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
45
46 #define MAX_BLOCKS 8 ///< maximum number of blocks per frame
47 #define MAX_LSPS 16
///< maximum filter order
48 #define MAX_LSPS_ALIGN16 16
///< same as #MAX_LSPS; needs to be multiple
49 ///< of 16 for ASM input buffer alignment
50 #define MAX_FRAMES 3
///< maximum number of frames per superframe
51 #define MAX_FRAMESIZE 160
///< maximum number of samples per frame
52 #define MAX_SIGNAL_HISTORY 416
///< maximum excitation signal history
53 #define MAX_SFRAMESIZE (MAX_FRAMESIZE * MAX_FRAMES)
54 ///< maximum number of samples per superframe
55 #define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that
56 ///< was split over two packets
57 #define VLC_NBITS 6
///< number of bits to read per VLC iteration
58
59 /**
60 * Frame type VLC coding.
61 */
63
64 /**
65 * Adaptive codebook types.
66 */
67 enum {
70 ///< we interpolate to get a per-sample pitch.
71 ///< Signal is generated using an asymmetric sinc
72 ///< window function
73 ///< @note see #wmavoice_ipol1_coeffs
75 ///< a Hamming sinc window function
76 ///< @note see #wmavoice_ipol2_coeffs
77 };
78
79 /**
80 * Fixed codebook types.
81 */
82 enum {
84 ///< generated from a hardcoded (fixed) codebook
85 ///< with per-frame (low) gain values
87 ///< gain values
89 ///< used in particular for low-bitrate streams
91 ///< combinations of either single pulses or
92 ///< pulse pairs
93 };
94
95 /**
96 * Description of frame types.
97 */
100 ///< (contains 160/#n_blocks samples)
105 ///< (rather than just one single pulse)
106 ///< only if #fcb_type == #FCB_TYPE_EXC_PULSES
107 uint16_t
frame_size;
///< the amount of bits that make up the block
108 ///< data (per frame)
127 };
128
129 /**
130 * WMA Voice decoding context.
131 */
133 /**
134 * @name Global values specified in the stream header / extradata or used all over.
135 * @{
136 */
138 ///< it contains the extradata from the
139 ///< demuxer. During decoding, it contains
140 ///< packet data.
141 int8_t vbm_tree[25];
///< converts VLC codes to frame type
142
143 int spillover_bitsize;
///< number of bits used to specify
144 ///< #spillover_nbits in the packet header
145 ///< = ceil(log2(ctx->block_align << 3))
146 int history_nsamples;
///< number of samples in history for signal
147 ///< prediction (through ACB)
148
149 /* postfilter specific values */
150 int do_apf;
///< whether to apply the averaged
151 ///< projection filter (APF)
152 int denoise_strength;
///< strength of denoising in Wiener filter
153 ///< [0-11]
154 int denoise_tilt_corr;
///< Whether to apply tilt correction to the
155 ///< Wiener filter coefficients (postfilter)
156 int dc_level;
///< Predicted amount of DC noise, based
157 ///< on which a DC removal filter is used
158
159 int lsps;
///< number of LSPs per frame [10 or 16]
161 int lsp_def_mode;
///< defines different sets of LSP defaults
162 ///< [0, 1]
163 int frame_lsp_bitsize;
///< size (in bits) of LSPs, when encoded
164 ///< per-frame (independent coding)
165 int sframe_lsp_bitsize;
///< size (in bits) of LSPs, when encoded
166 ///< per superframe (residual coding)
167
170 int pitch_nbits;
///< number of bits used to specify the
171 ///< pitch value in the frame header
172 int block_pitch_nbits;
///< number of bits used to specify the
173 ///< first block's pitch value
175 int block_delta_pitch_nbits;
///< number of bits used to specify the
176 ///< delta pitch between this and the last
177 ///< block's pitch value, used in all but
178 ///< first block
179 int block_delta_pitch_hrange;
///< 1/2 range of the delta (full range is
180 ///< from -this to +this-1)
181 uint16_t block_conv_table[4];
///< boundaries for block pitch unit/scale
182 ///< conversion
183
184 /**
185 * @}
186 *
187 * @name Packet values specified in the packet header or related to a packet.
188 *
189 * A packet is considered to be a single unit of data provided to this
190 * decoder by the demuxer.
191 * @{
192 */
193 int spillover_nbits;
///< number of bits of the previous packet's
194 ///< last superframe preceding this
195 ///< packet's first full superframe (useful
196 ///< for re-synchronization also)
197 int has_residual_lsps;
///< if set, superframes contain one set of
198 ///< LSPs that cover all frames, encoded as
199 ///< independent and residual LSPs; if not
200 ///< set, each frame contains its own, fully
201 ///< independent, LSPs
202 int skip_bits_next;
///< number of bits to skip at the next call
203 ///< to #wmavoice_decode_packet() (since
204 ///< they're part of the previous superframe)
205
207 ///< cache for superframe data split over
208 ///< multiple packets
209 int sframe_cache_size;
///< set to >0 if we have data from an
210 ///< (incomplete) superframe from a previous
211 ///< packet that spilled over in the current
212 ///< packet; specifies the amount of bits in
213 ///< #sframe_cache
215
216 /**
217 * @}
218 *
219 * @name Frame and superframe values
220 * Superframe and frame data - these can change from frame to frame,
221 * although some of them do in that case serve as a cache / history for
222 * the next frame or superframe.
