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"
42 #define MAX_BLOCKS 8 ///< maximum number of blocks per frame
43 #define MAX_LSPS 16 ///< maximum filter order
44 #define MAX_LSPS_ALIGN16 16 ///< same as #MAX_LSPS; needs to be multiple
45 ///< of 16 for ASM input buffer alignment
46 #define MAX_FRAMES 3 ///< maximum number of frames per superframe
47 #define MAX_FRAMESIZE 160 ///< maximum number of samples per frame
48 #define MAX_SIGNAL_HISTORY 416 ///< maximum excitation signal history
49 #define MAX_SFRAMESIZE (MAX_FRAMESIZE * MAX_FRAMES)
50 ///< maximum number of samples per superframe
51 #define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that
52 ///< was split over two packets
53 #define VLC_NBITS 6 ///< number of bits to read per VLC iteration
56 * Frame type VLC coding.
58 static VLC frame_type_vlc;
61 * Adaptive codebook types.
64 ACB_TYPE_NONE = 0, ///< no adaptive codebook (only hardcoded fixed)
65 ACB_TYPE_ASYMMETRIC = 1, ///< adaptive codebook with per-frame pitch, which
66 ///< we interpolate to get a per-sample pitch.
67 ///< Signal is generated using an asymmetric sinc
69 ///< @note see #wmavoice_ipol1_coeffs
70 ACB_TYPE_HAMMING = 2 ///< Per-block pitch with signal generation using
71 ///< a Hamming sinc window function
72 ///< @note see #wmavoice_ipol2_coeffs
76 * Fixed codebook types.
79 FCB_TYPE_SILENCE = 0, ///< comfort noise during silence
80 ///< generated from a hardcoded (fixed) codebook
81 ///< with per-frame (low) gain values
82 FCB_TYPE_HARDCODED = 1, ///< hardcoded (fixed) codebook with per-block
84 FCB_TYPE_AW_PULSES = 2, ///< Pitch-adaptive window (AW) pulse signals,
85 ///< used in particular for low-bitrate streams
86 FCB_TYPE_EXC_PULSES = 3, ///< Innovation (fixed) codebook pulse sets in
87 ///< combinations of either single pulses or
92 * Description of frame types.
94 static const struct frame_type_desc {
95 uint8_t n_blocks; ///< amount of blocks per frame (each block
96 ///< (contains 160/#n_blocks samples)
97 uint8_t log_n_blocks; ///< log2(#n_blocks)
98 uint8_t acb_type; ///< Adaptive codebook type (ACB_TYPE_*)
99 uint8_t fcb_type; ///< Fixed codebook type (FCB_TYPE_*)
100 uint8_t dbl_pulses; ///< how many pulse vectors have pulse pairs
101 ///< (rather than just one single pulse)
102 ///< only if #fcb_type == #FCB_TYPE_EXC_PULSES
103 uint16_t frame_size; ///< the amount of bits that make up the block
104 ///< data (per frame)
105 } frame_descs[17] = {
106 { 1, 0, ACB_TYPE_NONE, FCB_TYPE_SILENCE, 0, 0 },
107 { 2, 1, ACB_TYPE_NONE, FCB_TYPE_HARDCODED, 0, 28 },
108 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_AW_PULSES, 0, 46 },
109 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 80 },
110 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 104 },
111 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 0, 108 },
112 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 132 },
113 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 168 },
114 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 64 },
115 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 80 },
116 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 104 },
117 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 108 },
118 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 132 },
119 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 168 },
120 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 176 },
121 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 208 },
122 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 256 }
126 * WMA Voice decoding context.
130 * @defgroup struct_global Global values
131 * Global values, specified in the stream header / extradata or used
135 GetBitContext gb; ///< packet bitreader. During decoder init,
136 ///< it contains the extradata from the
137 ///< demuxer. During decoding, it contains
139 int8_t vbm_tree[25]; ///< converts VLC codes to frame type
141 int spillover_bitsize; ///< number of bits used to specify
142 ///< #spillover_nbits in the packet header
143 ///< = ceil(log2(ctx->block_align << 3))
144 int history_nsamples; ///< number of samples in history for signal
145 ///< prediction (through ACB)
147 /* postfilter specific values */
148 int do_apf; ///< whether to apply the averaged
149 ///< projection filter (APF)
150 int denoise_strength; ///< strength of denoising in Wiener filter
152 int denoise_tilt_corr; ///< Whether to apply tilt correction to the
153 ///< Wiener filter coefficients (postfilter)
154 int dc_level; ///< Predicted amount of DC noise, based
155 ///< on which a DC removal filter is used
157 int lsps; ///< number of LSPs per frame [10 or 16]
158 int lsp_q_mode; ///< defines quantizer defaults [0, 1]
159 int lsp_def_mode; ///< defines different sets of LSP defaults
161 int frame_lsp_bitsize; ///< size (in bits) of LSPs, when encoded
162 ///< per-frame (independent coding)
163 int sframe_lsp_bitsize; ///< size (in bits) of LSPs, when encoded
164 ///< per superframe (residual coding)
166 int min_pitch_val; ///< base value for pitch parsing code
167 int max_pitch_val; ///< max value + 1 for pitch parsing
168 int pitch_nbits; ///< number of bits used to specify the
169 ///< pitch value in the frame header
170 int block_pitch_nbits; ///< number of bits used to specify the
171 ///< first block's pitch value
172 int block_pitch_range; ///< range of the block pitch
173 int block_delta_pitch_nbits; ///< number of bits used to specify the
174 ///< delta pitch between this and the last
175 ///< block's pitch value, used in all but
177 int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is
178 ///< from -this to +this-1)
179 uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale
184 * @defgroup struct_packet Packet values
185 * 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
215 * @defgroup struct_frame Frame and superframe values
216 * Superframe and frame data - these can change from frame to frame,
217 * although some of them do in that case serve as a cache / history for
218 * the next frame or superframe.
221 double prev_lsps[MAX_LSPS]; ///< LSPs of the last frame of the previous
223 int last_pitch_val; ///< pitch value of the previous frame
224 int last_acb_type; ///< frame type [0-2] of the previous frame
225 int pitch_diff_sh16; ///< ((cur_pitch_val - #last_pitch_val)
226 ///< << 16) / #MAX_FRAMESIZE
227 float silence_gain; ///< set for use in blocks if #ACB_TYPE_NONE
229 int aw_idx_is_ext; ///< whether the AW index was encoded in
230 ///< 8 bits (instead of 6)
231 int aw_pulse_range; ///< the range over which #aw_pulse_set1()
232 ///< can apply the pulse, relative to the
233 ///< value in aw_first_pulse_off. The exact
234 ///< position of the first AW-pulse is within
235 ///< [pulse_off, pulse_off + this], and
236 ///< depends on bitstream values; [16 or 24]
237 int aw_n_pulses[2]; ///< number of AW-pulses in each block; note
238 ///< that this number can be negative (in
239 ///< which case it basically means "zero")
240 int aw_first_pulse_off[2]; ///< index of first sample to which to
241 ///< apply AW-pulses, or -0xff if unset
242 int aw_next_pulse_off_cache; ///< the position (relative to start of the
243 ///< second block) at which pulses should
244 ///< start to be positioned, serves as a
245 ///< cache for pitch-adaptive window pulses
248 int frame_cntr; ///< current frame index [0 - 0xFFFE]; is
249 ///< only used for comfort noise in #pRNG()
250 float gain_pred_err[6]; ///< cache for gain prediction
251 float excitation_history[MAX_SIGNAL_HISTORY];
252 ///< cache of the signal of previous
253 ///< superframes, used as a history for
254 ///< signal generation
255 float synth_history[MAX_LSPS]; ///< see #excitation_history
258 * @defgroup post_filter Postfilter values
259 * Varibales used for postfilter implementation, mostly history for
260 * smoothing and so on, and context variables for FFT/iFFT.
