const double * prev_lsps,float * excitation,float * synth)1483 static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx,
1484 float *samples,
1485 const double *lsps, const double *prev_lsps,
1486 float *excitation, float *synth)
1487 {
1488 WMAVoiceContext *s = ctx->priv_data;
1489 int n, n_blocks_x2, log_n_blocks_x2, av_uninit(cur_pitch_val);
1490 int pitch[MAX_BLOCKS], av_uninit(last_block_pitch);
1491
1492 /* Parse frame type ("frame header"), see frame_descs */
1493 int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)], block_nsamples;
1494
1495 if (bd_idx < 0) {
1496 av_log(ctx, AV_LOG_ERROR,
1497 "Invalid frame type VLC code, skipping\n");
1498 return AVERROR_INVALIDDATA;
1499 }
1500
1501 block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks;
1502
1503 /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */
1504 if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) {
1505 /* Pitch is provided per frame, which is interpreted as the pitch of
1506 * the last sample of the last block of this frame. We can interpolate
1507 * the pitch of other blocks (and even pitch-per-sample) by gradually
1508 * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */
1509 n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1;
1510 log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1;
1511 cur_pitch_val = s->min_pitch_val + get_bits(gb, s->pitch_nbits);
1512 cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1);
1513 if (s->last_acb_type == ACB_TYPE_NONE ||
1514 20 * abs(cur_pitch_val - s->last_pitch_val) >
1515 (cur_pitch_val + s->last_pitch_val))
1516 s->last_pitch_val = cur_pitch_val;
1517
1518 /* pitch per block */
1519 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1520 int fac = n * 2 + 1;
1521
1522 pitch[n] = (MUL16(fac, cur_pitch_val) +
1523 MUL16((n_blocks_x2 - fac), s->last_pitch_val) +
1524 frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2;
1525 }
1526
1527 /* "pitch-diff-per-sample" for calculation of pitch per sample */
1528 s->pitch_diff_sh16 =
1529 (cur_pitch_val - s->last_pitch_val) * (1 << 16) / MAX_FRAMESIZE;
1530 }
1531
1532 /* Global gain (if silence) and pitch-adaptive window coordinates */
1533 switch (frame_descs[bd_idx].fcb_type) {
1534 case FCB_TYPE_SILENCE:
1535 s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)];
1536 break;
1537 case FCB_TYPE_AW_PULSES:
1538 aw_parse_coords(s, gb, pitch);
1539 break;
1540 }
1541
1542 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1543 int bl_pitch_sh2;
1544
1545 /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */
1546 switch (frame_descs[bd_idx].acb_type) {
1547 case ACB_TYPE_HAMMING: {
1548 /* Pitch is given per block. Per-block pitches are encoded as an
1549 * absolute value for the first block, and then delta values
1550 * relative to this value) for all subsequent blocks. The scale of
1551 * this pitch value is semi-logarithmic compared to its use in the
1552 * decoder, so we convert it to normal scale also. */
1553 int block_pitch,
1554 t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2,
1555 t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1,
1556 t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1;
1557
1558 if (n == 0) {
1559 block_pitch = get_bits(gb, s->block_pitch_nbits);
1560 } else
1561 block_pitch = last_block_pitch - s->block_delta_pitch_hrange +
1562 get_bits(gb, s->block_delta_pitch_nbits);
1563 /* Convert last_ so that any next delta is within _range */
1564 last_block_pitch = av_clip(block_pitch,
1565 s->block_delta_pitch_hrange,
1566 s->block_pitch_range -
1567 s->block_delta_pitch_hrange);
1568
1569 /* Convert semi-log-style scale back to normal scale */
1570 if (block_pitch < t1) {
1571 bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch;
1572 } else {
1573 block_pitch -= t1;
1574 if (block_pitch < t2) {
1575 bl_pitch_sh2 =
1576 (s->block_conv_table[1] << 2) + (block_pitch << 1);
1577 } else {
1578 block_pitch -= t2;
1579 if (block_pitch < t3) {
1580 bl_pitch_sh2 =
1581 (s->block_conv_table[2] + block_pitch) << 2;
1582 } else