223 * @{
224 */
225 double prev_lsps[
MAX_LSPS];
///< LSPs of the last frame of the previous
226 ///< superframe
229 int pitch_diff_sh16;
///< ((cur_pitch_val - #last_pitch_val)
230 ///< << 16) / #MAX_FRAMESIZE
232
233 int aw_idx_is_ext;
///< whether the AW index was encoded in
234 ///< 8 bits (instead of 6)
235 int aw_pulse_range;
///< the range over which #aw_pulse_set1()
236 ///< can apply the pulse, relative to the
237 ///< value in aw_first_pulse_off. The exact
238 ///< position of the first AW-pulse is within
239 ///< [pulse_off, pulse_off + this], and
240 ///< depends on bitstream values; [16 or 24]
241 int aw_n_pulses[2];
///< number of AW-pulses in each block; note
242 ///< that this number can be negative (in
243 ///< which case it basically means "zero")
244 int aw_first_pulse_off[2];
///< index of first sample to which to
245 ///< apply AW-pulses, or -0xff if unset
246 int aw_next_pulse_off_cache;
///< the position (relative to start of the
247 ///< second block) at which pulses should
248 ///< start to be positioned, serves as a
249 ///< cache for pitch-adaptive window pulses
250 ///< between blocks
251
252 int frame_cntr;
///< current frame index [0 - 0xFFFE]; is
253 ///< only used for comfort noise in #pRNG()
254 float gain_pred_err[6];
///< cache for gain prediction
256 ///< cache of the signal of previous
257 ///< superframes, used as a history for
258 ///< signal generation
259 float synth_history[
MAX_LSPS];
///< see #excitation_history
260 /**
261 * @}
262 *
263 * @name Postfilter values
264 *
265 * Variables used for postfilter implementation, mostly history for
266 * smoothing and so on, and context variables for FFT/iFFT.
267 * @{
268 */
270 ///< postfilter (for denoise filter)
271 DCTContext dct, dst;
///< contexts for phase shift (in Hilbert
272 ///< transform, part of postfilter)
273 float sin[511], cos[511];
///< 8-bit cosine/sine windows over [-pi,pi]
274 ///< range
275 float postfilter_agc;
///< gain control memory, used in
276 ///< #adaptive_gain_control()
277 float dcf_mem[2];
///< DC filter history
279 ///< zero filter output (i.e. excitation)
280 ///< by postfilter
284 ///< aligned buffer for LPC tilting
286 ///< aligned buffer for denoise coefficients
288 ///< aligned buffer for postfilter speech
289 ///< synthesis
290 /**
291 * @}
292 */
294
295 /**
296 * Set up the variable bit mode (VBM) tree from container extradata.
297 * @param gb bit I/O context.
298 * The bit context (s->gb) should be loaded with byte 23-46 of the
299 * container extradata (i.e. the ones containing the VBM tree).
300 * @param vbm_tree pointer to array to which the decoded VBM tree will be
301 * written.
302 * @return 0 on success, <0 on error.
303 */
305 {
306 int cntr[8] = { 0 },
n,
res;
307
308 memset(vbm_tree, 0xff, sizeof(vbm_tree[0]) * 25);
309 for (
n = 0;
n < 17;
n++) {
311 if (cntr[res] > 3) // should be >= 3 + (res == 7))
312 return -1;
313 vbm_tree[res * 3 + cntr[
res]++] =
n;
314 }
315 return 0;
316 }
317
319 {
321 2, 2, 2, 4, 4, 4,
322 6, 6, 6, 8, 8, 8,
323 10, 10, 10, 12, 12, 12,
324 14, 14, 14, 14
325 };
326 static const uint16_t codes[] = {
327 0x0000, 0x0001, 0x0002, // 00/01/10
328 0x000c, 0x000d, 0x000e, // 11+00/01/10
329 0x003c, 0x003d, 0x003e, // 1111+00/01/10
330 0x00fc, 0x00fd, 0x00fe, // 111111+00/01/10
331 0x03fc, 0x03fd, 0x03fe, // 11111111+00/01/10
332 0x0ffc, 0x0ffd, 0x0ffe, // 1111111111+00/01/10
333 0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx
334 };
335
337 bits, 1, 1, codes, 2, 2, 132);
338 }
339
340 /**
341 * Set up decoder with parameters from demuxer (extradata etc.).
342 */
344 {
345 int n,
flags, pitch_range, lsp16_flag;
347
348 /**
349 * Extradata layout:
350 * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c),
351 * - byte 19-22: flags field (annoyingly in LE; see below for known
352 * values),
353 * - byte 23-46: variable bitmode tree (really just 17 * 3 bits,
354 * rest is 0).
355 */
358 "Invalid extradata size %d (should be 46)\n",
361 }
370
372 memcpy(&s->
sin[255], s->
cos, 256 *
sizeof(s->
cos[0]));
373 for (n = 0; n < 255; n++) {
376 }
377 }
381 "Invalid denoise filter strength %d (max=11)\n",
384 }
389 lsp16_flag = flags & 0x1000;
390 if (lsp16_flag) {
394 } else {
398 }
399 for (n = 0; n < s->
lsps; n++)
401
406 }
407
411 if (pitch_range <= 0) {
414 }
419
421 int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8,
423
425 "Unsupported samplerate %d (min=%d, max=%d)\n",
427
429 }
430
439 }
445
449
450 return 0;
451 }
452
453 /**
454 * @name Postfilter functions
455 * Postfilter functions (gain control, wiener denoise filter, DC filter,
456 * kalman smoothening, plus surrounding code to wrap it)
457 * @{
458 */
459 /**
460 * Adaptive gain control (as used in postfilter).
461 *
462 * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except
463 * that the energy here is calculated using sum(abs(...)), whereas the
464 * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)).
465 *
466 * @param out output buffer for filtered samples
467 * @param in input buffer containing the samples as they are after the
468 * postfilter steps so far
469 * @param speech_synth input buffer containing speech synth before postfilter
470 * @param size input buffer size
471 * @param alpha exponential filter factor
472 * @param gain_mem pointer to filter memory (single float)
473 */
475 const float *speech_synth,
477 {
478 int i;
479 float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor;
480 float mem = *gain_mem;
481
482 for (i = 0; i <
size; i++) {
483 speech_energy += fabsf(speech_synth[i]);
484 postfilter_energy += fabsf(in[i]);
485 }
486 gain_scale_factor = (1.0 -
alpha) * speech_energy / postfilter_energy;
487
488 for (i = 0; i <
size; i++) {
489 mem = alpha * mem + gain_scale_factor;
490 out[i] = in[i] *
mem;
491 }
492
494 }
495
496 /**
497 * Kalman smoothing function.
498 *
499 * This function looks back pitch +/- 3 samples back into history to find
500 * the best fitting curve (that one giving the optimal gain of the two
501 * signals, i.e. the highest dot product between the two), and then
502 * uses that signal history to smoothen the output of the speech synthesis
503 * filter.
