2 * Windows Media Audio Voice decoder.
3 * Copyright (c) 2009 Ronald S. Bultje
5 * This file is part of FFmpeg.
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.
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.
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
24 * @brief Windows Media Audio Voice compatible decoder
25 * @author Ronald S. Bultje <rsbultje@gmail.com>
32 #include "wmavoice_data.h"
33 #include "celp_math.h"
34 #include "celp_filters.h"
35 #include "acelp_vectors.h"
36 #include "acelp_filters.h"
38 #include "libavutil/lzo.h"
43 #define MAX_BLOCKS 8 ///< maximum number of blocks per frame
44 #define MAX_LSPS 16 ///< maximum filter order
45 #define MAX_LSPS_ALIGN16 16 ///< same as #MAX_LSPS; needs to be multiple
46 ///< of 16 for ASM input buffer alignment
47 #define MAX_FRAMES 3 ///< maximum number of frames per superframe
48 #define MAX_FRAMESIZE 160 ///< maximum number of samples per frame
49 #define MAX_SIGNAL_HISTORY 416 ///< maximum excitation signal history
50 #define MAX_SFRAMESIZE (MAX_FRAMESIZE * MAX_FRAMES)
51 ///< maximum number of samples per superframe
52 #define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that
53 ///< was split over two packets
54 #define VLC_NBITS 6 ///< number of bits to read per VLC iteration
57 * Frame type VLC coding.
59 static VLC frame_type_vlc;
62 * Adaptive codebook types.
65 ACB_TYPE_NONE = 0, ///< no adaptive codebook (only hardcoded fixed)
66 ACB_TYPE_ASYMMETRIC = 1, ///< adaptive codebook with per-frame pitch, which
67 ///< we interpolate to get a per-sample pitch.
68 ///< Signal is generated using an asymmetric sinc
70 ///< @note see #wmavoice_ipol1_coeffs
71 ACB_TYPE_HAMMING = 2 ///< Per-block pitch with signal generation using
72 ///< a Hamming sinc window function
73 ///< @note see #wmavoice_ipol2_coeffs
77 * Fixed codebook types.
80 FCB_TYPE_SILENCE = 0, ///< comfort noise during silence
81 ///< generated from a hardcoded (fixed) codebook
82 ///< with per-frame (low) gain values
83 FCB_TYPE_HARDCODED = 1, ///< hardcoded (fixed) codebook with per-block
85 FCB_TYPE_AW_PULSES = 2, ///< Pitch-adaptive window (AW) pulse signals,
86 ///< used in particular for low-bitrate streams
87 FCB_TYPE_EXC_PULSES = 3, ///< Innovation (fixed) codebook pulse sets in
88 ///< combinations of either single pulses or
93 * Description of frame types.
95 static const struct frame_type_desc {
96 uint8_t n_blocks; ///< amount of blocks per frame (each block
97 ///< (contains 160/#n_blocks samples)
98 uint8_t log_n_blocks; ///< log2(#n_blocks)
99 uint8_t acb_type; ///< Adaptive codebook type (ACB_TYPE_*)
100 uint8_t fcb_type; ///< Fixed codebook type (FCB_TYPE_*)
101 uint8_t dbl_pulses; ///< how many pulse vectors have pulse pairs
102 ///< (rather than just one single pulse)
103 ///< only if #fcb_type == #FCB_TYPE_EXC_PULSES
104 uint16_t frame_size; ///< the amount of bits that make up the block
105 ///< data (per frame)
106 } frame_descs[17] = {
107 { 1, 0, ACB_TYPE_NONE, FCB_TYPE_SILENCE, 0, 0 },
108 { 2, 1, ACB_TYPE_NONE, FCB_TYPE_HARDCODED, 0, 28 },
109 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_AW_PULSES, 0, 46 },
110 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 80 },
111 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 104 },
112 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 0, 108 },
113 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 132 },
114 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 168 },
115 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 64 },
116 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 80 },
117 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 104 },
118 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 108 },
119 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 132 },
120 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 168 },
121 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 176 },
122 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 208 },
123 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 256 }
127 * WMA Voice decoding context.
131 * @name Global values specified in the stream header / extradata or used all over.
134 GetBitContext gb; ///< packet bitreader. During decoder init,
135 ///< it contains the extradata from the
136 ///< demuxer. During decoding, it contains
138 int8_t vbm_tree[25]; ///< converts VLC codes to frame type
140 int spillover_bitsize; ///< number of bits used to specify
141 ///< #spillover_nbits in the packet header
142 ///< = ceil(log2(ctx->block_align << 3))
143 int history_nsamples; ///< number of samples in history for signal
144 ///< prediction (through ACB)
146 /* postfilter specific values */
147 int do_apf; ///< whether to apply the averaged
148 ///< projection filter (APF)
149 int denoise_strength; ///< strength of denoising in Wiener filter
151 int denoise_tilt_corr; ///< Whether to apply tilt correction to the
152 ///< Wiener filter coefficients (postfilter)
153 int dc_level; ///< Predicted amount of DC noise, based
154 ///< on which a DC removal filter is used
156 int lsps; ///< number of LSPs per frame [10 or 16]
157 int lsp_q_mode; ///< defines quantizer defaults [0, 1]
158 int lsp_def_mode; ///< defines different sets of LSP defaults
160 int frame_lsp_bitsize; ///< size (in bits) of LSPs, when encoded
161 ///< per-frame (independent coding)
162 int sframe_lsp_bitsize; ///< size (in bits) of LSPs, when encoded
163 ///< per superframe (residual coding)
165 int min_pitch_val; ///< base value for pitch parsing code
166 int max_pitch_val; ///< max value + 1 for pitch parsing
167 int pitch_nbits; ///< number of bits used to specify the
168 ///< pitch value in the frame header
169 int block_pitch_nbits; ///< number of bits used to specify the
170 ///< first block's pitch value
171 int block_pitch_range; ///< range of the block pitch
172 int block_delta_pitch_nbits; ///< number of bits used to specify the
173 ///< delta pitch between this and the last
174 ///< block's pitch value, used in all but
176 int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is
177 ///< from -this to +this-1)
178 uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale
184 * @name Packet values specified in the packet header or related to a packet.
186 * A packet is considered to be a single unit of data provided to this
187 * decoder by the demuxer.
190 int spillover_nbits; ///< number of bits of the previous packet's
191 ///< last superframe preceeding this
192 ///< packet's first full superframe (useful
193 ///< for re-synchronization also)
194 int has_residual_lsps; ///< if set, superframes contain one set of
195 ///< LSPs that cover all frames, encoded as
196 ///< independent and residual LSPs; if not
197 ///< set, each frame contains its own, fully
198 ///< independent, LSPs
199 int skip_bits_next; ///< number of bits to skip at the next call
200 ///< to #wmavoice_decode_packet() (since
201 ///< they're part of the previous superframe)
203 uint8_t sframe_cache[SFRAME_CACHE_MAXSIZE + FF_INPUT_BUFFER_PADDING_SIZE];
204 ///< cache for superframe data split over
205 ///< multiple packets
206 int sframe_cache_size; ///< set to >0 if we have data from an
207 ///< (incomplete) superframe from a previous
208 ///< packet that spilled over in the current
209 ///< packet; specifies the amount of bits in
211 PutBitContext pb; ///< bitstream writer for #sframe_cache
216 * @name Frame and superframe values
217 * Superframe and frame data - these can change from frame to frame,
218 * although some of them do in that case serve as a cache / history for
219 * the next frame or superframe.
222 double prev_lsps[MAX_LSPS]; ///< LSPs of the last frame of the previous
224 int last_pitch_val; ///< pitch value of the previous frame
225 int last_acb_type; ///< frame type [0-2] of the previous frame
226 int pitch_diff_sh16; ///< ((cur_pitch_val - #last_pitch_val)
227 ///< << 16) / #MAX_FRAMESIZE
228 float silence_gain; ///< set for use in blocks if #ACB_TYPE_NONE
230 int aw_idx_is_ext; ///< whether the AW index was encoded in
231 ///< 8 bits (instead of 6)
232 int aw_pulse_range; ///< the range over which #aw_pulse_set1()
233 ///< can apply the pulse, relative to the
234 ///< value in aw_first_pulse_off. The exact
235 ///< position of the first AW-pulse is within
236 ///< [pulse_off, pulse_off + this], and
237 ///< depends on bitstream values; [16 or 24]
238 int aw_n_pulses[2]; ///< number of AW-pulses in each block; note
239 ///< that this number can be negative (in
240 ///< which case it basically means "zero")
241 int aw_first_pulse_off[2]; ///< index of first sample to which to
242 ///< apply AW-pulses, or -0xff if unset
243 int aw_next_pulse_off_cache; ///< the position (relative to start of the
244 ///< second block) at which pulses should
245 ///< start to be positioned, serves as a
246 ///< cache for pitch-adaptive window pulses
249 int frame_cntr; ///< current frame index [0 - 0xFFFE]; is
250 ///< only used for comfort noise in #pRNG()
251 float gain_pred_err[6]; ///< cache for gain prediction
252 float excitation_history[MAX_SIGNAL_HISTORY];
253 ///< cache of the signal of previous
254 ///< superframes, used as a history for
255 ///< signal generation
256 float synth_history[MAX_LSPS]; ///< see #excitation_history
260 * @name Postfilter values
262 * Variables used for postfilter implementation, mostly history for
263 * smoothing and so on, and context variables for FFT/iFFT.