263 RDFTContext rdft, irdft; ///< contexts for FFT-calculation in the
264 ///< postfilter (for denoise filter)
265 DCTContext dct, dst; ///< contexts for phase shift (in Hilbert
266 ///< transform, part of postfilter)
267 float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi]
269 float postfilter_agc; ///< gain control memory, used in
270 ///< #adaptive_gain_control()
271 float dcf_mem[2]; ///< DC filter history
272 float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE];
273 ///< zero filter output (i.e. excitation)
275 float denoise_filter_cache[MAX_FRAMESIZE];
276 int denoise_filter_cache_size; ///< samples in #denoise_filter_cache
277 DECLARE_ALIGNED(16, float, tilted_lpcs_pf)[0x80];
278 ///< aligned buffer for LPC tilting
279 DECLARE_ALIGNED(16, float, denoise_coeffs_pf)[0x80];
280 ///< aligned buffer for denoise coefficients
281 DECLARE_ALIGNED(16, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16];
282 ///< aligned buffer for postfilter speech
290 * Set up the variable bit mode (VBM) tree from container extradata.
291 * @param gb bit I/O context.
292 * The bit context (s->gb) should be loaded with byte 23-46 of the
293 * container extradata (i.e. the ones containing the VBM tree).
294 * @param vbm_tree pointer to array to which the decoded VBM tree will be
296 * @return 0 on success, <0 on error.
298 static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25])
300 static const uint8_t bits[] = {
303 10, 10, 10, 12, 12, 12,
306 static const uint16_t codes[] = {
307 0x0000, 0x0001, 0x0002, // 00/01/10
308 0x000c, 0x000d, 0x000e, // 11+00/01/10
309 0x003c, 0x003d, 0x003e, // 1111+00/01/10
310 0x00fc, 0x00fd, 0x00fe, // 111111+00/01/10
311 0x03fc, 0x03fd, 0x03fe, // 11111111+00/01/10
312 0x0ffc, 0x0ffd, 0x0ffe, // 1111111111+00/01/10
313 0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx
317 memset(vbm_tree, 0xff, sizeof(vbm_tree));
318 memset(cntr, 0, sizeof(cntr));
319 for (n = 0; n < 17; n++) {
320 res = get_bits(gb, 3);
321 if (cntr[res] > 3) // should be >= 3 + (res == 7))
323 vbm_tree[res * 3 + cntr[res]++] = n;
325 INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits),
326 bits, 1, 1, codes, 2, 2, 132);
331 * Set up decoder with parameters from demuxer (extradata etc.).
333 static av_cold int wmavoice_decode_init(AVCodecContext *ctx)
335 int n, flags, pitch_range, lsp16_flag;
336 WMAVoiceContext *s = ctx->priv_data;
340 * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c),
341 * - byte 19-22: flags field (annoyingly in LE; see below for known
343 * - byte 23-46: variable bitmode tree (really just 17 * 3 bits,
346 if (ctx->extradata_size != 46) {
347 av_log(ctx, AV_LOG_ERROR,
348 "Invalid extradata size %d (should be 46)\n",
349 ctx->extradata_size);
352 flags = AV_RL32(ctx->extradata + 18);
353 s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align);
354 s->do_apf = flags & 0x1;
356 ff_rdft_init(&s->rdft, 7, DFT_R2C);
357 ff_rdft_init(&s->irdft, 7, IDFT_C2R);
358 ff_dct_init(&s->dct, 6, DCT_I);
359 ff_dct_init(&s->dst, 6, DST_I);
361 ff_sine_window_init(s->cos, 256);
362 memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0]));
363 for (n = 0; n < 255; n++) {
364 s->sin[n] = -s->sin[510 - n];
365 s->cos[510 - n] = s->cos[n];
368 s->denoise_strength = (flags >> 2) & 0xF;
369 if (s->denoise_strength >= 12) {
370 av_log(ctx, AV_LOG_ERROR,
371 "Invalid denoise filter strength %d (max=11)\n",
372 s->denoise_strength);
375 s->denoise_tilt_corr = !!(flags & 0x40);
376 s->dc_level = (flags >> 7) & 0xF;
377 s->lsp_q_mode = !!(flags & 0x2000);
378 s->lsp_def_mode = !!(flags & 0x4000);
379 lsp16_flag = flags & 0x1000;
382 s->frame_lsp_bitsize = 34;
383 s->sframe_lsp_bitsize = 60;
386 s->frame_lsp_bitsize = 24;
387 s->sframe_lsp_bitsize = 48;
389 for (n = 0; n < s->lsps; n++)
390 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
392 init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3);
393 if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) {
394 av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n");
398 s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8;
399 s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8;
400 pitch_range = s->max_pitch_val - s->min_pitch_val;
401 s->pitch_nbits = av_ceil_log2(pitch_range);
402 s->last_pitch_val = 40;
403 s->last_acb_type = ACB_TYPE_NONE;
404 s->history_nsamples = s->max_pitch_val + 8;
406 if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) {
407 int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8,
408 max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8;
410 av_log(ctx, AV_LOG_ERROR,
411 "Unsupported samplerate %d (min=%d, max=%d)\n",
412 ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz
417 s->block_conv_table[0] = s->min_pitch_val;
418 s->block_conv_table[1] = (pitch_range * 25) >> 6;
419 s->block_conv_table[2] = (pitch_range * 44) >> 6;
420 s->block_conv_table[3] = s->max_pitch_val - 1;
421 s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF;
422 s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange);
423 s->block_pitch_range = s->block_conv_table[2] +
424 s->block_conv_table[3] + 1 +
425 2 * (s->block_conv_table[1] - 2 * s->min_pitch_val);
426 s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range);
428 ctx->sample_fmt = SAMPLE_FMT_FLT;
434 * @defgroup postfilter Postfilter functions
435 * Postfilter functions (gain control, wiener denoise filter, DC filter,
436 * kalman smoothening, plus surrounding code to wrap it)
440 * Adaptive gain control (as used in postfilter).
442 * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except
443 * that the energy here is calculated using sum(abs(...)), whereas the
444 * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)).
446 * @param out output buffer for filtered samples
447 * @param in input buffer containing the samples as they are after the
448 * postfilter steps so far
449 * @param speech_synth input buffer containing speech synth before postfilter
450 * @param size input buffer size
451 * @param alpha exponential filter factor
452 * @param gain_mem pointer to filter memory (single float)
454 static void adaptive_gain_control(float *out, const float *in,
455 const float *speech_synth,
456 int size, float alpha, float *gain_mem)
459 float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor;
460 float mem = *gain_mem;
462 for (i = 0; i < size; i++) {
463 speech_energy += fabsf(speech_synth[i]);
464 postfilter_energy += fabsf(in[i]);
466 gain_scale_factor = (1.0 - alpha) * speech_energy / postfilter_energy;
468 for (i = 0; i < size; i++) {
469 mem = alpha * mem + gain_scale_factor;
470 out[i] = in[i] * mem;
477 * Kalman smoothing function.
479 * This function looks back pitch +/- 3 samples back into history to find
480 * the best fitting curve (that one giving the optimal gain of the two
481 * signals, i.e. the highest dot product between the two), and then
482 * uses that signal history to smoothen the output of the speech synthesis
485 * @param s WMA Voice decoding context
486 * @param pitch pitch of the speech signal
487 * @param in input speech signal
488 * @param out output pointer for smoothened signal
489 * @param size input/output buffer size
491 * @returns -1 if no smoothening took place, e.g. because no optimal
492 * fit could be found, or 0 on success.
494 static int kalman_smoothen(WMAVoiceContext *s, int pitch,
495 const float *in, float *out, int size)
498 float optimal_gain = 0, dot;
499 const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)],
500 *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)],
503 /* find best fitting point in history */
505 dot = ff_dot_productf(in, ptr, size);
506 if (dot > optimal_gain) {
510 } while (--ptr >= end);
512 if (optimal_gain <= 0)
514 dot = ff_dot_productf(best_hist_ptr, best_hist_ptr, size);
515 if (dot <= 0) // would be 1.0
518 if (optimal_gain <= dot) {
519 dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000
523 /* actual smoothing */
524 for (n = 0; n < size; n++)
525 out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]);
531 * Get the tilt factor of a formant filter from its transfer function
532 * @see #tilt_factor() in amrnbdec.c, which does essentially the same,
533 * but somehow (??) it does a speech synthesis filter in the
534 * middle, which is missing here
536 * @param lpcs LPC coefficients
537 * @param n_lpcs Size of LPC buffer
538 * @returns the tilt factor
540 static float tilt_factor(const float *lpcs, int n_lpcs)
544 rh0 = 1.0 + ff_dot_productf(lpcs, lpcs, n_lpcs);
545 rh1 = lpcs[0] + ff_dot_productf(lpcs, &lpcs[1], n_lpcs - 1);
551 * Derive denoise filter coefficients (in real domain) from the LPCs.