504 *
505 * @param s WMA Voice decoding context
506 * @param pitch pitch of the speech signal
507 * @param in input speech signal
508 * @param out output pointer for smoothened signal
509 * @param size input/output buffer size
510 *
511 * @returns -1 if no smoothening took place, e.g. because no optimal
512 * fit could be found, or 0 on success.
513 */
516 {
518 float optimal_gain = 0, dot;
521 *best_hist_ptr = NULL;
522
523 /* find best fitting point in history */
524 do {
526 if (dot > optimal_gain) {
527 optimal_gain = dot;
528 best_hist_ptr = ptr;
529 }
530 } while (--ptr >= end);
531
532 if (optimal_gain <= 0)
533 return -1;
535 if (dot <= 0) // would be 1.0
536 return -1;
537
538 if (optimal_gain <= dot) {
539 dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000
540 } else
541 dot = 0.625;
542
543 /* actual smoothing */
544 for (n = 0; n <
size; n++)
545 out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]);
546
547 return 0;
548 }
549
550 /**
551 * Get the tilt factor of a formant filter from its transfer function
552 * @see #tilt_factor() in amrnbdec.c, which does essentially the same,
553 * but somehow (??) it does a speech synthesis filter in the
554 * middle, which is missing here
555 *
556 * @param lpcs LPC coefficients
557 * @param n_lpcs Size of LPC buffer
558 * @returns the tilt factor
559 */
561 {
562 float rh0, rh1;
563
566
567 return rh1 / rh0;
568 }
569
570 /**
571 * Derive denoise filter coefficients (in real domain) from the LPCs.
572 */
574 int fcb_type,
float *
coeffs,
int remainder)
575 {
577 float irange, angle_mul, gain_mul, range, sq;
579
580 /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */
582 #define log_range(var, assign) do { \
583 float tmp = log10f(assign); var = tmp; \
584 max = FFMAX(max, tmp); min = FFMIN(min, tmp); \
585 } while (0)
586 log_range(last_coeff, lpcs[1] * lpcs[1]);
587 for (n = 1; n < 64; n++)
588 log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] +
589 lpcs[n * 2 + 1] * lpcs[n * 2 + 1]);
591 #undef log_range
594
595 /* Now, use this spectrum to pick out these frequencies with higher
596 * (relative) power/energy (which we then take to be "not noise"),
597 * and set up a table (still in lpc[]) of (relative) gains per frequency.
598 * These frequencies will be maintained, while others ("noise") will be
599 * decreased in the filter output. */
600 irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63]
602 (5.0 / 14.7));
604 for (n = 0; n <= 64; n++) {
605 float pwr;
606
607 idx =
FFMAX(0,
lrint((max - lpcs[n]) * irange) - 1);
609 lpcs[
n] = angle_mul * pwr;
610
611 /* 70.57 =~ 1/log10(1.0331663) */
612 idx = (pwr * gain_mul - 0.0295) * 70.570526123;
613 if (idx > 127) { // fall back if index falls outside table range
615 powf(1.0331663, idx - 127);
616 } else
618 }
619
620 /* calculate the Hilbert transform of the gains, which we do (since this
621 * is a sine input) by doing a phase shift (in theory, H(sin())=cos()).
622 * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the
623 * "moment" of the LPCs in this filter. */
626
627 /* Split out the coefficient indexes into phase/magnitude pairs */
628 idx = 255 + av_clip(lpcs[64], -255, 255);
629 coeffs[0] = coeffs[0] * s->
cos[idx];
630 idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255);
631 last_coeff = coeffs[64] * s->
cos[idx];
632 for (n = 63;; n--) {
633 idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255);
634 coeffs[n * 2 + 1] = coeffs[
n] * s->
sin[idx];
635 coeffs[n * 2] = coeffs[
n] * s->
cos[idx];
636
637 if (!--n) break;
638
639 idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255);
640 coeffs[n * 2 + 1] = coeffs[
n] * s->
sin[idx];
641 coeffs[n * 2] = coeffs[
n] * s->
cos[idx];
642 }
644
645 /* move into real domain */
647
648 /* tilt correction and normalize scale */
649 memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder));
651 float tilt_mem = 0;
652
653 coeffs[remainder - 1] = 0;
656 coeffs, remainder);
657 }
659 remainder));
660 for (n = 0; n < remainder; n++)
661 coeffs[n] *= sq;
662 }
663
664 /**
665 * This function applies a Wiener filter on the (noisy) speech signal as
666 * a means to denoise it.
667 *
668 * - take RDFT of LPCs to get the power spectrum of the noise + speech;
669 * - using this power spectrum, calculate (for each frequency) the Wiener
670 * filter gain, which depends on the frequency power and desired level
671 * of noise subtraction (when set too high, this leads to artifacts)
672 * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse
673 * of 4-8kHz);
674 * - by doing a phase shift, calculate the Hilbert transform of this array
675 * of per-frequency filter-gains to get the filtering coefficients;
676 * - smoothen/normalize/de-tilt these filter coefficients as desired;
677 * - take RDFT of noisy sound, apply the coefficients and take its IRDFT
678 * to get the denoised speech signal;
679 * - the leftover (i.e. output of the IRDFT on denoised speech data beyond
680 * the frame boundary) are saved and applied to subsequent frames by an
681 * overlap-add method (otherwise you get clicking-artifacts).