266 RDFTContext rdft, irdft; ///< contexts for FFT-calculation in the
267 ///< postfilter (for denoise filter)
268 DCTContext dct, dst; ///< contexts for phase shift (in Hilbert
269 ///< transform, part of postfilter)
270 float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi]
272 float postfilter_agc; ///< gain control memory, used in
273 ///< #adaptive_gain_control()
274 float dcf_mem[2]; ///< DC filter history
275 float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE];
276 ///< zero filter output (i.e. excitation)
278 float denoise_filter_cache[MAX_FRAMESIZE];
279 int denoise_filter_cache_size; ///< samples in #denoise_filter_cache
280 DECLARE_ALIGNED(32, float, tilted_lpcs_pf)[0x80];
281 ///< aligned buffer for LPC tilting
282 DECLARE_ALIGNED(32, float, denoise_coeffs_pf)[0x80];
283 ///< aligned buffer for denoise coefficients
284 DECLARE_ALIGNED(32, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16];
285 ///< aligned buffer for postfilter speech
293 * Set up the variable bit mode (VBM) tree from container extradata.
294 * @param gb bit I/O context.
295 * The bit context (s->gb) should be loaded with byte 23-46 of the
296 * container extradata (i.e. the ones containing the VBM tree).
297 * @param vbm_tree pointer to array to which the decoded VBM tree will be
299 * @return 0 on success, <0 on error.
301 static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25])
303 static const uint8_t bits[] = {
306 10, 10, 10, 12, 12, 12,
309 static const uint16_t codes[] = {
310 0x0000, 0x0001, 0x0002, // 00/01/10
311 0x000c, 0x000d, 0x000e, // 11+00/01/10
312 0x003c, 0x003d, 0x003e, // 1111+00/01/10
313 0x00fc, 0x00fd, 0x00fe, // 111111+00/01/10
314 0x03fc, 0x03fd, 0x03fe, // 11111111+00/01/10
315 0x0ffc, 0x0ffd, 0x0ffe, // 1111111111+00/01/10
316 0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx
320 memset(vbm_tree, 0xff, sizeof(vbm_tree[0]) * 25);
321 memset(cntr, 0, sizeof(cntr));
322 for (n = 0; n < 17; n++) {
323 res = get_bits(gb, 3);
324 if (cntr[res] > 3) // should be >= 3 + (res == 7))
326 vbm_tree[res * 3 + cntr[res]++] = n;
328 INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits),
329 bits, 1, 1, codes, 2, 2, 132);
334 * Set up decoder with parameters from demuxer (extradata etc.).
336 static av_cold int wmavoice_decode_init(AVCodecContext *ctx)
338 int n, flags, pitch_range, lsp16_flag;
339 WMAVoiceContext *s = ctx->priv_data;
343 * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c),
344 * - byte 19-22: flags field (annoyingly in LE; see below for known
346 * - byte 23-46: variable bitmode tree (really just 17 * 3 bits,
349 if (ctx->extradata_size != 46) {
350 av_log(ctx, AV_LOG_ERROR,
351 "Invalid extradata size %d (should be 46)\n",
352 ctx->extradata_size);
355 flags = AV_RL32(ctx->extradata + 18);
356 s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align);
357 s->do_apf = flags & 0x1;
359 ff_rdft_init(&s->rdft, 7, DFT_R2C);
360 ff_rdft_init(&s->irdft, 7, IDFT_C2R);
361 ff_dct_init(&s->dct, 6, DCT_I);
362 ff_dct_init(&s->dst, 6, DST_I);
364 ff_sine_window_init(s->cos, 256);
365 memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0]));
366 for (n = 0; n < 255; n++) {
367 s->sin[n] = -s->sin[510 - n];
368 s->cos[510 - n] = s->cos[n];
371 s->denoise_strength = (flags >> 2) & 0xF;
372 if (s->denoise_strength >= 12) {
373 av_log(ctx, AV_LOG_ERROR,
374 "Invalid denoise filter strength %d (max=11)\n",
375 s->denoise_strength);
378 s->denoise_tilt_corr = !!(flags & 0x40);
379 s->dc_level = (flags >> 7) & 0xF;
380 s->lsp_q_mode = !!(flags & 0x2000);
381 s->lsp_def_mode = !!(flags & 0x4000);
382 lsp16_flag = flags & 0x1000;
385 s->frame_lsp_bitsize = 34;
386 s->sframe_lsp_bitsize = 60;
389 s->frame_lsp_bitsize = 24;
390 s->sframe_lsp_bitsize = 48;
392 for (n = 0; n < s->lsps; n++)
393 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
395 init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3);
396 if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) {
397 av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n");
401 s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8;
402 s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8;
403 pitch_range = s->max_pitch_val - s->min_pitch_val;
404 if (pitch_range <= 0) {
405 av_log(ctx, AV_LOG_ERROR, "Invalid pitch range; broken extradata?\n");
408 s->pitch_nbits = av_ceil_log2(pitch_range);
409 s->last_pitch_val = 40;
410 s->last_acb_type = ACB_TYPE_NONE;
411 s->history_nsamples = s->max_pitch_val + 8;
413 if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) {
414 int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8,
415 max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8;
417 av_log(ctx, AV_LOG_ERROR,
418 "Unsupported samplerate %d (min=%d, max=%d)\n",
419 ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz
424 s->block_conv_table[0] = s->min_pitch_val;
425 s->block_conv_table[1] = (pitch_range * 25) >> 6;
426 s->block_conv_table[2] = (pitch_range * 44) >> 6;
427 s->block_conv_table[3] = s->max_pitch_val - 1;
428 s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF;
429 if (s->block_delta_pitch_hrange <= 0) {
430 av_log(ctx, AV_LOG_ERROR, "Invalid delta pitch hrange; broken extradata?\n");
433 s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange);
434 s->block_pitch_range = s->block_conv_table[2] +
435 s->block_conv_table[3] + 1 +
436 2 * (s->block_conv_table[1] - 2 * s->min_pitch_val);
437 s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range);
439 ctx->sample_fmt = AV_SAMPLE_FMT_FLT;
445 * @name Postfilter functions
446 * Postfilter functions (gain control, wiener denoise filter, DC filter,
447 * kalman smoothening, plus surrounding code to wrap it)
451 * Adaptive gain control (as used in postfilter).
453 * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except
454 * that the energy here is calculated using sum(abs(...)), whereas the
455 * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)).
457 * @param out output buffer for filtered samples
458 * @param in input buffer containing the samples as they are after the
459 * postfilter steps so far
460 * @param speech_synth input buffer containing speech synth before postfilter
461 * @param size input buffer size
462 * @param alpha exponential filter factor
463 * @param gain_mem pointer to filter memory (single float)
465 static void adaptive_gain_control(float *out, const float *in,
466 const float *speech_synth,
467 int size, float alpha, float *gain_mem)
470 float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor;
471 float mem = *gain_mem;
473 for (i = 0; i < size; i++) {
474 speech_energy += fabsf(speech_synth[i]);
475 postfilter_energy += fabsf(in[i]);
477 gain_scale_factor = (1.0 - alpha) * speech_energy / postfilter_energy;
479 for (i = 0; i < size; i++) {
480 mem = alpha * mem + gain_scale_factor;
481 out[i] = in[i] * mem;
488 * Kalman smoothing function.
490 * This function looks back pitch +/- 3 samples back into history to find
491 * the best fitting curve (that one giving the optimal gain of the two
492 * signals, i.e. the highest dot product between the two), and then
493 * uses that signal history to smoothen the output of the speech synthesis
496 * @param s WMA Voice decoding context
497 * @param pitch pitch of the speech signal
498 * @param in input speech signal
499 * @param out output pointer for smoothened signal
500 * @param size input/output buffer size
502 * @returns -1 if no smoothening took place, e.g. because no optimal
503 * fit could be found, or 0 on success.