553 static void calc_input_response(WMAVoiceContext *s, float *lpcs,
554 int fcb_type, float *coeffs, int remainder)
556 float last_coeff, min = 15.0, max = -15.0;
557 float irange, angle_mul, gain_mul, range, sq;
560 /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */
561 ff_rdft_calc(&s->rdft, lpcs);
562 #define log_range(var, assign) do { \
563 float tmp = log10f(assign); var = tmp; \
564 max = FFMAX(max, tmp); min = FFMIN(min, tmp); \
566 log_range(last_coeff, lpcs[1] * lpcs[1]);
567 for (n = 1; n < 64; n++)
568 log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] +
569 lpcs[n * 2 + 1] * lpcs[n * 2 + 1]);
570 log_range(lpcs[0], lpcs[0] * lpcs[0]);
573 lpcs[64] = last_coeff;
575 /* Now, use this spectrum to pick out these frequencies with higher
576 * (relative) power/energy (which we then take to be "not noise"),
577 * and set up a table (still in lpc[]) of (relative) gains per frequency.
578 * These frequencies will be maintained, while others ("noise") will be
579 * decreased in the filter output. */
580 irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63]
581 gain_mul = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) :
583 angle_mul = gain_mul * (8.0 * M_LN10 / M_PI);
584 for (n = 0; n <= 64; n++) {
587 idx = FFMAX(0, lrint((max - lpcs[n]) * irange) - 1);
588 pow = wmavoice_denoise_power_table[s->denoise_strength][idx];
589 lpcs[n] = angle_mul * pow;
591 /* 70.57 =~ 1/log10(1.0331663) */
592 idx = (pow * gain_mul - 0.0295) * 70.570526123;
593 if (idx > 127) { // fallback if index falls outside table range
594 coeffs[n] = wmavoice_energy_table[127] *
595 powf(1.0331663, idx - 127);
597 coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)];
600 /* calculate the Hilbert transform of the gains, which we do (since this
601 * is a sinus input) by doing a phase shift (in theory, H(sin())=cos()).
602 * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the
603 * "moment" of the LPCs in this filter. */
604 ff_dct_calc(&s->dct, lpcs);
605 ff_dct_calc(&s->dst, lpcs);
607 /* Split out the coefficient indexes into phase/magnitude pairs */
608 idx = 255 + av_clip(lpcs[64], -255, 255);
609 coeffs[0] = coeffs[0] * s->cos[idx];
610 idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255);
611 last_coeff = coeffs[64] * s->cos[idx];
613 idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255);
614 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
615 coeffs[n * 2] = coeffs[n] * s->cos[idx];
619 idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255);
620 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
621 coeffs[n * 2] = coeffs[n] * s->cos[idx];
623 coeffs[1] = last_coeff;
625 /* move into real domain */
626 ff_rdft_calc(&s->irdft, coeffs);
628 /* tilt correction and normalize scale */
629 memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder));
630 if (s->denoise_tilt_corr) {
633 coeffs[remainder - 1] = 0;
634 ff_tilt_compensation(&tilt_mem,
635 -1.8 * tilt_factor(coeffs, remainder - 1),
638 sq = (1.0 / 64.0) * sqrtf(1 / ff_dot_productf(coeffs, coeffs, remainder));
639 for (n = 0; n < remainder; n++)
644 * This function applies a Wiener filter on the (noisy) speech signal as
645 * a means to denoise it.
647 * - take RDFT of LPCs to get the power spectrum of the noise + speech;
648 * - using this power spectrum, calculate (for each frequency) the Wiener
649 * filter gain, which depends on the frequency power and desired level
650 * of noise subtraction (when set too high, this leads to artifacts)
651 * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse
653 * - by doing a phase shift, calculate the Hilbert transform of this array
654 * of per-frequency filter-gains to get the filtering coefficients;
655 * - smoothen/normalize/de-tilt these filter coefficients as desired;
656 * - take RDFT of noisy sound, apply the coefficients and take its IRDFT
657 * to get the denoised speech signal;
658 * - the leftover (i.e. output of the IRDFT on denoised speech data beyond
659 * the frame boundary) are saved and applied to subsequent frames by an
660 * overlap-add method (otherwise you get clicking-artifacts).
662 * @param s WMA Voice decoding context
663 * @param fcb_type Frame (codebook) type
664 * @param synth_pf input: the noisy speech signal, output: denoised speech
665 * data; should be 16-byte aligned (for ASM purposes)
666 * @param size size of the speech data
667 * @param lpcs LPCs used to synthesize this frame's speech data
669 static void wiener_denoise(WMAVoiceContext *s, int fcb_type,
670 float *synth_pf, int size,
673 int remainder, lim, n;
675 if (fcb_type != FCB_TYPE_SILENCE) {
676 float *tilted_lpcs = s->tilted_lpcs_pf,
677 *coeffs = s->denoise_coeffs_pf, tilt_mem = 0;
679 tilted_lpcs[0] = 1.0;
680 memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps);
681 memset(&tilted_lpcs[s->lsps + 1], 0,
682 sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1));
683 ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps),
684 tilted_lpcs, s->lsps + 2);
686 /* The IRDFT output (127 samples for 7-bit filter) beyond the frame
687 * size is applied to the next frame. All input beyond this is zero,
688 * and thus all output beyond this will go towards zero, hence we can
689 * limit to min(size-1, 127-size) as a performance consideration. */
690 remainder = FFMIN(127 - size, size - 1);
691 calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder);
693 /* apply coefficients (in frequency spectrum domain), i.e. complex
694 * number multiplication */
695 memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size));
696 ff_rdft_calc(&s->rdft, synth_pf);
697 ff_rdft_calc(&s->rdft, coeffs);
698 synth_pf[0] *= coeffs[0];
699 synth_pf[1] *= coeffs[1];
700 for (n = 1; n < 64; n++) {
701 float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1];
702 synth_pf[n * 2] = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1];
703 synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1];
705 ff_rdft_calc(&s->irdft, synth_pf);
708 /* merge filter output with the history of previous runs */
709 if (s->denoise_filter_cache_size) {
710 lim = FFMIN(s->denoise_filter_cache_size, size);
711 for (n = 0; n < lim; n++)
712 synth_pf[n] += s->denoise_filter_cache[n];
713 s->denoise_filter_cache_size -= lim;
714 memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size],
715 sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size);
718 /* move remainder of filter output into a cache for future runs */
719 if (fcb_type != FCB_TYPE_SILENCE) {
720 lim = FFMIN(remainder, s->denoise_filter_cache_size);
721 for (n = 0; n < lim; n++)
722 s->denoise_filter_cache[n] += synth_pf[size + n];
723 if (lim < remainder) {
724 memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim],
725 sizeof(s->denoise_filter_cache[0]) * (remainder - lim));
726 s->denoise_filter_cache_size = remainder;
732 * Averaging projection filter, the postfilter used in WMAVoice.