1583 bl_pitch_sh2 = s->block_conv_table[3] << 2;
1584 }
1585 }
1586 pitch[n] = bl_pitch_sh2 >> 2;
1587 break;
1588 }
1589
1590 case ACB_TYPE_ASYMMETRIC: {
1591 bl_pitch_sh2 = pitch[n] << 2;
1592 break;
1593 }
1594
1595 default: // ACB_TYPE_NONE has no pitch
1596 bl_pitch_sh2 = 0;
1597 break;
1598 }
1599
1600 synth_block(s, gb, n, block_nsamples, bl_pitch_sh2,
1601 lsps, prev_lsps, &frame_descs[bd_idx],
1602 &excitation[n * block_nsamples],
1603 &synth[n * block_nsamples]);
1604 }
1605
1606 /* Averaging projection filter, if applicable. Else, just copy samples
1607 * from synthesis buffer */
1608 if (s->do_apf) {
1609 double i_lsps[MAX_LSPS];
1610 float lpcs[MAX_LSPS];
1611
1612 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1613 i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n]));
1614 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1615 postfilter(s, synth, samples, 80, lpcs,
1616 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx],
1617 frame_descs[bd_idx].fcb_type, pitch[0]);
1618
1619 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1620 i_lsps[n] = cos(lsps[n]);
1621 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1622 postfilter(s, &synth[80], &samples[80], 80, lpcs,
1623 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80],
1624 frame_descs[bd_idx].fcb_type, pitch[0]);
1625 } else
1626 memcpy(samples, synth, 160 * sizeof(synth[0]));
1627
1628 /* Cache values for next frame */
1629 s->frame_cntr++;
1630 if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%)
1631 s->last_acb_type = frame_descs[bd_idx].acb_type;
1632 switch (frame_descs[bd_idx].acb_type) {
1633 case ACB_TYPE_NONE:
1634 s->last_pitch_val = 0;
1635 break;
1636 case ACB_TYPE_ASYMMETRIC:
1637 s->last_pitch_val = cur_pitch_val;
1638 break;
1639 case ACB_TYPE_HAMMING:
1640 s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1];
1641 break;
1642 }
1643
1644 return 0;
1645 }
1646
1647 /**
1648 * Ensure minimum value for first item, maximum value for last value,
1649 * proper spacing between each value and proper ordering.
1650 *
1651 * @param lsps array of LSPs
1652 * @param num size of LSP array
1653 *
1654 * @note basically a double version of #ff_acelp_reorder_lsf(), might be
1655 * useful to put in a generic location later on. Parts are also
1656 * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(),
1657 * which is in float.
1658 */
stabilize_lsps(double * lsps,int num)1659 static void stabilize_lsps(double *lsps, int num)
1660 {
1661 int n, m, l;
1662
1663 /* set minimum value for first, maximum value for last and minimum
1664 * spacing between LSF values.
1665 * Very similar to ff_set_min_dist_lsf(), but in double. */
1666 lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI);
1667 for (n = 1; n < num; n++)
1668 lsps[n] = FFMAX(lsps[n], lsps[n - 1] + 0.0125 * M_PI);
1669 lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI);
1670
1671 /* reorder (looks like one-time / non-recursed bubblesort).
1672 * Very similar to ff_sort_nearly_sorted_floats(), but in double. */
1673 for (n = 1; n < num; n++) {
1674 if (lsps[n] < lsps[n - 1]) {
1675 for (m = 1; m < num; m++) {
1676 double tmp = lsps[m];
1677 for (l = m - 1; l >= 0; l--) {
1678 if (lsps[l] <= tmp) break;
1679 lsps[l + 1] = lsps[l];
1680 }
1681 lsps[l + 1] = tmp;
1682 }
1683 break;
1684 }
1685 }
1686 }
1687
1688 /**
1689 * Synthesize output samples for a single superframe. If we have any data
1690 * cached in s->sframe_cache, that will be used instead of whatever is loaded
1691 * in s->gb.
1692 *
1693 * WMA Voice superframes contain 3 frames, each containing 160 audio samples,
1694 * to give a total of 480 samples per frame. See #synth_frame() for frame
1695 * parsing. In addition to 3 frames, superframes can also contain the LSPs
1696 * (if these are globally specified for all frames (residually); they can
1697 * also be specified individually per-frame. See the s->has_residual_lsps
1698 * option), and can specify the number of samples encoded in this superframe
1699 * (if less than 480), usually used to prevent blanks at track boundaries.