505 static int kalman_smoothen(WMAVoiceContext *s, int pitch,
506 const float *in, float *out, int size)
509 float optimal_gain = 0, dot;
510 const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)],
511 *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)],
514 /* find best fitting point in history */
516 dot = ff_dot_productf(in, ptr, size);
517 if (dot > optimal_gain) {
521 } while (--ptr >= end);
523 if (optimal_gain <= 0)
525 dot = ff_dot_productf(best_hist_ptr, best_hist_ptr, size);
526 if (dot <= 0) // would be 1.0
529 if (optimal_gain <= dot) {
530 dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000
534 /* actual smoothing */
535 for (n = 0; n < size; n++)
536 out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]);
542 * Get the tilt factor of a formant filter from its transfer function
543 * @see #tilt_factor() in amrnbdec.c, which does essentially the same,
544 * but somehow (??) it does a speech synthesis filter in the
545 * middle, which is missing here
547 * @param lpcs LPC coefficients
548 * @param n_lpcs Size of LPC buffer
549 * @returns the tilt factor
551 static float tilt_factor(const float *lpcs, int n_lpcs)
555 rh0 = 1.0 + ff_dot_productf(lpcs, lpcs, n_lpcs);
556 rh1 = lpcs[0] + ff_dot_productf(lpcs, &lpcs[1], n_lpcs - 1);
562 * Derive denoise filter coefficients (in real domain) from the LPCs.
564 static void calc_input_response(WMAVoiceContext *s, float *lpcs,
565 int fcb_type, float *coeffs, int remainder)
567 float last_coeff, min = 15.0, max = -15.0;
568 float irange, angle_mul, gain_mul, range, sq;
571 /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */
572 s->rdft.rdft_calc(&s->rdft, lpcs);
573 #define log_range(var, assign) do { \
574 float tmp = log10f(assign); var = tmp; \
575 max = FFMAX(max, tmp); min = FFMIN(min, tmp); \
577 log_range(last_coeff, lpcs[1] * lpcs[1]);
578 for (n = 1; n < 64; n++)
579 log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] +
580 lpcs[n * 2 + 1] * lpcs[n * 2 + 1]);
581 log_range(lpcs[0], lpcs[0] * lpcs[0]);
584 lpcs[64] = last_coeff;
586 /* Now, use this spectrum to pick out these frequencies with higher
587 * (relative) power/energy (which we then take to be "not noise"),
588 * and set up a table (still in lpc[]) of (relative) gains per frequency.
589 * These frequencies will be maintained, while others ("noise") will be
590 * decreased in the filter output. */
591 irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63]
592 gain_mul = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) :
594 angle_mul = gain_mul * (8.0 * M_LN10 / M_PI);
595 for (n = 0; n <= 64; n++) {
598 idx = FFMAX(0, lrint((max - lpcs[n]) * irange) - 1);
599 pwr = wmavoice_denoise_power_table[s->denoise_strength][idx];
600 lpcs[n] = angle_mul * pwr;
602 /* 70.57 =~ 1/log10(1.0331663) */
603 idx = (pwr * gain_mul - 0.0295) * 70.570526123;
604 if (idx > 127) { // fallback if index falls outside table range
605 coeffs[n] = wmavoice_energy_table[127] *
606 powf(1.0331663, idx - 127);
608 coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)];
611 /* calculate the Hilbert transform of the gains, which we do (since this
612 * is a sinus input) by doing a phase shift (in theory, H(sin())=cos()).
613 * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the
614 * "moment" of the LPCs in this filter. */
615 s->dct.dct_calc(&s->dct, lpcs);
616 s->dst.dct_calc(&s->dst, lpcs);
618 /* Split out the coefficient indexes into phase/magnitude pairs */
619 idx = 255 + av_clip(lpcs[64], -255, 255);
620 coeffs[0] = coeffs[0] * s->cos[idx];
621 idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255);
622 last_coeff = coeffs[64] * s->cos[idx];
624 idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255);
625 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
626 coeffs[n * 2] = coeffs[n] * s->cos[idx];
630 idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255);
631 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
632 coeffs[n * 2] = coeffs[n] * s->cos[idx];
634 coeffs[1] = last_coeff;
636 /* move into real domain */
637 s->irdft.rdft_calc(&s->irdft, coeffs);
639 /* tilt correction and normalize scale */
640 memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder));
641 if (s->denoise_tilt_corr) {
644 coeffs[remainder - 1] = 0;
645 ff_tilt_compensation(&tilt_mem,
646 -1.8 * tilt_factor(coeffs, remainder - 1),
649 sq = (1.0 / 64.0) * sqrtf(1 / ff_dot_productf(coeffs, coeffs, remainder));
650 for (n = 0; n < remainder; n++)
655 * This function applies a Wiener filter on the (noisy) speech signal as
656 * a means to denoise it.
658 * - take RDFT of LPCs to get the power spectrum of the noise + speech;
659 * - using this power spectrum, calculate (for each frequency) the Wiener
660 * filter gain, which depends on the frequency power and desired level
661 * of noise subtraction (when set too high, this leads to artifacts)
662 * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse
664 * - by doing a phase shift, calculate the Hilbert transform of this array
665 * of per-frequency filter-gains to get the filtering coefficients;
666 * - smoothen/normalize/de-tilt these filter coefficients as desired;
667 * - take RDFT of noisy sound, apply the coefficients and take its IRDFT
668 * to get the denoised speech signal;
669 * - the leftover (i.e. output of the IRDFT on denoised speech data beyond
670 * the frame boundary) are saved and applied to subsequent frames by an
671 * overlap-add method (otherwise you get clicking-artifacts).
673 * @param s WMA Voice decoding context
674 * @param fcb_type Frame (codebook) type
675 * @param synth_pf input: the noisy speech signal, output: denoised speech
676 * data; should be 16-byte aligned (for ASM purposes)
677 * @param size size of the speech data
678 * @param lpcs LPCs used to synthesize this frame's speech data
680 static void wiener_denoise(WMAVoiceContext *s, int fcb_type,
681 float *synth_pf, int size,
684 int remainder, lim, n;
686 if (fcb_type != FCB_TYPE_SILENCE) {
687 float *tilted_lpcs = s->tilted_lpcs_pf,
688 *coeffs = s->denoise_coeffs_pf, tilt_mem = 0;
690 tilted_lpcs[0] = 1.0;
691 memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps);
692 memset(&tilted_lpcs[s->lsps + 1], 0,
693 sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1));
694 ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps),
695 tilted_lpcs, s->lsps + 2);
697 /* The IRDFT output (127 samples for 7-bit filter) beyond the frame
698 * size is applied to the next frame. All input beyond this is zero,
699 * and thus all output beyond this will go towards zero, hence we can
700 * limit to min(size-1, 127-size) as a performance consideration. */
701 remainder = FFMIN(127 - size, size - 1);
702 calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder);
704 /* apply coefficients (in frequency spectrum domain), i.e. complex
705 * number multiplication */
706 memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size));
707 s->rdft.rdft_calc(&s->rdft, synth_pf);
708 s->rdft.rdft_calc(&s->rdft, coeffs);
709 synth_pf[0] *= coeffs[0];
710 synth_pf[1] *= coeffs[1];
711 for (n = 1; n < 64; n++) {
712 float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1];
713 synth_pf[n * 2] = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1];
714 synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1];
716 s->irdft.rdft_calc(&s->irdft, synth_pf);
719 /* merge filter output with the history of previous runs */
720 if (s->denoise_filter_cache_size) {
721 lim = FFMIN(s->denoise_filter_cache_size, size);
722 for (n = 0; n < lim; n++)
723 synth_pf[n] += s->denoise_filter_cache[n];
724 s->denoise_filter_cache_size -= lim;
725 memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size],
726 sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size);
729 /* move remainder of filter output into a cache for future runs */
730 if (fcb_type != FCB_TYPE_SILENCE) {
731 lim = FFMIN(remainder, s->denoise_filter_cache_size);
732 for (n = 0; n < lim; n++)
733 s->denoise_filter_cache[n] += synth_pf[size + n];
734 if (lim < remainder) {
735 memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim],
736 sizeof(s->denoise_filter_cache[0]) * (remainder - lim));
737 s->denoise_filter_cache_size = remainder;
743 * Averaging projection filter, the postfilter used in WMAVoice.