734 * This uses the following steps:
735 * - A zero-synthesis filter (generate excitation from synth signal)
736 * - Kalman smoothing on excitation, based on pitch
737 * - Re-synthesized smoothened output
738 * - Iterative Wiener denoise filter
739 * - Adaptive gain filter
742 * @param s WMAVoice decoding context
743 * @param synth Speech synthesis output (before postfilter)
744 * @param samples Output buffer for filtered samples
745 * @param size Buffer size of synth & samples
746 * @param lpcs Generated LPCs used for speech synthesis
747 * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned)
748 * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses)
749 * @param pitch Pitch of the input signal
751 static void postfilter(WMAVoiceContext *s, const float *synth,
752 float *samples, int size,
753 const float *lpcs, float *zero_exc_pf,
754 int fcb_type, int pitch)
756 float synth_filter_in_buf[MAX_FRAMESIZE / 2],
757 *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16],
758 *synth_filter_in = zero_exc_pf;
760 assert(size <= MAX_FRAMESIZE / 2);
762 /* generate excitation from input signal */
763 ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps);
765 if (fcb_type >= FCB_TYPE_AW_PULSES &&
766 !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size))
767 synth_filter_in = synth_filter_in_buf;
769 /* re-synthesize speech after smoothening, and keep history */
770 ff_celp_lp_synthesis_filterf(synth_pf, lpcs,
771 synth_filter_in, size, s->lsps);
772 memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps],
773 sizeof(synth_pf[0]) * s->lsps);
775 wiener_denoise(s, fcb_type, synth_pf, size, lpcs);
777 adaptive_gain_control(samples, synth_pf, synth, size, 0.99,
780 if (s->dc_level > 8) {
781 /* remove ultra-low frequency DC noise / highpass filter;
782 * coefficients are identical to those used in SIPR decoding,
783 * and very closely resemble those used in AMR-NB decoding. */
784 ff_acelp_apply_order_2_transfer_function(samples, samples,
785 (const float[2]) { -1.99997, 1.0 },
786 (const float[2]) { -1.9330735188, 0.93589198496 },
787 0.93980580475, s->dcf_mem, size);
796 * @param lsps output pointer to the array that will hold the LSPs
797 * @param num number of LSPs to be dequantized
798 * @param values quantized values, contains n_stages values
799 * @param sizes range (i.e. max value) of each quantized value
800 * @param n_stages number of dequantization runs
801 * @param table dequantization table to be used
802 * @param mul_q LSF multiplier
803 * @param base_q base (lowest) LSF values
805 static void dequant_lsps(double *lsps, int num,
806 const uint16_t *values,
807 const uint16_t *sizes,
808 int n_stages, const uint8_t *table,
810 const double *base_q)
814 memset(lsps, 0, num * sizeof(*lsps));
815 for (n = 0; n < n_stages; n++) {
816 const uint8_t *t_off = &table[values[n] * num];
817 double base = base_q[n], mul = mul_q[n];
819 for (m = 0; m < num; m++)
820 lsps[m] += base + mul * t_off[m];
822 table += sizes[n] * num;
827 * @defgroup lsp_dequant LSP dequantization routines
828 * LSP dequantization routines, for 10/16LSPs and independent/residual coding.
829 * @note we assume enough bits are available, caller should check.
830 * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;
831 * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits.
835 * Parse 10 independently-coded LSPs.
837 static void dequant_lsp10i(GetBitContext *gb, double *lsps)
839 static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 };
840 static const double mul_lsf[4] = {
841 5.2187144800e-3, 1.4626986422e-3,
842 9.6179549166e-4, 1.1325736225e-3
844 static const double base_lsf[4] = {
845 M_PI * -2.15522e-1, M_PI * -6.1646e-2,
846 M_PI * -3.3486e-2, M_PI * -5.7408e-2
850 v[0] = get_bits(gb, 8);
851 v[1] = get_bits(gb, 6);
852 v[2] = get_bits(gb, 5);
853 v[3] = get_bits(gb, 5);
855 dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i,
860 * Parse 10 independently-coded LSPs, and then derive the tables to
861 * generate LSPs for the other frames from them (residual coding).
863 static void dequant_lsp10r(GetBitContext *gb,
864 double *i_lsps, const double *old,
865 double *a1, double *a2, int q_mode)
867 static const uint16_t vec_sizes[3] = { 128, 64, 64 };
868 static const double mul_lsf[3] = {
869 2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3
871 static const double base_lsf[3] = {
872 M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2
874 const float (*ipol_tab)[2][10] = q_mode ?
875 wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a;
876 uint16_t interpol, v[3];
879 dequant_lsp10i(gb, i_lsps);
881 interpol = get_bits(gb, 5);
882 v[0] = get_bits(gb, 7);
883 v[1] = get_bits(gb, 6);
884 v[2] = get_bits(gb, 6);
886 for (n = 0; n < 10; n++) {
887 double delta = old[n] - i_lsps[n];
888 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
889 a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
892 dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r,
897 * Parse 16 independently-coded LSPs.
899 static void dequant_lsp16i(GetBitContext *gb, double *lsps)
901 static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 };
902 static const double mul_lsf[5] = {
903 3.3439586280e-3, 6.9908173703e-4,
904 3.3216608306e-3, 1.0334960326e-3,
907 static const double base_lsf[5] = {
908 M_PI * -1.27576e-1, M_PI * -2.4292e-2,
909 M_PI * -1.28094e-1, M_PI * -3.2128e-2,
914 v[0] = get_bits(gb, 8);
915 v[1] = get_bits(gb, 6);
916 v[2] = get_bits(gb, 7);
917 v[3] = get_bits(gb, 6);
918 v[4] = get_bits(gb, 7);
920 dequant_lsps( lsps, 5, v, vec_sizes, 2,
921 wmavoice_dq_lsp16i1, mul_lsf, base_lsf);
922 dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2,
923 wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]);
924 dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1,
925 wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]);
929 * Parse 16 independently-coded LSPs, and then derive the tables to
930 * generate LSPs for the other frames from them (residual coding).
932 static void dequant_lsp16r(GetBitContext *gb,
933 double *i_lsps, const double *old,
934 double *a1, double *a2, int q_mode)
936 static const uint16_t vec_sizes[3] = { 128, 128, 128 };
937 static const double mul_lsf[3] = {
938 1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3
940 static const double base_lsf[3] = {
941 M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2
943 const float (*ipol_tab)[2][16] = q_mode ?
944 wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a;
945 uint16_t interpol, v[3];
948 dequant_lsp16i(gb, i_lsps);
950 interpol = get_bits(gb, 5);
951 v[0] = get_bits(gb, 7);
952 v[1] = get_bits(gb, 7);
953 v[2] = get_bits(gb, 7);
955 for (n = 0; n < 16; n++) {
956 double delta = old[n] - i_lsps[n];
957 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
958 a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
961 dequant_lsps( a2, 10, v, vec_sizes, 1,
962 wmavoice_dq_lsp16r1, mul_lsf, base_lsf);
963 dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1,
964 wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]);
965 dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1,
966 wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]);
971 * @defgroup aw Pitch-adaptive window coding functions
972 * The next few functions are for pitch-adaptive window coding.
976 * Parse the offset of the first pitch-adaptive window pulses, and
977 * the distribution of pulses between the two blocks in this frame.
978 * @param s WMA Voice decoding context private data
979 * @param gb bit I/O context
980 * @param pitch pitch for each block in this frame
982 static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb,
985 static const int16_t start_offset[94] = {
986 -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11,
987 13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26,
988 27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43,
989 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67,
990 69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91,
991 93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115,
992 117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139,
993 141, 143, 145, 147, 149, 151, 153, 155, 157, 159
997 /* position of pulse */
998 s->aw_idx_is_ext = 0;
999 if ((bits = get_bits(gb, 6)) >= 54) {
1000 s->aw_idx_is_ext = 1;
1001 bits += (bits - 54) * 3 + get_bits(gb, 2);
1004 /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count
1005 * the distribution of the pulses in each block contained in this frame. */
1006 s->aw_pulse_range = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16;
1007 for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ;
1008 s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0];
1009 s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2;
1010 offset += s->aw_n_pulses[0] * pitch[0];
1011 s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1];
1012 s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2;
1014 /* if continuing from a position before the block, reset position to
1015 * start of block (when corrected for the range over which it can be
1016 * spread in aw_pulse_set1()). */
1017 if (start_offset[bits] < MAX_FRAMESIZE / 2) {
1018 while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0)
1019 s->aw_first_pulse_off[1] -= pitch[1];
1020 if (start_offset[bits] < 0)
1021 while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0)
1022 s->aw_first_pulse_off[0] -= pitch[0];
1027 * Apply second set of pitch-adaptive window pulses.