682 *
683 * @param s WMA Voice decoding context
684 * @param fcb_type Frame (codebook) type
685 * @param synth_pf input: the noisy speech signal, output: denoised speech
686 * data; should be 16-byte aligned (for ASM purposes)
687 * @param size size of the speech data
688 * @param lpcs LPCs used to synthesize this frame's speech data
689 */
691 float *synth_pf,
int size,
692 const float *lpcs)
693 {
694 int remainder, lim,
n;
695
699
700 tilted_lpcs[0] = 1.0;
701 memcpy(&tilted_lpcs[1], lpcs,
sizeof(lpcs[0]) * s->
lsps);
702 memset(&tilted_lpcs[s->
lsps + 1], 0,
703 sizeof(tilted_lpcs[0]) * (128 - s->
lsps - 1));
705 tilted_lpcs, s->
lsps + 2);
706
707 /* The IRDFT output (127 samples for 7-bit filter) beyond the frame
708 * size is applied to the next frame. All input beyond this is zero,
709 * and thus all output beyond this will go towards zero, hence we can
710 * limit to min(size-1, 127-size) as a performance consideration. */
711 remainder =
FFMIN(127 - size, size - 1);
713
714 /* apply coefficients (in frequency spectrum domain), i.e. complex
715 * number multiplication */
716 memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size));
721 for (n = 1; n < 64; n++) {
722 float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1];
723 synth_pf[n * 2] = v1 *
coeffs[n * 2] - v2 *
coeffs[n * 2 + 1];
724 synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1];
725 }
727 }
728
729 /* merge filter output with the history of previous runs */
732 for (n = 0; n < lim; n++)
737 }
738
739 /* move remainder of filter output into a cache for future runs */
742 for (n = 0; n < lim; n++)
744 if (lim < remainder) {
748 }
749 }
750 }
751
752 /**
753 * Averaging projection filter, the postfilter used in WMAVoice.
754 *
755 * This uses the following steps:
756 * - A zero-synthesis filter (generate excitation from synth signal)
757 * - Kalman smoothing on excitation, based on pitch
758 * - Re-synthesized smoothened output
759 * - Iterative Wiener denoise filter
760 * - Adaptive gain filter
761 * - DC filter
762 *
763 * @param s WMAVoice decoding context
764 * @param synth Speech synthesis output (before postfilter)
765 * @param samples Output buffer for filtered samples
766 * @param size Buffer size of synth & samples
767 * @param lpcs Generated LPCs used for speech synthesis
768 * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned)
769 * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses)
770 * @param pitch Pitch of the input signal
771 */
773 float *samples,
int size,
774 const float *lpcs, float *zero_exc_pf,
775 int fcb_type, int pitch)
776 {
779 *synth_filter_in = zero_exc_pf;
780
782
783 /* generate excitation from input signal */
785
788 synth_filter_in = synth_filter_in_buf;
789
790 /* re-synthesize speech after smoothening, and keep history */
792 synth_filter_in, size, s->
lsps);
793 memcpy(&synth_pf[-s->
lsps], &synth_pf[size - s->
lsps],
794 sizeof(synth_pf[0]) * s->
lsps);
795
797
800
802 /* remove ultra-low frequency DC noise / highpass filter;
803 * coefficients are identical to those used in SIPR decoding,
804 * and very closely resemble those used in AMR-NB decoding. */
806 (const float[2]) { -1.99997, 1.0 },
807 (const float[2]) { -1.9330735188, 0.93589198496 },
809 }
810 }
811 /**
812 * @}
813 */
814
815 /**
816 * Dequantize LSPs
817 * @param lsps output pointer to the array that will hold the LSPs
818 * @param num number of LSPs to be dequantized
819 * @param values quantized values, contains n_stages values
820 * @param sizes range (i.e. max value) of each quantized value
821 * @param n_stages number of dequantization runs
822 * @param table dequantization table to be used
823 * @param mul_q LSF multiplier
824 * @param base_q base (lowest) LSF values
825 */
827 const uint16_t *values,
828 const uint16_t *
sizes,
830 const double *mul_q,
831 const double *base_q)
832 {
834
835 memset(lsps, 0, num * sizeof(*lsps));
836 for (n = 0; n < n_stages; n++) {
837 const uint8_t *t_off = &table[values[
n] * num];
838 double base = base_q[
n], mul = mul_q[
n];
839
840 for (m = 0; m < num; m++)
841 lsps[m] += base + mul * t_off[m];
842
843 table += sizes[
n] * num;
844 }
845 }
846
847 /**
848 * @name LSP dequantization routines
849 * LSP dequantization routines, for 10/16LSPs and independent/residual coding.
850 * @note we assume enough bits are available, caller should check.
851 * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;
852 * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits.
853 * @{
854 */
855 /**
856 * Parse 10 independently-coded LSPs.
857 */
859 {
860 static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 };
861 static const double mul_lsf[4] = {
862 5.2187144800e-3, 1.4626986422e-3,
863 9.6179549166e-4, 1.1325736225e-3
864 };
865 static const double base_lsf[4] = {
866 M_PI * -2.15522e-1,
M_PI * -6.1646e-2,
867 M_PI * -3.3486e-2,
M_PI * -5.7408e-2
868 };
870
875
877 mul_lsf, base_lsf);
878 }
879
880 /**
881 * Parse 10 independently-coded LSPs, and then derive the tables to
882 * generate LSPs for the other frames from them (residual coding).
883 */
885 double *i_lsps, const double *old,
886 double *
a1,
double *
a2,
int q_mode)
887 {
888 static const uint16_t vec_sizes[3] = { 128, 64, 64 };
889 static const double mul_lsf[3] = {
890 2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3
891 };
892 static const double base_lsf[3] = {
893 M_PI * -1.07448e-1,
M_PI * -5.2706e-2,
M_PI * -5.1634e-2
894 };
895 const float (*ipol_tab)[2][10] = q_mode ?
899
901
906
907 for (n = 0; n < 10; n++) {
908 double delta = old[
n] - i_lsps[
n];
909 a1[
n] = ipol_tab[
interpol][0][
n] * delta + i_lsps[
n];
910 a1[10 +
n] = ipol_tab[
interpol][1][
n] * delta + i_lsps[
n];
911 }
912
914 mul_lsf, base_lsf);
915 }
916
917 /**
918 * Parse 16 independently-coded LSPs.
919 */
921 {
922 static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 };
923 static const double mul_lsf[5] = {
924 3.3439586280e-3, 6.9908173703e-4,
925 3.3216608306e-3, 1.0334960326e-3,
926 3.1899104283e-3
927 };
928 static const double base_lsf[5] = {
929 M_PI * -1.27576e-1,
M_PI * -2.4292e-2,
930 M_PI * -1.28094e-1,
M_PI * -3.2128e-2,
932 };
934
940
947 }
948
949 /**
950 * Parse 16 independently-coded LSPs, and then derive the tables to
951 * generate LSPs for the other frames from them (residual coding).