745 * This uses the following steps:
746 * - A zero-synthesis filter (generate excitation from synth signal)
747 * - Kalman smoothing on excitation, based on pitch
748 * - Re-synthesized smoothened output
749 * - Iterative Wiener denoise filter
750 * - Adaptive gain filter
753 * @param s WMAVoice decoding context
754 * @param synth Speech synthesis output (before postfilter)
755 * @param samples Output buffer for filtered samples
756 * @param size Buffer size of synth & samples
757 * @param lpcs Generated LPCs used for speech synthesis
758 * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned)
759 * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses)
760 * @param pitch Pitch of the input signal
762 static void postfilter(WMAVoiceContext *s, const float *synth,
763 float *samples, int size,
764 const float *lpcs, float *zero_exc_pf,
765 int fcb_type, int pitch)
767 float synth_filter_in_buf[MAX_FRAMESIZE / 2],
768 *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16],
769 *synth_filter_in = zero_exc_pf;
771 assert(size <= MAX_FRAMESIZE / 2);
773 /* generate excitation from input signal */
774 ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps);
776 if (fcb_type >= FCB_TYPE_AW_PULSES &&
777 !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size))
778 synth_filter_in = synth_filter_in_buf;
780 /* re-synthesize speech after smoothening, and keep history */
781 ff_celp_lp_synthesis_filterf(synth_pf, lpcs,
782 synth_filter_in, size, s->lsps);
783 memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps],
784 sizeof(synth_pf[0]) * s->lsps);
786 wiener_denoise(s, fcb_type, synth_pf, size, lpcs);
788 adaptive_gain_control(samples, synth_pf, synth, size, 0.99,
791 if (s->dc_level > 8) {
792 /* remove ultra-low frequency DC noise / highpass filter;
793 * coefficients are identical to those used in SIPR decoding,
794 * and very closely resemble those used in AMR-NB decoding. */
795 ff_acelp_apply_order_2_transfer_function(samples, samples,
796 (const float[2]) { -1.99997, 1.0 },
797 (const float[2]) { -1.9330735188, 0.93589198496 },
798 0.93980580475, s->dcf_mem, size);
807 * @param lsps output pointer to the array that will hold the LSPs
808 * @param num number of LSPs to be dequantized
809 * @param values quantized values, contains n_stages values
810 * @param sizes range (i.e. max value) of each quantized value
811 * @param n_stages number of dequantization runs
812 * @param table dequantization table to be used
813 * @param mul_q LSF multiplier
814 * @param base_q base (lowest) LSF values
816 static void dequant_lsps(double *lsps, int num,
817 const uint16_t *values,
818 const uint16_t *sizes,
819 int n_stages, const uint8_t *table,
821 const double *base_q)
825 memset(lsps, 0, num * sizeof(*lsps));
826 for (n = 0; n < n_stages; n++) {
827 const uint8_t *t_off = &table[values[n] * num];
828 double base = base_q[n], mul = mul_q[n];
830 for (m = 0; m < num; m++)
831 lsps[m] += base + mul * t_off[m];
833 table += sizes[n] * num;
838 * @name LSP dequantization routines
839 * LSP dequantization routines, for 10/16LSPs and independent/residual coding.
840 * @note we assume enough bits are available, caller should check.
841 * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;
842 * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits.
846 * Parse 10 independently-coded LSPs.
848 static void dequant_lsp10i(GetBitContext *gb, double *lsps)
850 static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 };
851 static const double mul_lsf[4] = {
852 5.2187144800e-3, 1.4626986422e-3,
853 9.6179549166e-4, 1.1325736225e-3
855 static const double base_lsf[4] = {
856 M_PI * -2.15522e-1, M_PI * -6.1646e-2,
857 M_PI * -3.3486e-2, M_PI * -5.7408e-2
861 v[0] = get_bits(gb, 8);
862 v[1] = get_bits(gb, 6);
863 v[2] = get_bits(gb, 5);
864 v[3] = get_bits(gb, 5);
866 dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i,
871 * Parse 10 independently-coded LSPs, and then derive the tables to
872 * generate LSPs for the other frames from them (residual coding).
874 static void dequant_lsp10r(GetBitContext *gb,
875 double *i_lsps, const double *old,
876 double *a1, double *a2, int q_mode)
878 static const uint16_t vec_sizes[3] = { 128, 64, 64 };
879 static const double mul_lsf[3] = {
880 2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3
882 static const double base_lsf[3] = {
883 M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2
885 const float (*ipol_tab)[2][10] = q_mode ?
886 wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a;
887 uint16_t interpol, v[3];
890 dequant_lsp10i(gb, i_lsps);
892 interpol = get_bits(gb, 5);
893 v[0] = get_bits(gb, 7);
894 v[1] = get_bits(gb, 6);
895 v[2] = get_bits(gb, 6);
897 for (n = 0; n < 10; n++) {
898 double delta = old[n] - i_lsps[n];
899 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
900 a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
903 dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r,
908 * Parse 16 independently-coded LSPs.
910 static void dequant_lsp16i(GetBitContext *gb, double *lsps)
912 static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 };
913 static const double mul_lsf[5] = {
914 3.3439586280e-3, 6.9908173703e-4,
915 3.3216608306e-3, 1.0334960326e-3,
918 static const double base_lsf[5] = {
919 M_PI * -1.27576e-1, M_PI * -2.4292e-2,
920 M_PI * -1.28094e-1, M_PI * -3.2128e-2,
925 v[0] = get_bits(gb, 8);
926 v[1] = get_bits(gb, 6);
927 v[2] = get_bits(gb, 7);
928 v[3] = get_bits(gb, 6);
929 v[4] = get_bits(gb, 7);
931 dequant_lsps( lsps, 5, v, vec_sizes, 2,
932 wmavoice_dq_lsp16i1, mul_lsf, base_lsf);
933 dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2,
934 wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]);
935 dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1,
936 wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]);
940 * Parse 16 independently-coded LSPs, and then derive the tables to
941 * generate LSPs for the other frames from them (residual coding).
943 static void dequant_lsp16r(GetBitContext *gb,
944 double *i_lsps, const double *old,
945 double *a1, double *a2, int q_mode)
947 static const uint16_t vec_sizes[3] = { 128, 128, 128 };
948 static const double mul_lsf[3] = {
949 1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3
951 static const double base_lsf[3] = {
952 M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2
954 const float (*ipol_tab)[2][16] = q_mode ?
955 wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a;
956 uint16_t interpol, v[3];
959 dequant_lsp16i(gb, i_lsps);
961 interpol = get_bits(gb, 5);
962 v[0] = get_bits(gb, 7);
963 v[1] = get_bits(gb, 7);
964 v[2] = get_bits(gb, 7);
966 for (n = 0; n < 16; n++) {
967 double delta = old[n] - i_lsps[n];
968 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
969 a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
972 dequant_lsps( a2, 10, v, vec_sizes, 1,
973 wmavoice_dq_lsp16r1, mul_lsf, base_lsf);
974 dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1,
975 wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]);
976 dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1,
977 wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]);
982 * @name Pitch-adaptive window coding functions
983 * The next few functions are for pitch-adaptive window coding.
987 * Parse the offset of the first pitch-adaptive window pulses, and
988 * the distribution of pulses between the two blocks in this frame.
989 * @param s WMA Voice decoding context private data
990 * @param gb bit I/O context
991 * @param pitch pitch for each block in this frame
993 static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb,
996 static const int16_t start_offset[94] = {
997 -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11,
998 13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26,
999 27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43,
1000 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67,
1001 69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91,
1002 93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115,
1003 117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139,
1004 141, 143, 145, 147, 149, 151, 153, 155, 157, 159
1008 /* position of pulse */
1009 s->aw_idx_is_ext = 0;
1010 if ((bits = get_bits(gb, 6)) >= 54) {
1011 s->aw_idx_is_ext = 1;
1012 bits += (bits - 54) * 3 + get_bits(gb, 2);
1015 /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count
1016 * the distribution of the pulses in each block contained in this frame. */
1017 s->aw_pulse_range = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16;
1018 for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ;
1019 s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0];
1020 s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2;
1021 offset += s->aw_n_pulses[0] * pitch[0];
1022 s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1];
1023 s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2;
1025 /* if continuing from a position before the block, reset position to
1026 * start of block (when corrected for the range over which it can be
1027 * spread in aw_pulse_set1()). */
1028 if (start_offset[bits] < MAX_FRAMESIZE / 2) {
1029 while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0)
1030 s->aw_first_pulse_off[1] -= pitch[1];
1031 if (start_offset[bits] < 0)
1032 while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0)
1033 s->aw_first_pulse_off[0] -= pitch[0];
1038 * Apply second set of pitch-adaptive window pulses.