1028 * @param s WMA Voice decoding context private data
1029 * @param gb bit I/O context
1030 * @param block_idx block index in frame [0, 1]
1031 * @param fcb structure containing fixed codebook vector info
1033 static void aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb,
1034 int block_idx, AMRFixed *fcb)
1036 uint16_t use_mask[7]; // only 5 are used, rest is padding
1037 /* in this function, idx is the index in the 80-bit (+ padding) use_mask
1038 * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits
1039 * of idx are the position of the bit within a particular item in the
1040 * array (0 being the most significant bit, and 15 being the least
1041 * significant bit), and the remainder (>> 4) is the index in the
1042 * use_mask[]-array. This is faster and uses less memory than using a
1043 * 80-byte/80-int array. */
1044 int pulse_off = s->aw_first_pulse_off[block_idx],
1045 pulse_start, n, idx, range, aidx, start_off = 0;
1047 /* set offset of first pulse to within this block */
1048 if (s->aw_n_pulses[block_idx] > 0)
1049 while (pulse_off + s->aw_pulse_range < 1)
1050 pulse_off += fcb->pitch_lag;
1052 /* find range per pulse */
1053 if (s->aw_n_pulses[0] > 0) {
1054 if (block_idx == 0) {
1056 } else /* block_idx = 1 */ {
1058 if (s->aw_n_pulses[block_idx] > 0)
1059 pulse_off = s->aw_next_pulse_off_cache;
1063 pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0;
1065 /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,
1066 * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus
1067 * we exclude that range from being pulsed again in this function. */
1068 memset( use_mask, -1, 5 * sizeof(use_mask[0]));
1069 memset(&use_mask[5], 0, 2 * sizeof(use_mask[0]));
1070 if (s->aw_n_pulses[block_idx] > 0)
1071 for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) {
1072 int excl_range = s->aw_pulse_range; // always 16 or 24
1073 uint16_t *use_mask_ptr = &use_mask[idx >> 4];
1074 int first_sh = 16 - (idx & 15);
1075 *use_mask_ptr++ &= 0xFFFF << first_sh;
1076 excl_range -= first_sh;
1077 if (excl_range >= 16) {
1078 *use_mask_ptr++ = 0;
1079 *use_mask_ptr &= 0xFFFF >> (excl_range - 16);
1081 *use_mask_ptr &= 0xFFFF >> excl_range;
1084 /* find the 'aidx'th offset that is not excluded */
1085 aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4);
1086 for (n = 0; n <= aidx; pulse_start++) {
1087 for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ;
1088 if (idx >= MAX_FRAMESIZE / 2) { // find from zero
1089 if (use_mask[0]) idx = 0x0F;
1090 else if (use_mask[1]) idx = 0x1F;
1091 else if (use_mask[2]) idx = 0x2F;
1092 else if (use_mask[3]) idx = 0x3F;
1093 else if (use_mask[4]) idx = 0x4F;
1095 idx -= av_log2_16bit(use_mask[idx >> 4]);
1097 if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) {
1098 use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15));
1104 fcb->x[fcb->n] = start_off;
1105 fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0;
1108 /* set offset for next block, relative to start of that block */
1109 n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag;
1110 s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0;
1114 * Apply first set of pitch-adaptive window pulses.
1115 * @param s WMA Voice decoding context private data
1116 * @param gb bit I/O context
1117 * @param block_idx block index in frame [0, 1]
1118 * @param fcb storage location for fixed codebook pulse info
1120 static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb,
1121 int block_idx, AMRFixed *fcb)
1123 int val = get_bits(gb, 12 - 2 * (s->aw_idx_is_ext && !block_idx));
1126 if (s->aw_n_pulses[block_idx] > 0) {
1127 int n, v_mask, i_mask, sh, n_pulses;
1129 if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each
1134 } else { // 4 pulses, 1:sign + 2:index each
1141 for (n = n_pulses - 1; n >= 0; n--, val >>= sh) {
1142 fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0;
1143 fcb->x[fcb->n] = (val & i_mask) * n_pulses + n +
1144 s->aw_first_pulse_off[block_idx];
1145 while (fcb->x[fcb->n] < 0)
1146 fcb->x[fcb->n] += fcb->pitch_lag;
1147 if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2)
1151 int num2 = (val & 0x1FF) >> 1, delta, idx;
1153 if (num2 < 1 * 79) { delta = 1; idx = num2 + 1; }
1154 else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; }
1155 else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; }
1156 else { delta = 7; idx = num2 + 1 - 3 * 75; }
1157 v = (val & 0x200) ? -1.0 : 1.0;
1159 fcb->no_repeat_mask |= 3 << fcb->n;
1160 fcb->x[fcb->n] = idx - delta;
1162 fcb->x[fcb->n + 1] = idx;
1163 fcb->y[fcb->n + 1] = (val & 1) ? -v : v;
1171 * Generate a random number from frame_cntr and block_idx, which will lief
1172 * in the range [0, 1000 - block_size] (so it can be used as an index in a
1173 * table of size 1000 of which you want to read block_size entries).
1175 * @param frame_cntr current frame number
1176 * @param block_num current block index
1177 * @param block_size amount of entries we want to read from a table
1178 * that has 1000 entries
1179 * @return a (non-)random number in the [0, 1000 - block_size] range.
1181 static int pRNG(int frame_cntr, int block_num, int block_size)
1183 /* array to simplify the calculation of z:
1184 * y = (x % 9) * 5 + 6;
1185 * z = (49995 * x) / y;
1186 * Since y only has 9 values, we can remove the division by using a
1187 * LUT and using FASTDIV-style divisions. For each of the 9 values
1188 * of y, we can rewrite z as:
1189 * z = x * (49995 / y) + x * ((49995 % y) / y)
1190 * In this table, each col represents one possible value of y, the
1191 * first number is 49995 / y, and the second is the FASTDIV variant
1192 * of 49995 % y / y. */
1193 static const unsigned int div_tbl[9][2] = {
1194 { 8332, 3 * 715827883U }, // y = 6
1195 { 4545, 0 * 390451573U }, // y = 11
1196 { 3124, 11 * 268435456U }, // y = 16
1197 { 2380, 15 * 204522253U }, // y = 21
1198 { 1922, 23 * 165191050U }, // y = 26
1199 { 1612, 23 * 138547333U }, // y = 31
1200 { 1388, 27 * 119304648U }, // y = 36
1201 { 1219, 16 * 104755300U }, // y = 41
1202 { 1086, 39 * 93368855U } // y = 46
1204 unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr;
1205 if (x >= 0xFFFF) x -= 0xFFFF; // max value of x is 8*1877+0xFFFE=0x13AA6,
1206 // so this is effectively a modulo (%)
1207 y = x - 9 * MULH(477218589, x); // x % 9
1208 z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1]));
1209 // z = x * 49995 / (y * 5 + 6)
1210 return z % (1000 - block_size);
1214 * Parse hardcoded signal for a single block.
1215 * @note see #synth_block().
1217 static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb,
1218 int block_idx, int size,
1219 const struct frame_type_desc *frame_desc,
1225 assert(size <= MAX_FRAMESIZE);
1227 /* Set the offset from which we start reading wmavoice_std_codebook */
1228 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1229 r_idx = pRNG(s->frame_cntr, block_idx, size);
1230 gain = s->silence_gain;
1231 } else /* FCB_TYPE_HARDCODED */ {
1232 r_idx = get_bits(gb, 8);
1233 gain = wmavoice_gain_universal[get_bits(gb, 6)];
1236 /* Clear gain prediction parameters */
1237 memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err));
1239 /* Apply gain to hardcoded codebook and use that as excitation signal */
1240 for (n = 0; n < size; n++)
1241 excitation[n] = wmavoice_std_codebook[r_idx + n] * gain;
1245 * Parse FCB/ACB signal for a single block.
1246 * @note see #synth_block().
1248 static void synth_block_fcb_acb(WMAVoiceContext *s, GetBitContext *gb,
1249 int block_idx, int size,
1250 int block_pitch_sh2,
1251 const struct frame_type_desc *frame_desc,
1254 static const float gain_coeff[6] = {
1255 0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458
1257 float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain;
1258 int n, idx, gain_weight;
1261 assert(size <= MAX_FRAMESIZE / 2);
1262 memset(pulses, 0, sizeof(*pulses) * size);
1264 fcb.pitch_lag = block_pitch_sh2 >> 2;
1265 fcb.pitch_fac = 1.0;
1266 fcb.no_repeat_mask = 0;
1269 /* For the other frame types, this is where we apply the innovation
1270 * (fixed) codebook pulses of the speech signal. */
1271 if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1272 aw_pulse_set1(s, gb, block_idx, &fcb);
1273 aw_pulse_set2(s, gb, block_idx, &fcb);
1274 } else /* FCB_TYPE_EXC_PULSES */ {
1275 int offset_nbits = 5 - frame_desc->log_n_blocks;
1277 fcb.no_repeat_mask = -1;
1278 /* similar to ff_decode_10_pulses_35bits(), but with single pulses
1279 * (instead of double) for a subset of pulses */
1280 for (n = 0; n < 5; n++) {
1284 sign = get_bits1(gb) ? 1.0 : -1.0;
1285 pos1 = get_bits(gb, offset_nbits);
1286 fcb.x[fcb.n] = n + 5 * pos1;
1287 fcb.y[fcb.n++] = sign;
1288 if (n < frame_desc->dbl_pulses) {
1289 pos2 = get_bits(gb, offset_nbits);
1290 fcb.x[fcb.n] = n + 5 * pos2;
1291 fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign;
1295 ff_set_fixed_vector(pulses, &fcb, 1.0, size);
1297 /* Calculate gain for adaptive & fixed codebook signal.