1700 *
1701 * @param ctx WMA Voice decoder context
1702 * @return 0 on success, <0 on error or 1 if there was not enough data to
1703 * fully parse the superframe
1704 */
synth_superframe(AVCodecContext * ctx,AVFrame * frame,int * got_frame_ptr)1705 static int synth_superframe(AVCodecContext *ctx, AVFrame *frame,
1706 int *got_frame_ptr)
1707 {
1708 WMAVoiceContext *s = ctx->priv_data;
1709 GetBitContext *gb = &s->gb, s_gb;
1710 int n, res, n_samples = MAX_SFRAMESIZE;
1711 double lsps[MAX_FRAMES][MAX_LSPS];
1712 const double *mean_lsf = s->lsps == 16 ?
1713 wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode];
1714 float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12];
1715 float synth[MAX_LSPS + MAX_SFRAMESIZE];
1716 float *samples;
1717
1718 memcpy(synth, s->synth_history,
1719 s->lsps * sizeof(*synth));
1720 memcpy(excitation, s->excitation_history,
1721 s->history_nsamples * sizeof(*excitation));
1722
1723 if (s->sframe_cache_size > 0) {
1724 gb = &s_gb;
1725 init_get_bits(gb, s->sframe_cache, s->sframe_cache_size);
1726 s->sframe_cache_size = 0;
1727 }
1728
1729 /* First bit is speech/music bit, it differentiates between WMAVoice
1730 * speech samples (the actual codec) and WMAVoice music samples, which
1731 * are really WMAPro-in-WMAVoice-superframes. I've never seen those in
1732 * the wild yet. */
1733 if (!get_bits1(gb)) {
1734 avpriv_request_sample(ctx, "WMAPro-in-WMAVoice");
1735 return AVERROR_PATCHWELCOME;
1736 }
1737
1738 /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */
1739 if (get_bits1(gb)) {
1740 if ((n_samples = get_bits(gb, 12)) > MAX_SFRAMESIZE) {
1741 av_log(ctx, AV_LOG_ERROR,
1742 "Superframe encodes > %d samples (%d), not allowed\n",
1743 MAX_SFRAMESIZE, n_samples);
1744 return AVERROR_INVALIDDATA;
1745 }
1746 }
1747
1748 /* Parse LSPs, if global for the superframe (can also be per-frame). */
1749 if (s->has_residual_lsps) {
1750 double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2];
1751
1752 for (n = 0; n < s->lsps; n++)
1753 prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n];
1754
1755 if (s->lsps == 10) {
1756 dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1757 } else /* s->lsps == 16 */
1758 dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1759
1760 for (n = 0; n < s->lsps; n++) {
1761 lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]);
1762 lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]);
1763 lsps[2][n] += mean_lsf[n];
1764 }
1765 for (n = 0; n < 3; n++)
1766 stabilize_lsps(lsps[n], s->lsps);
1767 }
1768
1769 /* synth_superframe can run multiple times per packet
1770 * free potential previous frame */
1771 av_frame_unref(frame);
1772
1773 /* get output buffer */
1774 frame->nb_samples = MAX_SFRAMESIZE;
1775 if ((res = ff_get_buffer(ctx, frame, 0)) < 0)
1776 return res;
1777 frame->nb_samples = n_samples;
1778 samples = (float *)frame->data[0];
1779
1780 /* Parse frames, optionally preceded by per-frame (independent) LSPs. */
1781 for (n = 0; n < 3; n++) {
1782 if (!s->has_residual_lsps) {
1783 int m;
1784
1785 if (s->lsps == 10) {
1786 dequant_lsp10i(gb, lsps[n]);
1787 } else /* s->lsps == 16 */
1788 dequant_lsp16i(gb, lsps[n]);
1789
1790 for (m = 0; m < s->lsps; m++)
1791 lsps[n][m] += mean_lsf[m];
1792 stabilize_lsps(lsps[n], s->lsps);
1793 }
1794
1795 if ((res = synth_frame(ctx, gb, n,
1796 &samples[n * MAX_FRAMESIZE],
1797 lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1],
1798 &excitation[s->history_nsamples + n * MAX_FRAMESIZE],
1799 &synth[s->lsps + n * MAX_FRAMESIZE]))) {
1800 *got_frame_ptr = 0;
1801 return res;
1802 }
1803 }
1804
1805 /* Statistics? FIXME - we don't check for length, a slight overrun
1806 * will be caught by internal buffer padding, and anything else
1807 * will be skipped, not read. */
1808 if (get_bits1(gb)) {
1809 res = get_bits(gb, 4);
1810 skip_bits(gb, 10 * (res + 1));
1811 }
1812
1813 if (get_bits_left(gb) < 0) {
1814 wmavoice_flush(ctx);
1815 return AVERROR_INVALIDDATA;
1816 }
1817
1818 *got_frame_ptr = 1;
1819
1820 /* Update history */
1821 memcpy(s->prev_lsps, lsps[2],
1822 s->lsps * sizeof(*s->prev_lsps));
1823 memcpy(s->synth_history, &synth[MAX_SFRAMESIZE],
1824 s->lsps * sizeof(*synth));
1825 memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE],
1826 s->history_nsamples * sizeof(*excitation));
1827 if (s->do_apf)
1828 memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE],
1829 s->history_nsamples * sizeof(*s->zero_exc_pf));
1830
1831 return 0;
1832 }
1833
1834 /**
1835 * Parse the packet header at the start of each packet (input data to this
1836 * decoder).