1039 * @param s WMA Voice decoding context private data
1040 * @param gb bit I/O context
1041 * @param block_idx block index in frame [0, 1]
1042 * @param fcb structure containing fixed codebook vector info
1044 static void aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb,
1045 int block_idx, AMRFixed *fcb)
1047 uint16_t use_mask_mem[9]; // only 5 are used, rest is padding
1048 uint16_t *use_mask = use_mask_mem + 2;
1049 /* in this function, idx is the index in the 80-bit (+ padding) use_mask
1050 * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits
1051 * of idx are the position of the bit within a particular item in the
1052 * array (0 being the most significant bit, and 15 being the least
1053 * significant bit), and the remainder (>> 4) is the index in the
1054 * use_mask[]-array. This is faster and uses less memory than using a
1055 * 80-byte/80-int array. */
1056 int pulse_off = s->aw_first_pulse_off[block_idx],
1057 pulse_start, n, idx, range, aidx, start_off = 0;
1059 /* set offset of first pulse to within this block */
1060 if (s->aw_n_pulses[block_idx] > 0)
1061 while (pulse_off + s->aw_pulse_range < 1)
1062 pulse_off += fcb->pitch_lag;
1064 /* find range per pulse */
1065 if (s->aw_n_pulses[0] > 0) {
1066 if (block_idx == 0) {
1068 } else /* block_idx = 1 */ {
1070 if (s->aw_n_pulses[block_idx] > 0)
1071 pulse_off = s->aw_next_pulse_off_cache;
1075 pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0;
1077 /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,
1078 * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus
1079 * we exclude that range from being pulsed again in this function. */
1080 memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0]));
1081 memset( use_mask, -1, 5 * sizeof(use_mask[0]));
1082 memset(&use_mask[5], 0, 2 * sizeof(use_mask[0]));
1083 if (s->aw_n_pulses[block_idx] > 0)
1084 for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) {
1085 int excl_range = s->aw_pulse_range; // always 16 or 24
1086 uint16_t *use_mask_ptr = &use_mask[idx >> 4];
1087 int first_sh = 16 - (idx & 15);
1088 *use_mask_ptr++ &= 0xFFFF << first_sh;
1089 excl_range -= first_sh;
1090 if (excl_range >= 16) {
1091 *use_mask_ptr++ = 0;
1092 *use_mask_ptr &= 0xFFFF >> (excl_range - 16);
1094 *use_mask_ptr &= 0xFFFF >> excl_range;
1097 /* find the 'aidx'th offset that is not excluded */
1098 aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4);
1099 for (n = 0; n <= aidx; pulse_start++) {
1100 for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ;
1101 if (idx >= MAX_FRAMESIZE / 2) { // find from zero
1102 if (use_mask[0]) idx = 0x0F;
1103 else if (use_mask[1]) idx = 0x1F;
1104 else if (use_mask[2]) idx = 0x2F;
1105 else if (use_mask[3]) idx = 0x3F;
1106 else if (use_mask[4]) idx = 0x4F;
1108 idx -= av_log2_16bit(use_mask[idx >> 4]);
1110 if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) {
1111 use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15));
1117 fcb->x[fcb->n] = start_off;
1118 fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0;
1121 /* set offset for next block, relative to start of that block */
1122 n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag;
1123 s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0;
1127 * Apply first set of pitch-adaptive window pulses.
1128 * @param s WMA Voice decoding context private data
1129 * @param gb bit I/O context
1130 * @param block_idx block index in frame [0, 1]
1131 * @param fcb storage location for fixed codebook pulse info
1133 static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb,
1134 int block_idx, AMRFixed *fcb)
1136 int val = get_bits(gb, 12 - 2 * (s->aw_idx_is_ext && !block_idx));
1139 if (s->aw_n_pulses[block_idx] > 0) {
1140 int n, v_mask, i_mask, sh, n_pulses;
1142 if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each
1147 } else { // 4 pulses, 1:sign + 2:index each
1154 for (n = n_pulses - 1; n >= 0; n--, val >>= sh) {
1155 fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0;
1156 fcb->x[fcb->n] = (val & i_mask) * n_pulses + n +
1157 s->aw_first_pulse_off[block_idx];
1158 while (fcb->x[fcb->n] < 0)
1159 fcb->x[fcb->n] += fcb->pitch_lag;
1160 if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2)
1164 int num2 = (val & 0x1FF) >> 1, delta, idx;
1166 if (num2 < 1 * 79) { delta = 1; idx = num2 + 1; }
1167 else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; }
1168 else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; }
1169 else { delta = 7; idx = num2 + 1 - 3 * 75; }
1170 v = (val & 0x200) ? -1.0 : 1.0;
1172 fcb->no_repeat_mask |= 3 << fcb->n;
1173 fcb->x[fcb->n] = idx - delta;
1175 fcb->x[fcb->n + 1] = idx;
1176 fcb->y[fcb->n + 1] = (val & 1) ? -v : v;
1184 * Generate a random number from frame_cntr and block_idx, which will lief
1185 * in the range [0, 1000 - block_size] (so it can be used as an index in a
1186 * table of size 1000 of which you want to read block_size entries).
1188 * @param frame_cntr current frame number
1189 * @param block_num current block index
1190 * @param block_size amount of entries we want to read from a table
1191 * that has 1000 entries
1192 * @return a (non-)random number in the [0, 1000 - block_size] range.
1194 static int pRNG(int frame_cntr, int block_num, int block_size)
1196 /* array to simplify the calculation of z:
1197 * y = (x % 9) * 5 + 6;
1198 * z = (49995 * x) / y;
1199 * Since y only has 9 values, we can remove the division by using a
1200 * LUT and using FASTDIV-style divisions. For each of the 9 values
1201 * of y, we can rewrite z as:
1202 * z = x * (49995 / y) + x * ((49995 % y) / y)
1203 * In this table, each col represents one possible value of y, the
1204 * first number is 49995 / y, and the second is the FASTDIV variant
1205 * of 49995 % y / y. */
1206 static const unsigned int div_tbl[9][2] = {
1207 { 8332, 3 * 715827883U }, // y = 6
1208 { 4545, 0 * 390451573U }, // y = 11
1209 { 3124, 11 * 268435456U }, // y = 16
1210 { 2380, 15 * 204522253U }, // y = 21
1211 { 1922, 23 * 165191050U }, // y = 26
1212 { 1612, 23 * 138547333U }, // y = 31
1213 { 1388, 27 * 119304648U }, // y = 36
1214 { 1219, 16 * 104755300U }, // y = 41
1215 { 1086, 39 * 93368855U } // y = 46
1217 unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr;
1218 if (x >= 0xFFFF) x -= 0xFFFF; // max value of x is 8*1877+0xFFFE=0x13AA6,
1219 // so this is effectively a modulo (%)
1220 y = x - 9 * MULH(477218589, x); // x % 9
1221 z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1]));
1222 // z = x * 49995 / (y * 5 + 6)
1223 return z % (1000 - block_size);
1227 * Parse hardcoded signal for a single block.
1228 * @note see #synth_block().
1230 static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb,
1231 int block_idx, int size,
1232 const struct frame_type_desc *frame_desc,
1238 assert(size <= MAX_FRAMESIZE);
1240 /* Set the offset from which we start reading wmavoice_std_codebook */
1241 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1242 r_idx = pRNG(s->frame_cntr, block_idx, size);
1243 gain = s->silence_gain;
1244 } else /* FCB_TYPE_HARDCODED */ {
1245 r_idx = get_bits(gb, 8);
1246 gain = wmavoice_gain_universal[get_bits(gb, 6)];
1249 /* Clear gain prediction parameters */
1250 memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err));
1252 /* Apply gain to hardcoded codebook and use that as excitation signal */
1253 for (n = 0; n < size; n++)
1254 excitation[n] = wmavoice_std_codebook[r_idx + n] * gain;
1258 * Parse FCB/ACB signal for a single block.
1259 * @note see #synth_block().
1261 static void synth_block_fcb_acb(WMAVoiceContext *s, GetBitContext *gb,
1262 int block_idx, int size,
1263 int block_pitch_sh2,
1264 const struct frame_type_desc *frame_desc,
1267 static const float gain_coeff[6] = {
1268 0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458
1270 float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain;
1271 int n, idx, gain_weight;
1274 assert(size <= MAX_FRAMESIZE / 2);
1275 memset(pulses, 0, sizeof(*pulses) * size);
1277 fcb.pitch_lag = block_pitch_sh2 >> 2;
1278 fcb.pitch_fac = 1.0;
1279 fcb.no_repeat_mask = 0;
1282 /* For the other frame types, this is where we apply the innovation
1283 * (fixed) codebook pulses of the speech signal. */
1284 if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1285 aw_pulse_set1(s, gb, block_idx, &fcb);
1286 aw_pulse_set2(s, gb, block_idx, &fcb);
1287 } else /* FCB_TYPE_EXC_PULSES */ {
1288 int offset_nbits = 5 - frame_desc->log_n_blocks;
1290 fcb.no_repeat_mask = -1;
1291 /* similar to ff_decode_10_pulses_35bits(), but with single pulses
1292 * (instead of double) for a subset of pulses */
1293 for (n = 0; n < 5; n++) {
1297 sign = get_bits1(gb) ? 1.0 : -1.0;
1298 pos1 = get_bits(gb, offset_nbits);
1299 fcb.x[fcb.n] = n + 5 * pos1;
1300 fcb.y[fcb.n++] = sign;
1301 if (n < frame_desc->dbl_pulses) {
1302 pos2 = get_bits(gb, offset_nbits);
1303 fcb.x[fcb.n] = n + 5 * pos2;
1304 fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign;
1308 ff_set_fixed_vector(pulses, &fcb, 1.0, size);
1310 /* Calculate gain for adaptive & fixed codebook signal.