1298 * see ff_amr_set_fixed_gain(). */
1299 idx = get_bits(gb, 7);
1300 fcb_gain = expf(ff_dot_productf(s->gain_pred_err, gain_coeff, 6) -
1301 5.2409161640 + wmavoice_gain_codebook_fcb[idx]);
1302 acb_gain = wmavoice_gain_codebook_acb[idx];
1303 pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx],
1304 -2.9957322736 /* log(0.05) */,
1305 1.6094379124 /* log(5.0) */);
1307 gain_weight = 8 >> frame_desc->log_n_blocks;
1308 memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err,
1309 sizeof(*s->gain_pred_err) * (6 - gain_weight));
1310 for (n = 0; n < gain_weight; n++)
1311 s->gain_pred_err[n] = pred_err;
1313 /* Calculation of adaptive codebook */
1314 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1316 for (n = 0; n < size; n += len) {
1318 int abs_idx = block_idx * size + n;
1319 int pitch_sh16 = (s->last_pitch_val << 16) +
1320 s->pitch_diff_sh16 * abs_idx;
1321 int pitch = (pitch_sh16 + 0x6FFF) >> 16;
1322 int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000;
1323 idx = idx_sh16 >> 16;
1324 if (s->pitch_diff_sh16) {
1325 if (s->pitch_diff_sh16 > 0) {
1326 next_idx_sh16 = (idx_sh16) &~ 0xFFFF;
1328 next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF;
1329 len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8,
1334 ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch],
1335 wmavoice_ipol1_coeffs, 17,
1338 } else /* ACB_TYPE_HAMMING */ {
1339 int block_pitch = block_pitch_sh2 >> 2;
1340 idx = block_pitch_sh2 & 3;
1342 ff_acelp_interpolatef(excitation, &excitation[-block_pitch],
1343 wmavoice_ipol2_coeffs, 4,
1346 av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch,
1347 sizeof(float) * size);
1350 /* Interpolate ACB/FCB and use as excitation signal */
1351 ff_weighted_vector_sumf(excitation, excitation, pulses,
1352 acb_gain, fcb_gain, size);
1356 * Parse data in a single block.
1357 * @note we assume enough bits are available, caller should check.
1359 * @param s WMA Voice decoding context private data
1360 * @param gb bit I/O context
1361 * @param block_idx index of the to-be-read block
1362 * @param size amount of samples to be read in this block
1363 * @param block_pitch_sh2 pitch for this block << 2
1364 * @param lsps LSPs for (the end of) this frame
1365 * @param prev_lsps LSPs for the last frame
1366 * @param frame_desc frame type descriptor
1367 * @param excitation target memory for the ACB+FCB interpolated signal
1368 * @param synth target memory for the speech synthesis filter output
1369 * @return 0 on success, <0 on error.
1371 static void synth_block(WMAVoiceContext *s, GetBitContext *gb,
1372 int block_idx, int size,
1373 int block_pitch_sh2,
1374 const double *lsps, const double *prev_lsps,
1375 const struct frame_type_desc *frame_desc,
1376 float *excitation, float *synth)
1378 double i_lsps[MAX_LSPS];
1379 float lpcs[MAX_LSPS];
1383 if (frame_desc->acb_type == ACB_TYPE_NONE)
1384 synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation);
1386 synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2,
1387 frame_desc, excitation);
1389 /* convert interpolated LSPs to LPCs */
1390 fac = (block_idx + 0.5) / frame_desc->n_blocks;
1391 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1392 i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n]));
1393 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1395 /* Speech synthesis */
1396 ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps);
1400 * Synthesize output samples for a single frame.
1401 * @note we assume enough bits are available, caller should check.
1403 * @param ctx WMA Voice decoder context
1404 * @param gb bit I/O context (s->gb or one for cross-packet superframes)
1405 * @param frame_idx Frame number within superframe [0-2]
1406 * @param samples pointer to output sample buffer, has space for at least 160
1408 * @param lsps LSP array
1409 * @param prev_lsps array of previous frame's LSPs
1410 * @param excitation target buffer for excitation signal
1411 * @param synth target buffer for synthesized speech data
1412 * @return 0 on success, <0 on error.
1414 static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx,
1416 const double *lsps, const double *prev_lsps,
1417 float *excitation, float *synth)
1419 WMAVoiceContext *s = ctx->priv_data;
1420 int n, n_blocks_x2, log_n_blocks_x2, cur_pitch_val;
1421 int pitch[MAX_BLOCKS], last_block_pitch;
1423 /* Parse frame type ("frame header"), see frame_descs */
1424 int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)],
1425 block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks;
1428 av_log(ctx, AV_LOG_ERROR,
1429 "Invalid frame type VLC code, skipping\n");
1433 /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */
1434 if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) {
1435 /* Pitch is provided per frame, which is interpreted as the pitch of
1436 * the last sample of the last block of this frame. We can interpolate
1437 * the pitch of other blocks (and even pitch-per-sample) by gradually
1438 * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */
1439 n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1;
1440 log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1;
1441 cur_pitch_val = s->min_pitch_val + get_bits(gb, s->pitch_nbits);
1442 cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1);
1443 if (s->last_acb_type == ACB_TYPE_NONE ||
1444 20 * abs(cur_pitch_val - s->last_pitch_val) >
1445 (cur_pitch_val + s->last_pitch_val))
1446 s->last_pitch_val = cur_pitch_val;
1448 /* pitch per block */
1449 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1450 int fac = n * 2 + 1;
1452 pitch[n] = (MUL16(fac, cur_pitch_val) +
1453 MUL16((n_blocks_x2 - fac), s->last_pitch_val) +
1454 frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2;
1457 /* "pitch-diff-per-sample" for calculation of pitch per sample */
1458 s->pitch_diff_sh16 =
1459 ((cur_pitch_val - s->last_pitch_val) << 16) / MAX_FRAMESIZE;
1462 /* Global gain (if silence) and pitch-adaptive window coordinates */
1463 switch (frame_descs[bd_idx].fcb_type) {
1464 case FCB_TYPE_SILENCE:
1465 s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)];
1467 case FCB_TYPE_AW_PULSES:
1468 aw_parse_coords(s, gb, pitch);
1472 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1475 /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */
1476 switch (frame_descs[bd_idx].acb_type) {
1477 case ACB_TYPE_HAMMING: {
1478 /* Pitch is given per block. Per-block pitches are encoded as an
1479 * absolute value for the first block, and then delta values
1480 * relative to this value) for all subsequent blocks. The scale of
1481 * this pitch value is semi-logaritmic compared to its use in the
1482 * decoder, so we convert it to normal scale also. */
1484 t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2,
1485 t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1,
1486 t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1;
1489 block_pitch = get_bits(gb, s->block_pitch_nbits);
1491 block_pitch = last_block_pitch - s->block_delta_pitch_hrange +
1492 get_bits(gb, s->block_delta_pitch_nbits);
1493 /* Convert last_ so that any next delta is within _range */
1494 last_block_pitch = av_clip(block_pitch,
1495 s->block_delta_pitch_hrange,
1496 s->block_pitch_range -
1497 s->block_delta_pitch_hrange);
1499 /* Convert semi-log-style scale back to normal scale */
1500 if (block_pitch < t1) {
1501 bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch;
1504 if (block_pitch < t2) {
1506 (s->block_conv_table[1] << 2) + (block_pitch << 1);
1509 if (block_pitch < t3) {
1511 (s->block_conv_table[2] + block_pitch) << 2;
1513 bl_pitch_sh2 = s->block_conv_table[3] << 2;
1516 pitch[n] = bl_pitch_sh2 >> 2;
1520 case ACB_TYPE_ASYMMETRIC: {
1521 bl_pitch_sh2 = pitch[n] << 2;
1525 default: // ACB_TYPE_NONE has no pitch
1530 synth_block(s, gb, n, block_nsamples, bl_pitch_sh2,
1531 lsps, prev_lsps, &frame_descs[bd_idx],
1532 &excitation[n * block_nsamples],
1533 &synth[n * block_nsamples]);
1536 /* Averaging projection filter, if applicable. Else, just copy samples
1537 * from synthesis buffer */
1539 double i_lsps[MAX_LSPS];
1540 float lpcs[MAX_LSPS];
1542 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1543 i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n]));
1544 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1545 postfilter(s, synth, samples, 80, lpcs,
1546 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx],
1547 frame_descs[bd_idx].fcb_type, pitch[0]);
1549 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1550 i_lsps[n] = cos(lsps[n]);
1551 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1552 postfilter(s, &synth[80], &samples[80], 80, lpcs,
1553 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80],
1554 frame_descs[bd_idx].fcb_type, pitch[0]);
1556 memcpy(samples, synth, 160 * sizeof(synth[0]));
1558 /* Cache values for next frame */
1560 if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%)
1561 s->last_acb_type = frame_descs[bd_idx].acb_type;
1562 switch (frame_descs[bd_idx].acb_type) {
1564 s->last_pitch_val = 0;
1566 case ACB_TYPE_ASYMMETRIC:
1567 s->last_pitch_val = cur_pitch_val;
1569 case ACB_TYPE_HAMMING:
1570 s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1];
1578 * Ensure minimum value for first item, maximum value for last value,
1579 * proper spacing between each value and proper ordering.