952 */
954 double *i_lsps, const double *old,
955 double *
a1,
double *
a2,
int q_mode)
956 {
957 static const uint16_t vec_sizes[3] = { 128, 128, 128 };
958 static const double mul_lsf[3] = {
959 1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3
960 };
961 static const double base_lsf[3] = {
963 };
964 const float (*ipol_tab)[2][16] = q_mode ?
968
970
975
976 for (n = 0; n < 16; n++) {
977 double delta = old[
n] - i_lsps[
n];
978 a1[
n] = ipol_tab[
interpol][0][
n] * delta + i_lsps[
n];
979 a1[16 +
n] = ipol_tab[
interpol][1][
n] * delta + i_lsps[
n];
980 }
981
988 }
989
990 /**
991 * @}
992 * @name Pitch-adaptive window coding functions
993 * The next few functions are for pitch-adaptive window coding.
994 * @{
995 */
996 /**
997 * Parse the offset of the first pitch-adaptive window pulses, and
998 * the distribution of pulses between the two blocks in this frame.
999 * @param s WMA Voice decoding context private data
1000 * @param gb bit I/O context
1001 * @param pitch pitch for each block in this frame
1002 */
1004 const int *pitch)
1005 {
1006 static const int16_t start_offset[94] = {
1007 -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11,
1008 13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26,
1009 27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43,
1010 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67,
1011 69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91,
1012 93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115,
1013 117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139,
1014 141, 143, 145, 147, 149, 151, 153, 155, 157, 159
1015 };
1017
1018 /* position of pulse */
1020 if ((bits =
get_bits(gb, 6)) >= 54) {
1022 bits += (bits - 54) * 3 +
get_bits(gb, 2);
1023 }
1024
1025 /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count
1026 * the distribution of the pulses in each block contained in this frame. */
1028 for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ;
1034
1035 /* if continuing from a position before the block, reset position to
1036 * start of block (when corrected for the range over which it can be
1037 * spread in aw_pulse_set1()). */
1041 if (start_offset[bits] < 0)
1044 }
1045 }
1046
1047 /**
1048 * Apply second set of pitch-adaptive window pulses.
1049 * @param s WMA Voice decoding context private data
1050 * @param gb bit I/O context
1051 * @param block_idx block index in frame [0, 1]
1052 * @param fcb structure containing fixed codebook vector info
1053 * @return -1 on error, 0 otherwise
1054 */
1057 {
1058 uint16_t use_mask_mem[9]; // only 5 are used, rest is padding
1059 uint16_t *use_mask = use_mask_mem + 2;
1060 /* in this function, idx is the index in the 80-bit (+ padding) use_mask
1061 * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits
1062 * of idx are the position of the bit within a particular item in the
1063 * array (0 being the most significant bit, and 15 being the least
1064 * significant bit), and the remainder (>> 4) is the index in the
1065 * use_mask[]-array. This is faster and uses less memory than using a
1066 * 80-byte/80-int array. */
1068 pulse_start,
n, idx, range, aidx, start_off = 0;
1069
1070 /* set offset of first pulse to within this block */
1074
1075 /* find range per pulse */
1077 if (block_idx == 0) {
1078 range = 32;
1079 } else /* block_idx = 1 */ {
1080 range = 8;
1083 }
1084 } else
1085 range = 16;
1086 pulse_start = s->
aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0;
1087
1088 /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,
1089 * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus
1090 * we exclude that range from being pulsed again in this function. */
1091 memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0]));
1092 memset( use_mask, -1, 5 * sizeof(use_mask[0]));
1093 memset(&use_mask[5], 0, 2 * sizeof(use_mask[0]));
1097 uint16_t *use_mask_ptr = &use_mask[idx >> 4];
1098 int first_sh = 16 - (idx & 15);
1099 *use_mask_ptr++ &= 0xFFFF
u << first_sh;
1100 excl_range -= first_sh;
1101 if (excl_range >= 16) {
1102 *use_mask_ptr++ = 0;
1103 *use_mask_ptr &= 0xFFFF >> (excl_range - 16);
1104 } else
1105 *use_mask_ptr &= 0xFFFF >> excl_range;
1106 }
1107
1108 /* find the 'aidx'th offset that is not excluded */
1110 for (n = 0; n <= aidx; pulse_start++) {
1111 for (idx = pulse_start; idx < 0; idx += fcb->
pitch_lag) ;
1113 if (use_mask[0]) idx = 0x0F;
1114 else if (use_mask[1]) idx = 0x1F;
1115 else if (use_mask[2]) idx = 0x2F;
1116 else if (use_mask[3]) idx = 0x3F;
1117 else if (use_mask[4]) idx = 0x4F;
1118 else return -1;
1120 }
1121 if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) {
1122 use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15));
1123 n++;
1124 start_off = idx;
1125 }
1126 }
1127
1128 fcb->
x[fcb->
n] = start_off;
1131
1132 /* set offset for next block, relative to start of that block */
1135 return 0;
1136 }
1137
1138 /**
1139 * Apply first set of pitch-adaptive window pulses.
1140 * @param s WMA Voice decoding context private data
1141 * @param gb bit I/O context
1142 * @param block_idx block index in frame [0, 1]
1143 * @param fcb storage location for fixed codebook pulse info
1144 */
1147 {
1150
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 +
1170 while (fcb->
x[fcb->
n] < 0)
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
1187 fcb->
x[fcb->
n + 1] = idx;
1188 fcb->
y[fcb->
n + 1] = (val & 1) ? -v : v;
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 * 715827883
U },
// y = 6
1220 { 4545, 0 * 390451573
U },
// y = 11
1221 { 3124, 11 * 268435456
U },
// y = 16
1222 { 2380, 15 * 204522253
U },
// y = 21
1223 { 1922, 23 * 165191050
U },
// y = 26
1224 { 1612, 23 * 138547333
U },
// y = 31
1225 { 1388, 27 * 119304648
U },
// y = 36
1226 { 1219, 16 * 104755300
U },
// y = 41
1227 { 1086, 39 * 93368855
U }
// 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 */
1243 int block_idx,
int size,
1245 float *excitation)
1246 {
1247 float gain;
1249
1251
1252 /* Set the offset from which we start reading wmavoice_std_codebook */
1256 } else /* FCB_TYPE_HARDCODED */ {
1259 }
1260
1261 /* Clear gain prediction parameters */
1263
1264 /* Apply gain to hardcoded codebook and use that as excitation signal */
1265 for (n = 0; n <
size; n++)
1267 }
1268
1269 /**
1270 * Parse FCB/ACB signal for a single block.