1311 * see ff_amr_set_fixed_gain(). */
1312 idx = get_bits(gb, 7);
1313 fcb_gain = expf(ff_dot_productf(s->gain_pred_err, gain_coeff, 6) -
1314 5.2409161640 + wmavoice_gain_codebook_fcb[idx]);
1315 acb_gain = wmavoice_gain_codebook_acb[idx];
1316 pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx],
1317 -2.9957322736 /* log(0.05) */,
1318 1.6094379124 /* log(5.0) */);
1320 gain_weight = 8 >> frame_desc->log_n_blocks;
1321 memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err,
1322 sizeof(*s->gain_pred_err) * (6 - gain_weight));
1323 for (n = 0; n < gain_weight; n++)
1324 s->gain_pred_err[n] = pred_err;
1326 /* Calculation of adaptive codebook */
1327 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1329 for (n = 0; n < size; n += len) {
1331 int abs_idx = block_idx * size + n;
1332 int pitch_sh16 = (s->last_pitch_val << 16) +
1333 s->pitch_diff_sh16 * abs_idx;
1334 int pitch = (pitch_sh16 + 0x6FFF) >> 16;
1335 int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000;
1336 idx = idx_sh16 >> 16;
1337 if (s->pitch_diff_sh16) {
1338 if (s->pitch_diff_sh16 > 0) {
1339 next_idx_sh16 = (idx_sh16) &~ 0xFFFF;
1341 next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF;
1342 len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8,
1347 ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch],
1348 wmavoice_ipol1_coeffs, 17,
1351 } else /* ACB_TYPE_HAMMING */ {
1352 int block_pitch = block_pitch_sh2 >> 2;
1353 idx = block_pitch_sh2 & 3;
1355 ff_acelp_interpolatef(excitation, &excitation[-block_pitch],
1356 wmavoice_ipol2_coeffs, 4,
1359 av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch,
1360 sizeof(float) * size);
1363 /* Interpolate ACB/FCB and use as excitation signal */
1364 ff_weighted_vector_sumf(excitation, excitation, pulses,
1365 acb_gain, fcb_gain, size);
1369 * Parse data in a single block.
1370 * @note we assume enough bits are available, caller should check.
1372 * @param s WMA Voice decoding context private data
1373 * @param gb bit I/O context
1374 * @param block_idx index of the to-be-read block
1375 * @param size amount of samples to be read in this block
1376 * @param block_pitch_sh2 pitch for this block << 2
1377 * @param lsps LSPs for (the end of) this frame
1378 * @param prev_lsps LSPs for the last frame
1379 * @param frame_desc frame type descriptor
1380 * @param excitation target memory for the ACB+FCB interpolated signal
1381 * @param synth target memory for the speech synthesis filter output
1382 * @return 0 on success, <0 on error.
1384 static void synth_block(WMAVoiceContext *s, GetBitContext *gb,
1385 int block_idx, int size,
1386 int block_pitch_sh2,
1387 const double *lsps, const double *prev_lsps,
1388 const struct frame_type_desc *frame_desc,
1389 float *excitation, float *synth)
1391 double i_lsps[MAX_LSPS];
1392 float lpcs[MAX_LSPS];
1396 if (frame_desc->acb_type == ACB_TYPE_NONE)
1397 synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation);
1399 synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2,
1400 frame_desc, excitation);
1402 /* convert interpolated LSPs to LPCs */
1403 fac = (block_idx + 0.5) / frame_desc->n_blocks;
1404 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1405 i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n]));
1406 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1408 /* Speech synthesis */
1409 ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps);
1413 * Synthesize output samples for a single frame.
1414 * @note we assume enough bits are available, caller should check.
1416 * @param ctx WMA Voice decoder context
1417 * @param gb bit I/O context (s->gb or one for cross-packet superframes)
1418 * @param frame_idx Frame number within superframe [0-2]
1419 * @param samples pointer to output sample buffer, has space for at least 160
1421 * @param lsps LSP array
1422 * @param prev_lsps array of previous frame's LSPs
1423 * @param excitation target buffer for excitation signal
1424 * @param synth target buffer for synthesized speech data
1425 * @return 0 on success, <0 on error.
1427 static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx,
1429 const double *lsps, const double *prev_lsps,
1430 float *excitation, float *synth)
1432 WMAVoiceContext *s = ctx->priv_data;
1433 int n, n_blocks_x2, log_n_blocks_x2, cur_pitch_val;
1434 int pitch[MAX_BLOCKS], last_block_pitch;
1436 /* Parse frame type ("frame header"), see frame_descs */
1437 int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)],
1438 block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks;
1441 av_log(ctx, AV_LOG_ERROR,
1442 "Invalid frame type VLC code, skipping\n");
1446 /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */
1447 if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) {
1448 /* Pitch is provided per frame, which is interpreted as the pitch of
1449 * the last sample of the last block of this frame. We can interpolate
1450 * the pitch of other blocks (and even pitch-per-sample) by gradually
1451 * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */
1452 n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1;
1453 log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1;
1454 cur_pitch_val = s->min_pitch_val + get_bits(gb, s->pitch_nbits);
1455 cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1);
1456 if (s->last_acb_type == ACB_TYPE_NONE ||
1457 20 * abs(cur_pitch_val - s->last_pitch_val) >
1458 (cur_pitch_val + s->last_pitch_val))
1459 s->last_pitch_val = cur_pitch_val;
1461 /* pitch per block */
1462 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1463 int fac = n * 2 + 1;
1465 pitch[n] = (MUL16(fac, cur_pitch_val) +
1466 MUL16((n_blocks_x2 - fac), s->last_pitch_val) +
1467 frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2;
1470 /* "pitch-diff-per-sample" for calculation of pitch per sample */
1471 s->pitch_diff_sh16 =
1472 ((cur_pitch_val - s->last_pitch_val) << 16) / MAX_FRAMESIZE;
1475 /* Global gain (if silence) and pitch-adaptive window coordinates */
1476 switch (frame_descs[bd_idx].fcb_type) {
1477 case FCB_TYPE_SILENCE:
1478 s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)];
1480 case FCB_TYPE_AW_PULSES:
1481 aw_parse_coords(s, gb, pitch);
1485 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1488 /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */
1489 switch (frame_descs[bd_idx].acb_type) {
1490 case ACB_TYPE_HAMMING: {
1491 /* Pitch is given per block. Per-block pitches are encoded as an
1492 * absolute value for the first block, and then delta values
1493 * relative to this value) for all subsequent blocks. The scale of
1494 * this pitch value is semi-logaritmic compared to its use in the
1495 * decoder, so we convert it to normal scale also. */
1497 t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2,
1498 t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1,
1499 t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1;
1502 block_pitch = get_bits(gb, s->block_pitch_nbits);
1504 block_pitch = last_block_pitch - s->block_delta_pitch_hrange +
1505 get_bits(gb, s->block_delta_pitch_nbits);
1506 /* Convert last_ so that any next delta is within _range */
1507 last_block_pitch = av_clip(block_pitch,
1508 s->block_delta_pitch_hrange,
1509 s->block_pitch_range -
1510 s->block_delta_pitch_hrange);
1512 /* Convert semi-log-style scale back to normal scale */
1513 if (block_pitch < t1) {
1514 bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch;
1517 if (block_pitch < t2) {
1519 (s->block_conv_table[1] << 2) + (block_pitch << 1);
1522 if (block_pitch < t3) {
1524 (s->block_conv_table[2] + block_pitch) << 2;
1526 bl_pitch_sh2 = s->block_conv_table[3] << 2;
1529 pitch[n] = bl_pitch_sh2 >> 2;
1533 case ACB_TYPE_ASYMMETRIC: {
1534 bl_pitch_sh2 = pitch[n] << 2;
1538 default: // ACB_TYPE_NONE has no pitch
1543 synth_block(s, gb, n, block_nsamples, bl_pitch_sh2,
1544 lsps, prev_lsps, &frame_descs[bd_idx],
1545 &excitation[n * block_nsamples],
1546 &synth[n * block_nsamples]);
1549 /* Averaging projection filter, if applicable. Else, just copy samples
1550 * from synthesis buffer */
1552 double i_lsps[MAX_LSPS];
1553 float lpcs[MAX_LSPS];
1555 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1556 i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n]));
1557 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1558 postfilter(s, synth, samples, 80, lpcs,
1559 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx],
1560 frame_descs[bd_idx].fcb_type, pitch[0]);
1562 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1563 i_lsps[n] = cos(lsps[n]);
1564 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1565 postfilter(s, &synth[80], &samples[80], 80, lpcs,
1566 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80],
1567 frame_descs[bd_idx].fcb_type, pitch[0]);
1569 memcpy(samples, synth, 160 * sizeof(synth[0]));
1571 /* Cache values for next frame */
1573 if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%)
1574 s->last_acb_type = frame_descs[bd_idx].acb_type;
1575 switch (frame_descs[bd_idx].acb_type) {
1577 s->last_pitch_val = 0;
1579 case ACB_TYPE_ASYMMETRIC:
1580 s->last_pitch_val = cur_pitch_val;
1582 case ACB_TYPE_HAMMING:
1583 s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1];
1591 * Ensure minimum value for first item, maximum value for last value,
1592 * proper spacing between each value and proper ordering.