1581 * @param lsps array of LSPs
1582 * @param num size of LSP array
1584 * @note basically a double version of #ff_acelp_reorder_lsf(), might be
1585 * useful to put in a generic location later on. Parts are also
1586 * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(),
1587 * which is in float.
1589 static void stabilize_lsps(double *lsps, int num)
1593 /* set minimum value for first, maximum value for last and minimum
1594 * spacing between LSF values.
1595 * Very similar to ff_set_min_dist_lsf(), but in double. */
1596 lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI);
1597 for (n = 1; n < num; n++)
1598 lsps[n] = FFMAX(lsps[n], lsps[n - 1] + 0.0125 * M_PI);
1599 lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI);
1601 /* reorder (looks like one-time / non-recursed bubblesort).
1602 * Very similar to ff_sort_nearly_sorted_floats(), but in double. */
1603 for (n = 1; n < num; n++) {
1604 if (lsps[n] < lsps[n - 1]) {
1605 for (m = 1; m < num; m++) {
1606 double tmp = lsps[m];
1607 for (l = m - 1; l >= 0; l--) {
1608 if (lsps[l] <= tmp) break;
1609 lsps[l + 1] = lsps[l];
1619 * Test if there's enough bits to read 1 superframe.
1621 * @param orig_gb bit I/O context used for reading. This function
1622 * does not modify the state of the bitreader; it
1623 * only uses it to copy the current stream position
1624 * @param s WMA Voice decoding context private data
1625 * @return -1 if unsupported, 1 on not enough bits or 0 if OK.
1627 static int check_bits_for_superframe(GetBitContext *orig_gb,
1630 GetBitContext s_gb, *gb = &s_gb;
1631 int n, need_bits, bd_idx;
1632 const struct frame_type_desc *frame_desc;
1634 /* initialize a copy */
1635 init_get_bits(gb, orig_gb->buffer, orig_gb->size_in_bits);
1636 skip_bits_long(gb, get_bits_count(orig_gb));
1637 assert(get_bits_left(gb) == get_bits_left(orig_gb));
1639 /* superframe header */
1640 if (get_bits_left(gb) < 14)
1643 return -1; // WMAPro-in-WMAVoice superframe
1644 if (get_bits1(gb)) skip_bits(gb, 12); // number of samples in superframe
1645 if (s->has_residual_lsps) { // residual LSPs (for all frames)
1646 if (get_bits_left(gb) < s->sframe_lsp_bitsize)
1648 skip_bits_long(gb, s->sframe_lsp_bitsize);
1652 for (n = 0; n < MAX_FRAMES; n++) {
1653 int aw_idx_is_ext = 0;
1655 if (!s->has_residual_lsps) { // independent LSPs (per-frame)
1656 if (get_bits_left(gb) < s->frame_lsp_bitsize) return 1;
1657 skip_bits_long(gb, s->frame_lsp_bitsize);
1659 bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)];
1661 return -1; // invalid frame type VLC code
1662 frame_desc = &frame_descs[bd_idx];
1663 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1664 if (get_bits_left(gb) < s->pitch_nbits)
1666 skip_bits_long(gb, s->pitch_nbits);
1668 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1670 } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1671 int tmp = get_bits(gb, 6);
1679 if (frame_desc->acb_type == ACB_TYPE_HAMMING) {
1680 need_bits = s->block_pitch_nbits +
1681 (frame_desc->n_blocks - 1) * s->block_delta_pitch_nbits;
1682 } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1683 need_bits = 2 * !aw_idx_is_ext;
1686 need_bits += frame_desc->frame_size;
1687 if (get_bits_left(gb) < need_bits)
1689 skip_bits_long(gb, need_bits);
1696 * Synthesize output samples for a single superframe. If we have any data
1697 * cached in s->sframe_cache, that will be used instead of whatever is loaded
1700 * WMA Voice superframes contain 3 frames, each containing 160 audio samples,
1701 * to give a total of 480 samples per frame. See #synth_frame() for frame
1702 * parsing. In addition to 3 frames, superframes can also contain the LSPs
1703 * (if these are globally specified for all frames (residually); they can
1704 * also be specified individually per-frame. See the s->has_residual_lsps
1705 * option), and can specify the number of samples encoded in this superframe
1706 * (if less than 480), usually used to prevent blanks at track boundaries.
1708 * @param ctx WMA Voice decoder context
1709 * @param samples pointer to output buffer for voice samples
1710 * @param data_size pointer containing the size of #samples on input, and the
1711 * amount of #samples filled on output
1712 * @return 0 on success, <0 on error or 1 if there was not enough data to
1713 * fully parse the superframe
1715 static int synth_superframe(AVCodecContext *ctx,
1716 float *samples, int *data_size)
1718 WMAVoiceContext *s = ctx->priv_data;
1719 GetBitContext *gb = &s->gb, s_gb;
1720 int n, res, n_samples = 480;
1721 double lsps[MAX_FRAMES][MAX_LSPS];
1722 const double *mean_lsf = s->lsps == 16 ?
1723 wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode];
1724 float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12];
1725 float synth[MAX_LSPS + MAX_SFRAMESIZE];
1727 memcpy(synth, s->synth_history,
1728 s->lsps * sizeof(*synth));
1729 memcpy(excitation, s->excitation_history,
1730 s->history_nsamples * sizeof(*excitation));
1732 if (s->sframe_cache_size > 0) {
1734 init_get_bits(gb, s->sframe_cache, s->sframe_cache_size);
1735 s->sframe_cache_size = 0;
1738 if ((res = check_bits_for_superframe(gb, s)) == 1) return 1;
1740 /* First bit is speech/music bit, it differentiates between WMAVoice
1741 * speech samples (the actual codec) and WMAVoice music samples, which
1742 * are really WMAPro-in-WMAVoice-superframes. I've never seen those in
1744 if (!get_bits1(gb)) {
1745 av_log_missing_feature(ctx, "WMAPro-in-WMAVoice support", 1);
1749 /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */
1750 if (get_bits1(gb)) {
1751 if ((n_samples = get_bits(gb, 12)) > 480) {
1752 av_log(ctx, AV_LOG_ERROR,
1753 "Superframe encodes >480 samples (%d), not allowed\n",
1758 /* Parse LSPs, if global for the superframe (can also be per-frame). */
1759 if (s->has_residual_lsps) {
1760 double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2];
1762 for (n = 0; n < s->lsps; n++)
1763 prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n];
1765 if (s->lsps == 10) {
1766 dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1767 } else /* s->lsps == 16 */
1768 dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1770 for (n = 0; n < s->lsps; n++) {
1771 lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]);
1772 lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]);
1773 lsps[2][n] += mean_lsf[n];
1775 for (n = 0; n < 3; n++)
1776 stabilize_lsps(lsps[n], s->lsps);
1779 /* Parse frames, optionally preceeded by per-frame (independent) LSPs. */
1780 for (n = 0; n < 3; n++) {
1781 if (!s->has_residual_lsps) {
1784 if (s->lsps == 10) {
1785 dequant_lsp10i(gb, lsps[n]);
1786 } else /* s->lsps == 16 */
1787 dequant_lsp16i(gb, lsps[n]);
1789 for (m = 0; m < s->lsps; m++)
1790 lsps[n][m] += mean_lsf[m];
1791 stabilize_lsps(lsps[n], s->lsps);
1794 if ((res = synth_frame(ctx, gb, n,
1795 &samples[n * MAX_FRAMESIZE],
1796 lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1],
1797 &excitation[s->history_nsamples + n * MAX_FRAMESIZE],
1798 &synth[s->lsps + n * MAX_FRAMESIZE])))
1802 /* Statistics? FIXME - we don't check for length, a slight overrun
1803 * will be caught by internal buffer padding, and anything else
1804 * will be skipped, not read. */
1805 if (get_bits1(gb)) {
1806 res = get_bits(gb, 4);
1807 skip_bits(gb, 10 * (res + 1));
1810 /* Specify nr. of output samples */
1811 *data_size = n_samples * sizeof(float);
1813 /* Update history */
1814 memcpy(s->prev_lsps, lsps[2],
1815 s->lsps * sizeof(*s->prev_lsps));
1816 memcpy(s->synth_history, &synth[MAX_SFRAMESIZE],
1817 s->lsps * sizeof(*synth));
1818 memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE],
1819 s->history_nsamples * sizeof(*excitation));
1821 memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE],
1822 s->history_nsamples * sizeof(*s->zero_exc_pf));
1828 * Parse the packet header at the start of each packet (input data to this
1831 * @param s WMA Voice decoding context private data
1832 * @return 1 if not enough bits were available, or 0 on success.