1271 * @note see #synth_block().
1272 */
1274 int block_idx,
int size,
1275 int block_pitch_sh2,
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 };
1283 int n, idx, gain_weight;
1285
1287 memset(pulses, 0, sizeof(*pulses) * size);
1288
1293
1294 /* For the other frame types, this is where we apply the innovation
1295 * (fixed) codebook pulses of the speech signal. */
1299 /* Conceal the block with silence and return.
1300 * Skip the correct amount of bits to read the next
1301 * block from the correct offset. */
1303
1304 for (n = 0; n <
size; n++)
1305 excitation[n] =
1308 return;
1309 }
1310 } else /* FCB_TYPE_EXC_PULSES */ {
1312
1314 /* similar to ff_decode_10_pulses_35bits(), but with single pulses
1315 * (instead of double) for a subset of pulses */
1316 for (n = 0; n < 5; n++) {
1317 float sign;
1318 int pos1, pos2;
1319
1322 fcb.
x[fcb.
n] = n + 5 * pos1;
1323 fcb.
y[fcb.
n++] = sign;
1324 if (n < frame_desc->dbl_pulses) {
1326 fcb.
x[fcb.
n] = n + 5 * pos2;
1327 fcb.
y[fcb.
n++] = (pos1 < pos2) ? -sign : sign;
1328 }
1329 }
1330 }
1332
1333 /* Calculate gain for adaptive & fixed codebook signal.
1334 * see ff_amr_set_fixed_gain(). */
1337 gain_coeff, 6) -
1341 -2.9957322736 /* log(0.05) */,
1342 1.6094379124 /* log(5.0) */);
1343
1347 for (n = 0; n < gain_weight; n++)
1349
1350 /* Calculation of adaptive codebook */
1353 for (n = 0; n <
size; n +=
len) {
1354 int next_idx_sh16;
1355 int abs_idx = block_idx * size +
n;
1358 int pitch = (pitch_sh16 + 0x6FFF) >> 16;
1359 int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000;
1360 idx = idx_sh16 >> 16;
1363 next_idx_sh16 = (idx_sh16) &~ 0xFFFF;
1364 } else
1365 next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF;
1367 1, size - n);
1368 } else
1370
1373 idx, 9, len);
1374 }
1375 } else /* ACB_TYPE_HAMMING */ {
1376 int block_pitch = block_pitch_sh2 >> 2;
1377 idx = block_pitch_sh2 & 3;
1378 if (idx) {
1381 idx, 8, size);
1382 } else
1384 sizeof(float) * size);
1385 }
1386
1387 /* Interpolate ACB/FCB and use as excitation signal */
1389 acb_gain, fcb_gain, size);
1390 }
1391
1392 /**
1393 * Parse data in a single block.
1394 * @note we assume enough bits are available, caller should check.
1395 *
1396 * @param s WMA Voice decoding context private data
1397 * @param gb bit I/O context
1398 * @param block_idx index of the to-be-read block
1399 * @param size amount of samples to be read in this block
1400 * @param block_pitch_sh2 pitch for this block << 2
1401 * @param lsps LSPs for (the end of) this frame
1402 * @param prev_lsps LSPs for the last frame
1403 * @param frame_desc frame type descriptor
1404 * @param excitation target memory for the ACB+FCB interpolated signal
1405 * @param synth target memory for the speech synthesis filter output
1406 * @return 0 on success, <0 on error.
1407 */
1409 int block_idx,
int size,
1410 int block_pitch_sh2,
1411 const double *lsps, const double *prev_lsps,
1413 float *excitation, float *synth)
1414 {
1417 float fac;
1419
1422 else
1424 frame_desc, excitation);
1425
1426 /* convert interpolated LSPs to LPCs */
1427 fac = (block_idx + 0.5) / frame_desc->
n_blocks;
1428 for (n = 0; n < s->
lsps; n++)
// LSF -> LSP
1429 i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n]));
1431
1432 /* Speech synthesis */
1434 }
1435
1436 /**
1437 * Synthesize output samples for a single frame.
1438 * @note we assume enough bits are available, caller should check.
1439 *
1440 * @param ctx WMA Voice decoder context
1441 * @param gb bit I/O context (s->gb or one for cross-packet superframes)
1442 * @param frame_idx Frame number within superframe [0-2]
1443 * @param samples pointer to output sample buffer, has space for at least 160
1444 * samples
1445 * @param lsps LSP array
1446 * @param prev_lsps array of previous frame's LSPs
1447 * @param excitation target buffer for excitation signal
1448 * @param synth target buffer for synthesized speech data
1449 * @return 0 on success, <0 on error.