1594 * @param lsps array of LSPs
1595 * @param num size of LSP array
1597 * @note basically a double version of #ff_acelp_reorder_lsf(), might be
1598 * useful to put in a generic location later on. Parts are also
1599 * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(),
1600 * which is in float.
1602 static void stabilize_lsps(double *lsps, int num)
1606 /* set minimum value for first, maximum value for last and minimum
1607 * spacing between LSF values.
1608 * Very similar to ff_set_min_dist_lsf(), but in double. */
1609 lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI);
1610 for (n = 1; n < num; n++)
1611 lsps[n] = FFMAX(lsps[n], lsps[n - 1] + 0.0125 * M_PI);
1612 lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI);
1614 /* reorder (looks like one-time / non-recursed bubblesort).
1615 * Very similar to ff_sort_nearly_sorted_floats(), but in double. */
1616 for (n = 1; n < num; n++) {
1617 if (lsps[n] < lsps[n - 1]) {
1618 for (m = 1; m < num; m++) {
1619 double tmp = lsps[m];
1620 for (l = m - 1; l >= 0; l--) {
1621 if (lsps[l] <= tmp) break;
1622 lsps[l + 1] = lsps[l];
1632 * Test if there's enough bits to read 1 superframe.
1634 * @param orig_gb bit I/O context used for reading. This function
1635 * does not modify the state of the bitreader; it
1636 * only uses it to copy the current stream position
1637 * @param s WMA Voice decoding context private data
1638 * @return -1 if unsupported, 1 on not enough bits or 0 if OK.
1640 static int check_bits_for_superframe(GetBitContext *orig_gb,
1643 GetBitContext s_gb, *gb = &s_gb;
1644 int n, need_bits, bd_idx;
1645 const struct frame_type_desc *frame_desc;
1647 /* initialize a copy */
1648 init_get_bits(gb, orig_gb->buffer, orig_gb->size_in_bits);
1649 skip_bits_long(gb, get_bits_count(orig_gb));
1650 assert(get_bits_left(gb) == get_bits_left(orig_gb));
1652 /* superframe header */
1653 if (get_bits_left(gb) < 14)
1656 return -1; // WMAPro-in-WMAVoice superframe
1657 if (get_bits1(gb)) skip_bits(gb, 12); // number of samples in superframe
1658 if (s->has_residual_lsps) { // residual LSPs (for all frames)
1659 if (get_bits_left(gb) < s->sframe_lsp_bitsize)
1661 skip_bits_long(gb, s->sframe_lsp_bitsize);
1665 for (n = 0; n < MAX_FRAMES; n++) {
1666 int aw_idx_is_ext = 0;
1668 if (!s->has_residual_lsps) { // independent LSPs (per-frame)
1669 if (get_bits_left(gb) < s->frame_lsp_bitsize) return 1;
1670 skip_bits_long(gb, s->frame_lsp_bitsize);
1672 bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)];
1674 return -1; // invalid frame type VLC code
1675 frame_desc = &frame_descs[bd_idx];
1676 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1677 if (get_bits_left(gb) < s->pitch_nbits)
1679 skip_bits_long(gb, s->pitch_nbits);
1681 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1683 } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1684 int tmp = get_bits(gb, 6);
1692 if (frame_desc->acb_type == ACB_TYPE_HAMMING) {
1693 need_bits = s->block_pitch_nbits +
1694 (frame_desc->n_blocks - 1) * s->block_delta_pitch_nbits;
1695 } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1696 need_bits = 2 * !aw_idx_is_ext;
1699 need_bits += frame_desc->frame_size;
1700 if (get_bits_left(gb) < need_bits)
1702 skip_bits_long(gb, need_bits);
1709 * Synthesize output samples for a single superframe. If we have any data
1710 * cached in s->sframe_cache, that will be used instead of whatever is loaded
1713 * WMA Voice superframes contain 3 frames, each containing 160 audio samples,
1714 * to give a total of 480 samples per frame. See #synth_frame() for frame
1715 * parsing. In addition to 3 frames, superframes can also contain the LSPs
1716 * (if these are globally specified for all frames (residually); they can
1717 * also be specified individually per-frame. See the s->has_residual_lsps
1718 * option), and can specify the number of samples encoded in this superframe
1719 * (if less than 480), usually used to prevent blanks at track boundaries.
1721 * @param ctx WMA Voice decoder context
1722 * @param samples pointer to output buffer for voice samples
1723 * @param data_size pointer containing the size of #samples on input, and the
1724 * amount of #samples filled on output
1725 * @return 0 on success, <0 on error or 1 if there was not enough data to
1726 * fully parse the superframe
1728 static int synth_superframe(AVCodecContext *ctx,
1729 float *samples, int *data_size)
1731 WMAVoiceContext *s = ctx->priv_data;
1732 GetBitContext *gb = &s->gb, s_gb;
1733 int n, res, n_samples = 480;
1734 double lsps[MAX_FRAMES][MAX_LSPS];
1735 const double *mean_lsf = s->lsps == 16 ?
1736 wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode];
1737 float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12];
1738 float synth[MAX_LSPS + MAX_SFRAMESIZE];
1740 memcpy(synth, s->synth_history,
1741 s->lsps * sizeof(*synth));
1742 memcpy(excitation, s->excitation_history,
1743 s->history_nsamples * sizeof(*excitation));
1745 if (s->sframe_cache_size > 0) {
1747 init_get_bits(gb, s->sframe_cache, s->sframe_cache_size);
1748 s->sframe_cache_size = 0;
1751 if ((res = check_bits_for_superframe(gb, s)) == 1) return 1;
1753 /* First bit is speech/music bit, it differentiates between WMAVoice
1754 * speech samples (the actual codec) and WMAVoice music samples, which
1755 * are really WMAPro-in-WMAVoice-superframes. I've never seen those in
1757 if (!get_bits1(gb)) {
1758 av_log_missing_feature(ctx, "WMAPro-in-WMAVoice support", 1);
1762 /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */
1763 if (get_bits1(gb)) {
1764 if ((n_samples = get_bits(gb, 12)) > 480) {
1765 av_log(ctx, AV_LOG_ERROR,
1766 "Superframe encodes >480 samples (%d), not allowed\n",
1771 /* Parse LSPs, if global for the superframe (can also be per-frame). */
1772 if (s->has_residual_lsps) {
1773 double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2];
1775 for (n = 0; n < s->lsps; n++)
1776 prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n];
1778 if (s->lsps == 10) {
1779 dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1780 } else /* s->lsps == 16 */
1781 dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1783 for (n = 0; n < s->lsps; n++) {
1784 lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]);
1785 lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]);
1786 lsps[2][n] += mean_lsf[n];
1788 for (n = 0; n < 3; n++)
1789 stabilize_lsps(lsps[n], s->lsps);
1792 /* Parse frames, optionally preceeded by per-frame (independent) LSPs. */
1793 for (n = 0; n < 3; n++) {
1794 if (!s->has_residual_lsps) {
1797 if (s->lsps == 10) {
1798 dequant_lsp10i(gb, lsps[n]);
1799 } else /* s->lsps == 16 */
1800 dequant_lsp16i(gb, lsps[n]);
1802 for (m = 0; m < s->lsps; m++)
1803 lsps[n][m] += mean_lsf[m];
1804 stabilize_lsps(lsps[n], s->lsps);
1807 if ((res = synth_frame(ctx, gb, n,
1808 &samples[n * MAX_FRAMESIZE],
1809 lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1],
1810 &excitation[s->history_nsamples + n * MAX_FRAMESIZE],
1811 &synth[s->lsps + n * MAX_FRAMESIZE])))
1815 /* Statistics? FIXME - we don't check for length, a slight overrun
1816 * will be caught by internal buffer padding, and anything else
1817 * will be skipped, not read. */
1818 if (get_bits1(gb)) {
1819 res = get_bits(gb, 4);
1820 skip_bits(gb, 10 * (res + 1));
1823 /* Specify nr. of output samples */
1824 *data_size = n_samples * sizeof(float);
1826 /* Update history */
1827 memcpy(s->prev_lsps, lsps[2],
1828 s->lsps * sizeof(*s->prev_lsps));
1829 memcpy(s->synth_history, &synth[MAX_SFRAMESIZE],
1830 s->lsps * sizeof(*synth));
1831 memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE],
1832 s->history_nsamples * sizeof(*excitation));
1834 memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE],
1835 s->history_nsamples * sizeof(*s->zero_exc_pf));
1841 * Parse the packet header at the start of each packet (input data to this
1844 * @param s WMA Voice decoding context private data
1845 * @return 1 if not enough bits were available, or 0 on success.