1837 *
1838 * @param s WMA Voice decoding context private data
1839 * @return <0 on error, nb_superframes on success.
1840 */
parse_packet_header(WMAVoiceContext * s)1841 static int parse_packet_header(WMAVoiceContext *s)
1842 {
1843 GetBitContext *gb = &s->gb;
1844 unsigned int res, n_superframes = 0;
1845
1846 skip_bits(gb, 4); // packet sequence number
1847 s->has_residual_lsps = get_bits1(gb);
1848 do {
1849 if (get_bits_left(gb) < 6 + s->spillover_bitsize)
1850 return AVERROR_INVALIDDATA;
1851
1852 res = get_bits(gb, 6); // number of superframes per packet
1853 // (minus first one if there is spillover)
1854 n_superframes += res;
1855 } while (res == 0x3F);
1856 s->spillover_nbits = get_bits(gb, s->spillover_bitsize);
1857
1858 return get_bits_left(gb) >= 0 ? n_superframes : AVERROR_INVALIDDATA;
1859 }
1860
1861 /**
1862 * Copy (unaligned) bits from gb/data/size to pb.
1863 *
1864 * @param pb target buffer to copy bits into
1865 * @param data source buffer to copy bits from
1866 * @param size size of the source data, in bytes
1867 * @param gb bit I/O context specifying the current position in the source.
1868 * data. This function might use this to align the bit position to
1869 * a whole-byte boundary before calling #ff_copy_bits() on aligned
1870 * source data
1871 * @param nbits the amount of bits to copy from source to target
1872 *
1873 * @note after calling this function, the current position in the input bit
1874 * I/O context is undefined.
1875 */
copy_bits(PutBitContext * pb,const uint8_t * data,int size,GetBitContext * gb,int nbits)1876 static void copy_bits(PutBitContext *pb,
1877 const uint8_t *data, int size,
1878 GetBitContext *gb, int nbits)
1879 {
1880 int rmn_bytes, rmn_bits;
1881
1882 rmn_bits = rmn_bytes = get_bits_left(gb);
1883 if (rmn_bits < nbits)
1884 return;
1885 if (nbits > pb->size_in_bits - put_bits_count(pb))
1886 return;
1887 rmn_bits &= 7; rmn_bytes >>= 3;
1888 if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0)
1889 put_bits(pb, rmn_bits, get_bits(gb, rmn_bits));
1890 ff_copy_bits(pb, data + size - rmn_bytes,
1891 FFMIN(nbits - rmn_bits, rmn_bytes << 3));
1892 }
1893
1894 /**
1895 * Packet decoding: a packet is anything that the (ASF) demuxer contains,
1896 * and we expect that the demuxer / application provides it to us as such
1897 * (else you'll probably get garbage as output). Every packet has a size of
1898 * ctx->block_align bytes, starts with a packet header (see
1899 * #parse_packet_header()), and then a series of superframes. Superframe
1900 * boundaries may exceed packets, i.e. superframes can split data over
1901 * multiple (two) packets.
1902 *
1903 * For more information about frames, see #synth_superframe().