1450 */
1452 float *samples,
1453 const double *lsps, const double *prev_lsps,
1454 float *excitation, float *synth)
1455 {
1457 int n, n_blocks_x2, log_n_blocks_x2,
av_uninit(cur_pitch_val);
1459
1460 /* Parse frame type ("frame header"), see frame_descs */
1462
1463 if (bd_idx < 0) {
1465 "Invalid frame type VLC code, skipping\n");
1467 }
1468
1470
1471 /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */
1473 /* Pitch is provided per frame, which is interpreted as the pitch of
1474 * the last sample of the last block of this frame. We can interpolate
1475 * the pitch of other blocks (and even pitch-per-sample) by gradually
1476 * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */
1485
1486 /* pitch per block */
1488 int fac = n * 2 + 1;
1489
1490 pitch[
n] = (
MUL16(fac, cur_pitch_val) +
1493 }
1494
1495 /* "pitch-diff-per-sample" for calculation of pitch per sample */
1498 }
1499
1500 /* Global gain (if silence) and pitch-adaptive window coordinates */
1504 break;
1507 break;
1508 }
1509
1511 int bl_pitch_sh2;
1512
1513 /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */
1516 /* Pitch is given per block. Per-block pitches are encoded as an
1517 * absolute value for the first block, and then delta values
1518 * relative to this value) for all subsequent blocks. The scale of
1519 * this pitch value is semi-logaritmic compared to its use in the
1520 * decoder, so we convert it to normal scale also. */
1521 int block_pitch,
1525
1526 if (n == 0) {
1528 } else
1531 /* Convert last_ so that any next delta is within _range */
1532 last_block_pitch = av_clip(block_pitch,
1536
1537 /* Convert semi-log-style scale back to normal scale */
1538 if (block_pitch < t1) {
1540 } else {
1542 if (block_pitch <
t2) {
1543 bl_pitch_sh2 =
1545 } else {
1547 if (block_pitch <
t3) {
1548 bl_pitch_sh2 =
1550 } else
1552 }
1553 }
1554 pitch[
n] = bl_pitch_sh2 >> 2;
1555 break;
1556 }
1557
1559 bl_pitch_sh2 = pitch[
n] << 2;
1560 break;
1561 }
1562
1563 default: // ACB_TYPE_NONE has no pitch
1564 bl_pitch_sh2 = 0;
1565 break;
1566 }
1567
1568 synth_block(s, gb, n, block_nsamples, bl_pitch_sh2,
1570 &excitation[n * block_nsamples],
1571 &synth[n * block_nsamples]);
1572 }
1573
1574 /* Averaging projection filter, if applicable. Else, just copy samples
1575 * from synthesis buffer */
1579
1580 for (n = 0; n < s->
lsps; n++)
// LSF -> LSP
1581 i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n]));
1586
1587 for (n = 0; n < s->
lsps; n++)
// LSF -> LSP
1588 i_lsps[n] = cos(lsps[n]);
1590 postfilter(s, &synth[80], &samples[80], 80, lpcs,
1593 } else
1594 memcpy(samples, synth, 160 * sizeof(synth[0]));
1595
1596 /* Cache values for next frame */
1603 break;
1606 break;
1609 break;
1610 }
1611
1612 return 0;
1613 }
1614
1615 /**
1616 * Ensure minimum value for first item, maximum value for last value,
1617 * proper spacing between each value and proper ordering.
1618 *
1619 * @param lsps array of LSPs
1620 * @param num size of LSP array
1621 *
1622 * @note basically a double version of #ff_acelp_reorder_lsf(), might be
1623 * useful to put in a generic location later on. Parts are also
1624 * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(),
1625 * which is in float.
1626 */
1628 {
1630
1631 /* set minimum value for first, maximum value for last and minimum
1632 * spacing between LSF values.
1633 * Very similar to ff_set_min_dist_lsf(), but in double. */
1634 lsps[0] =
FFMAX(lsps[0], 0.0015 *
M_PI);
1635 for (n = 1; n < num; n++)
1636 lsps[n] =
FFMAX(lsps[n], lsps[n - 1] + 0.0125 *
M_PI);
1637 lsps[num - 1] =
FFMIN(lsps[num - 1], 0.9985 *
M_PI);
1638
1639 /* reorder (looks like one-time / non-recursed bubblesort).
1640 * Very similar to ff_sort_nearly_sorted_floats(), but in double. */
1641 for (n = 1; n < num; n++) {
1642 if (lsps[n] < lsps[n - 1]) {
1643 for (m = 1; m < num; m++) {
1644 double tmp = lsps[
m];
1645 for (l = m - 1; l >= 0; l--) {
1646 if (lsps[l] <= tmp) break;
1647 lsps[l + 1] = lsps[l];
1648 }
1649 lsps[l + 1] = tmp;
1650 }
1651 break;
1652 }
1653 }
1654 }
1655
1656 /**
1657 * Test if there's enough bits to read 1 superframe.
1658 *
1659 * @param orig_gb bit I/O context used for reading. This function
1660 * does not modify the state of the bitreader; it
1661 * only uses it to copy the current stream position
1662 * @param s WMA Voice decoding context private data
1663 * @return < 0 on error, 1 on not enough bits or 0 if OK.
1664 */
1667 {
1669 int n, need_bits, bd_idx;
1671
1672 /* initialize a copy */
1676
1677 /* superframe header */
1679 return 1;
1681 return AVERROR(ENOSYS);
// WMAPro-in-WMAVoice superframe
1685 return 1;
1687 }
1688
1689 /* frames */
1691 int aw_idx_is_ext = 0;
1692
1696 }
1698 if (bd_idx < 0)
1703 return 1;
1705 }
1710 if (tmp >= 0x36) {
1712 aw_idx_is_ext = 1;
1713 }
1714 }
1715
1716 /* blocks */
1721 need_bits = 2 * !aw_idx_is_ext;
1722 } else
1723 need_bits = 0;
1726 return 1;
1728 }
1729
1730 return 0;
1731 }
1732
1733 /**
1734 * Synthesize output samples for a single superframe. If we have any data
1735 * cached in s->sframe_cache, that will be used instead of whatever is loaded
1736 * in s->gb.
1737 *
1738 * WMA Voice superframes contain 3 frames, each containing 160 audio samples,
1739 * to give a total of 480 samples per frame. See #synth_frame() for frame
1740 * parsing. In addition to 3 frames, superframes can also contain the LSPs
1741 * (if these are globally specified for all frames (residually); they can
1742 * also be specified individually per-frame. See the s->has_residual_lsps
1743 * option), and can specify the number of samples encoded in this superframe
1744 * (if less than 480), usually used to prevent blanks at track boundaries.