1847 static int parse_packet_header(WMAVoiceContext *s)
1849 GetBitContext *gb = &s->gb;
1852 if (get_bits_left(gb) < 11)
1854 skip_bits(gb, 4); // packet sequence number
1855 s->has_residual_lsps = get_bits1(gb);
1857 res = get_bits(gb, 6); // number of superframes per packet
1858 // (minus first one if there is spillover)
1859 if (get_bits_left(gb) < 6 * (res == 0x3F) + s->spillover_bitsize)
1861 } while (res == 0x3F);
1862 s->spillover_nbits = get_bits(gb, s->spillover_bitsize);
1868 * Copy (unaligned) bits from gb/data/size to pb.
1870 * @param pb target buffer to copy bits into
1871 * @param data source buffer to copy bits from
1872 * @param size size of the source data, in bytes
1873 * @param gb bit I/O context specifying the current position in the source.
1874 * data. This function might use this to align the bit position to
1875 * a whole-byte boundary before calling #ff_copy_bits() on aligned
1877 * @param nbits the amount of bits to copy from source to target
1879 * @note after calling this function, the current position in the input bit
1880 * I/O context is undefined.
1882 static void copy_bits(PutBitContext *pb,
1883 const uint8_t *data, int size,
1884 GetBitContext *gb, int nbits)
1886 int rmn_bytes, rmn_bits;
1888 rmn_bits = rmn_bytes = get_bits_left(gb);
1889 if (rmn_bits < nbits)
1891 if (nbits > pb->size_in_bits - put_bits_count(pb))
1893 rmn_bits &= 7; rmn_bytes >>= 3;
1894 if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0)
1895 put_bits(pb, rmn_bits, get_bits(gb, rmn_bits));
1896 ff_copy_bits(pb, data + size - rmn_bytes,
1897 FFMIN(nbits - rmn_bits, rmn_bytes << 3));
1901 * Packet decoding: a packet is anything that the (ASF) demuxer contains,
1902 * and we expect that the demuxer / application provides it to us as such
1903 * (else you'll probably get garbage as output). Every packet has a size of
1904 * ctx->block_align bytes, starts with a packet header (see
1905 * #parse_packet_header()), and then a series of superframes. Superframe
1906 * boundaries may exceed packets, i.e. superframes can split data over
1907 * multiple (two) packets.
1909 * For more information about frames, see #synth_superframe().
1911 static int wmavoice_decode_packet(AVCodecContext *ctx, void *data,
1912 int *data_size, AVPacket *avpkt)
1914 WMAVoiceContext *s = ctx->priv_data;
1915 GetBitContext *gb = &s->gb;
1918 if (*data_size < 480 * sizeof(float)) {
1919 av_log(ctx, AV_LOG_ERROR,
1920 "Output buffer too small (%d given - %zu needed)\n",
1921 *data_size, 480 * sizeof(float));
1926 /* Packets are sometimes a multiple of ctx->block_align, with a packet
1927 * header at each ctx->block_align bytes. However, FFmpeg's ASF demuxer
1928 * feeds us ASF packets, which may concatenate multiple "codec" packets
1929 * in a single "muxer" packet, so we artificially emulate that by
1930 * capping the packet size at ctx->block_align. */
1931 for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align);
1934 init_get_bits(&s->gb, avpkt->data, size << 3);
1936 /* size == ctx->block_align is used to indicate whether we are dealing with
1937 * a new packet or a packet of which we already read the packet header
1939 if (size == ctx->block_align) { // new packet header
1940 if ((res = parse_packet_header(s)) < 0)
1943 /* If the packet header specifies a s->spillover_nbits, then we want
1944 * to push out all data of the previous packet (+ spillover) before
1945 * continuing to parse new superframes in the current packet. */
1946 if (s->spillover_nbits > 0) {
1947 if (s->sframe_cache_size > 0) {
1948 int cnt = get_bits_count(gb);
1949 copy_bits(&s->pb, avpkt->data, size, gb, s->spillover_nbits);
1950 flush_put_bits(&s->pb);
1951 s->sframe_cache_size += s->spillover_nbits;
1952 if ((res = synth_superframe(ctx, data, data_size)) == 0 &&
1954 cnt += s->spillover_nbits;
1955 s->skip_bits_next = cnt & 7;
1958 skip_bits_long (gb, s->spillover_nbits - cnt +
1959 get_bits_count(gb)); // resync
1961 skip_bits_long(gb, s->spillover_nbits); // resync
1963 } else if (s->skip_bits_next)
1964 skip_bits(gb, s->skip_bits_next);
1966 /* Try parsing superframes in current packet */
1967 s->sframe_cache_size = 0;
1968 s->skip_bits_next = 0;
1969 pos = get_bits_left(gb);
1970 if ((res = synth_superframe(ctx, data, data_size)) < 0) {
1972 } else if (*data_size > 0) {
1973 int cnt = get_bits_count(gb);
1974 s->skip_bits_next = cnt & 7;
1976 } else if ((s->sframe_cache_size = pos) > 0) {
1977 /* rewind bit reader to start of last (incomplete) superframe... */
1978 init_get_bits(gb, avpkt->data, size << 3);
1979 skip_bits_long(gb, (size << 3) - pos);
1980 assert(get_bits_left(gb) == pos);
1982 /* ...and cache it for spillover in next packet */
1983 init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE);
1984 copy_bits(&s->pb, avpkt->data, size, gb, s->sframe_cache_size);
1985 // FIXME bad - just copy bytes as whole and add use the
1986 // skip_bits_next field
1992 static av_cold int wmavoice_decode_end(AVCodecContext *ctx)
1994 WMAVoiceContext *s = ctx->priv_data;
1997 ff_rdft_end(&s->rdft);
1998 ff_rdft_end(&s->irdft);
1999 ff_dct_end(&s->dct);
2000 ff_dct_end(&s->dst);
2006 static av_cold void wmavoice_flush(AVCodecContext *ctx)
2008 WMAVoiceContext *s = ctx->priv_data;
2011 s->postfilter_agc = 0;
2012 s->sframe_cache_size = 0;
2013 s->skip_bits_next = 0;
2014 for (n = 0; n < s->lsps; n++)
2015 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
2016 memset(s->excitation_history, 0,
2017 sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY);
2018 memset(s->synth_history, 0,
2019 sizeof(*s->synth_history) * MAX_LSPS);
2020 memset(s->gain_pred_err, 0,
2021 sizeof(s->gain_pred_err));
2024 memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0,
2025 sizeof(*s->synth_filter_out_buf) * s->lsps);
2026 memset(s->dcf_mem, 0,
2027 sizeof(*s->dcf_mem) * 2);
2028 memset(s->zero_exc_pf, 0,
2029 sizeof(*s->zero_exc_pf) * s->history_nsamples);
2030 memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache));
2034 AVCodec ff_wmavoice_decoder = {
2036 .type = AVMEDIA_TYPE_AUDIO,
2037 .id = CODEC_ID_WMAVOICE,
2038 .priv_data_size = sizeof(WMAVoiceContext),
2039 .init = wmavoice_decode_init,
2040 .close = wmavoice_decode_end,
2041 .decode = wmavoice_decode_packet,
2042 .capabilities = CODEC_CAP_SUBFRAMES,
2043 .flush = wmavoice_flush,
2044 .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"),