1904 */
wmavoice_decode_packet(AVCodecContext * ctx,void * data,int * got_frame_ptr,AVPacket * avpkt)1905 static int wmavoice_decode_packet(AVCodecContext *ctx, void *data,
1906 int *got_frame_ptr, AVPacket *avpkt)
1907 {
1908 WMAVoiceContext *s = ctx->priv_data;
1909 GetBitContext *gb = &s->gb;
1910 int size, res, pos;
1911
1912 /* Packets are sometimes a multiple of ctx->block_align, with a packet
1913 * header at each ctx->block_align bytes. However, FFmpeg's ASF demuxer
1914 * feeds us ASF packets, which may concatenate multiple "codec" packets
1915 * in a single "muxer" packet, so we artificially emulate that by
1916 * capping the packet size at ctx->block_align. */
1917 for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align);
1918 init_get_bits8(&s->gb, avpkt->data, size);
1919
1920 /* size == ctx->block_align is used to indicate whether we are dealing with
1921 * a new packet or a packet of which we already read the packet header
1922 * previously. */
1923 if (!(size % ctx->block_align)) { // new packet header
1924 if (!size) {
1925 s->spillover_nbits = 0;
1926 s->nb_superframes = 0;
1927 } else {
1928 if ((res = parse_packet_header(s)) < 0)
1929 return res;
1930 s->nb_superframes = res;
1931 }
1932
1933 /* If the packet header specifies a s->spillover_nbits, then we want
1934 * to push out all data of the previous packet (+ spillover) before
1935 * continuing to parse new superframes in the current packet. */
1936 if (s->sframe_cache_size > 0) {
1937 int cnt = get_bits_count(gb);
1938 if (cnt + s->spillover_nbits > avpkt->size * 8) {
1939 s->spillover_nbits = avpkt->size * 8 - cnt;
1940 }
1941 copy_bits(&s->pb, avpkt->data, size, gb, s->spillover_nbits);
1942 flush_put_bits(&s->pb);
1943 s->sframe_cache_size += s->spillover_nbits;
1944 if ((res = synth_superframe(ctx, data, got_frame_ptr)) == 0 &&
1945 *got_frame_ptr) {
1946 cnt += s->spillover_nbits;
1947 s->skip_bits_next = cnt & 7;
1948 res = cnt >> 3;
1949 return res;
1950 } else
1951 skip_bits_long (gb, s->spillover_nbits - cnt +
1952 get_bits_count(gb)); // resync
1953 } else if (s->spillover_nbits) {
1954 skip_bits_long(gb, s->spillover_nbits); // resync
1955 }
1956 } else if (s->skip_bits_next)
1957 skip_bits(gb, s->skip_bits_next);
1958
1959 /* Try parsing superframes in current packet */
1960 s->sframe_cache_size = 0;
1961 s->skip_bits_next = 0;
1962 pos = get_bits_left(gb);
1963 if (s->nb_superframes-- == 0) {
1964 *got_frame_ptr = 0;
1965 return size;
1966 } else if (s->nb_superframes > 0) {
1967 if ((res = synth_superframe(ctx, data, got_frame_ptr)) < 0) {
1968 return res;
1969 } else if (*got_frame_ptr) {
1970 int cnt = get_bits_count(gb);
1971 s->skip_bits_next = cnt & 7;
1972 res = cnt >> 3;
1973 return res;
1974 }
1975 } else if ((s->sframe_cache_size = pos) > 0) {
1976 /* ... cache it for spillover in next packet */
1977 init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE);
1978 copy_bits(&s->pb, avpkt->data, size, gb, s->sframe_cache_size);
1979 // FIXME bad - just copy bytes as whole and add use the
1980 // skip_bits_next field
1981 }
1982
1983 return size;
1984 }
1985
wmavoice_decode_end(AVCodecContext * ctx)1986 static av_cold int wmavoice_decode_end(AVCodecContext *ctx)
1987 {
1988 WMAVoiceContext *s = ctx->priv_data;
1989
1990 if (s->do_apf) {
1991 ff_rdft_end(&s->rdft);
1992 ff_rdft_end(&s->irdft);
1993 ff_dct_end(&s->dct);
1994 ff_dct_end(&s->dst);
1995 }
1996
1997 return 0;
1998 }
1999
2000 AVCodec ff_wmavoice_decoder = {
2001 .name = "wmavoice",
2002 .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"),
2003 .type = AVMEDIA_TYPE_AUDIO,
2004 .id = AV_CODEC_ID_WMAVOICE,
2005 .priv_data_size = sizeof(WMAVoiceContext),
2006 .init = wmavoice_decode_init,
2007 .close = wmavoice_decode_end,
2008 .decode = wmavoice_decode_packet,
2009 .capabilities = AV_CODEC_CAP_SUBFRAMES | AV_CODEC_CAP_DR1 | AV_CODEC_CAP_DELAY,
2010 .caps_internal = FF_CODEC_CAP_INIT_CLEANUP,
2011 .flush = wmavoice_flush,
2012 };
2013