1745 *
1746 * @param ctx WMA Voice decoder context
1747 * @return 0 on success, <0 on error or 1 if there was not enough data to
1748 * fully parse the superframe
1749 */
1751 int *got_frame_ptr)
1752 {
1755 int n,
res, n_samples = 480;
1761 float *samples;
1762
1764 s->
lsps *
sizeof(*synth));
1767
1769 gb = &s_gb;
1772 }
1773
1775 *got_frame_ptr = 0;
1776 return 1;
1777 } else if (res < 0)
1779
1780 /* First bit is speech/music bit, it differentiates between WMAVoice
1781 * speech samples (the actual codec) and WMAVoice music samples, which
1782 * are really WMAPro-in-WMAVoice-superframes. I've never seen those in
1783 * the wild yet. */
1787 }
1788
1789 /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */
1791 if ((n_samples =
get_bits(gb, 12)) > 480) {
1793 "Superframe encodes >480 samples (%d), not allowed\n",
1794 n_samples);
1796 }
1797 }
1798 /* Parse LSPs, if global for the superframe (can also be per-frame). */
1801
1802 for (n = 0; n < s->
lsps; n++)
1803 prev_lsps[n] = s->
prev_lsps[n] - mean_lsf[n];
1804
1807 } else /* s->lsps == 16 */
1809
1810 for (n = 0; n < s->
lsps; n++) {
1811 lsps[0][
n] = mean_lsf[
n] + (a1[
n] - a2[n * 2]);
1812 lsps[1][
n] = mean_lsf[
n] + (a1[s->
lsps +
n] - a2[n * 2 + 1]);
1813 lsps[2][
n] += mean_lsf[
n];
1814 }
1815 for (n = 0; n < 3; n++)
1817 }
1818
1819 /* get output buffer */
1824 samples = (
float *)frame->
data[0];
1825
1826 /* Parse frames, optionally preceded by per-frame (independent) LSPs. */
1827 for (n = 0; n < 3; n++) {
1830
1831 if (s->
lsps == 10) {
1833 } else /* s->lsps == 16 */
1835
1836 for (m = 0; m < s->
lsps; m++)
1837 lsps[n][m] += mean_lsf[m];
1839 }
1840
1843 lsps[n], n == 0 ? s->
prev_lsps : lsps[n - 1],
1845 &synth[s->
lsps + n * MAX_FRAMESIZE]))) {
1846 *got_frame_ptr = 0;
1848 }
1849 }
1850
1851 /* Statistics? FIXME - we don't check for length, a slight overrun
1852 * will be caught by internal buffer padding, and anything else
1853 * will be skipped, not read. */
1857 }
1858
1859 *got_frame_ptr = 1;
1860
1861 /* Update history */
1865 s->
lsps *
sizeof(*synth));
1871
1872 return 0;
1873 }
1874
1875 /**
1876 * Parse the packet header at the start of each packet (input data to this
1877 * decoder).
1878 *
1879 * @param s WMA Voice decoding context private data
1880 * @return 1 if not enough bits were available, or 0 on success.
1881 */
1883 {
1886
1888 return 1;
1889 skip_bits(gb, 4);
// packet sequence number
1891 do {
1892 res =
get_bits(gb, 6);
// number of superframes per packet
1893 // (minus first one if there is spillover)
1895 return 1;
1896 } while (res == 0x3F);
1898
1899 return 0;
1900 }
1901
1902 /**
1903 * Copy (unaligned) bits from gb/data/size to pb.
1904 *
1905 * @param pb target buffer to copy bits into
1906 * @param data source buffer to copy bits from
1907 * @param size size of the source data, in bytes
1908 * @param gb bit I/O context specifying the current position in the source.
1909 * data. This function might use this to align the bit position to
1910 * a whole-byte boundary before calling #avpriv_copy_bits() on aligned
1911 * source data
1912 * @param nbits the amount of bits to copy from source to target
1913 *
1914 * @note after calling this function, the current position in the input bit
1915 * I/O context is undefined.
1916 */
1920 {
1921 int rmn_bytes, rmn_bits;
1922
1924 if (rmn_bits < nbits)
1925 return;
1927 return;
1928 rmn_bits &= 7; rmn_bytes >>= 3;
1929 if ((rmn_bits =
FFMIN(rmn_bits, nbits)) > 0)
1932 FFMIN(nbits - rmn_bits, rmn_bytes << 3));
1933 }
1934
1935 /**
1936 * Packet decoding: a packet is anything that the (ASF) demuxer contains,
1937 * and we expect that the demuxer / application provides it to us as such
1938 * (else you'll probably get garbage as output). Every packet has a size of
1939 * ctx->block_align bytes, starts with a packet header (see
1940 * #parse_packet_header()), and then a series of superframes. Superframe
1941 * boundaries may exceed packets, i.e. superframes can split data over
1942 * multiple (two) packets.
1943 *
1944 * For more information about frames, see #synth_superframe().
1945 */
1947 int *got_frame_ptr,
AVPacket *avpkt)
1948 {
1952
1953 /* Packets are sometimes a multiple of ctx->block_align, with a packet
1954 * header at each ctx->block_align bytes. However, FFmpeg's ASF demuxer
1955 * feeds us ASF packets, which may concatenate multiple "codec" packets
1956 * in a single "muxer" packet, so we artificially emulate that by
1957 * capping the packet size at ctx->block_align. */
1959 if (!size) {
1960 *got_frame_ptr = 0;
1961 return 0;
1962 }
1964
1965 /* size == ctx->block_align is used to indicate whether we are dealing with
1966 * a new packet or a packet of which we already read the packet header
1967 * previously. */
1968 if (size == ctx->
block_align) {
// new packet header
1971
1972 /* If the packet header specifies a s->spillover_nbits, then we want
1973 * to push out all data of the previous packet (+ spillover) before
1974 * continuing to parse new superframes in the current packet. */
1982 *got_frame_ptr) {
1985 return cnt >> 3;
1986 } else
1989 } else
1991 }
1994
1995 /* Try parsing superframes in current packet */
2001 } else if (*got_frame_ptr) {
2004 return cnt >> 3;
2006 /* rewind bit reader to start of last (incomplete) superframe... */
2010
2011 /* ...and cache it for spillover in next packet */
2014 // FIXME bad - just copy bytes as whole and add use the
2015 // skip_bits_next field
2016 }
2017
2019 }
2020
2022 {
2024
2030 }
2031
2032 return 0;
2033 }
2034
2036 {
2039
2043 for (n = 0; n < s->
lsps; n++)
2051
2060 }
2061 }
2062
2075 };