1834 static int parse_packet_header(WMAVoiceContext *s)
1836 GetBitContext *gb = &s->gb;
1839 if (get_bits_left(gb) < 11)
1841 skip_bits(gb, 4); // packet sequence number
1842 s->has_residual_lsps = get_bits1(gb);
1844 res = get_bits(gb, 6); // number of superframes per packet
1845 // (minus first one if there is spillover)
1846 if (get_bits_left(gb) < 6 * (res == 0x3F) + s->spillover_bitsize)
1848 } while (res == 0x3F);
1849 s->spillover_nbits = get_bits(gb, s->spillover_bitsize);
1855 * Copy (unaligned) bits from gb/data/size to pb.
1857 * @param pb target buffer to copy bits into
1858 * @param data source buffer to copy bits from
1859 * @param size size of the source data, in bytes
1860 * @param gb bit I/O context specifying the current position in the source.
1861 * data. This function might use this to align the bit position to
1862 * a whole-byte boundary before calling #ff_copy_bits() on aligned
1864 * @param nbits the amount of bits to copy from source to target
1866 * @note after calling this function, the current position in the input bit
1867 * I/O context is undefined.
1869 static void copy_bits(PutBitContext *pb,
1870 const uint8_t *data, int size,
1871 GetBitContext *gb, int nbits)
1873 int rmn_bytes, rmn_bits;
1875 rmn_bits = rmn_bytes = get_bits_left(gb);
1876 if (rmn_bits < nbits)
1878 rmn_bits &= 7; rmn_bytes >>= 3;
1879 if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0)
1880 put_bits(pb, rmn_bits, get_bits(gb, rmn_bits));
1881 ff_copy_bits(pb, data + size - rmn_bytes,
1882 FFMIN(nbits - rmn_bits, rmn_bytes << 3));
1886 * Packet decoding: a packet is anything that the (ASF) demuxer contains,
1887 * and we expect that the demuxer / application provides it to us as such
1888 * (else you'll probably get garbage as output). Every packet has a size of
1889 * ctx->block_align bytes, starts with a packet header (see
1890 * #parse_packet_header()), and then a series of superframes. Superframe
1891 * boundaries may exceed packets, i.e. superframes can split data over
1892 * multiple (two) packets.
1894 * For more information about frames, see #synth_superframe().
1896 static int wmavoice_decode_packet(AVCodecContext *ctx, void *data,
1897 int *data_size, AVPacket *avpkt)
1899 WMAVoiceContext *s = ctx->priv_data;
1900 GetBitContext *gb = &s->gb;
1903 if (*data_size < 480 * sizeof(float)) {
1904 av_log(ctx, AV_LOG_ERROR,
1905 "Output buffer too small (%d given - %zu needed)\n",
1906 *data_size, 480 * sizeof(float));
1911 /* Packets are sometimes a multiple of ctx->block_align, with a packet
1912 * header at each ctx->block_align bytes. However, FFmpeg's ASF demuxer
1913 * feeds us ASF packets, which may concatenate multiple "codec" packets
1914 * in a single "muxer" packet, so we artificially emulate that by
1915 * capping the packet size at ctx->block_align. */
1916 for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align);
1919 init_get_bits(&s->gb, avpkt->data, size << 3);
1921 /* size == ctx->block_align is used to indicate whether we are dealing with
1922 * a new packet or a packet of which we already read the packet header
1924 if (size == ctx->block_align) { // new packet header
1925 if ((res = parse_packet_header(s)) < 0)
1928 /* If the packet header specifies a s->spillover_nbits, then we want
1929 * to push out all data of the previous packet (+ spillover) before
1930 * continuing to parse new superframes in the current packet. */
1931 if (s->spillover_nbits > 0) {
1932 if (s->sframe_cache_size > 0) {
1933 int cnt = get_bits_count(gb);
1934 copy_bits(&s->pb, avpkt->data, size, gb, s->spillover_nbits);
1935 flush_put_bits(&s->pb);
1936 s->sframe_cache_size += s->spillover_nbits;
1937 if ((res = synth_superframe(ctx, data, data_size)) == 0 &&
1939 cnt += s->spillover_nbits;
1940 s->skip_bits_next = cnt & 7;
1943 skip_bits_long (gb, s->spillover_nbits - cnt +
1944 get_bits_count(gb)); // resync
1946 skip_bits_long(gb, s->spillover_nbits); // resync
1948 } else if (s->skip_bits_next)
1949 skip_bits(gb, s->skip_bits_next);
1951 /* Try parsing superframes in current packet */
1952 s->sframe_cache_size = 0;
1953 s->skip_bits_next = 0;
1954 pos = get_bits_left(gb);
1955 if ((res = synth_superframe(ctx, data, data_size)) < 0) {
1957 } else if (*data_size > 0) {
1958 int cnt = get_bits_count(gb);
1959 s->skip_bits_next = cnt & 7;
1961 } else if ((s->sframe_cache_size = pos) > 0) {
1962 /* rewind bit reader to start of last (incomplete) superframe... */
1963 init_get_bits(gb, avpkt->data, size << 3);
1964 skip_bits_long(gb, (size << 3) - pos);
1965 assert(get_bits_left(gb) == pos);
1967 /* ...and cache it for spillover in next packet */
1968 init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE);
1969 copy_bits(&s->pb, avpkt->data, size, gb, s->sframe_cache_size);
1970 // FIXME bad - just copy bytes as whole and add use the
1971 // skip_bits_next field
1977 static av_cold int wmavoice_decode_end(AVCodecContext *ctx)
1979 WMAVoiceContext *s = ctx->priv_data;
1982 ff_rdft_end(&s->rdft);
1983 ff_rdft_end(&s->irdft);
1984 ff_dct_end(&s->dct);
1985 ff_dct_end(&s->dst);
1991 static av_cold void wmavoice_flush(AVCodecContext *ctx)
1993 WMAVoiceContext *s = ctx->priv_data;
1996 s->postfilter_agc = 0;
1997 s->sframe_cache_size = 0;
1998 s->skip_bits_next = 0;
1999 for (n = 0; n < s->lsps; n++)
2000 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
2001 memset(s->excitation_history, 0,
2002 sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY);
2003 memset(s->synth_history, 0,
2004 sizeof(*s->synth_history) * MAX_LSPS);
2005 memset(s->gain_pred_err, 0,
2006 sizeof(s->gain_pred_err));
2009 memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0,
2010 sizeof(*s->synth_filter_out_buf) * s->lsps);
2011 memset(s->dcf_mem, 0,
2012 sizeof(*s->dcf_mem) * 2);
2013 memset(s->zero_exc_pf, 0,
2014 sizeof(*s->zero_exc_pf) * s->history_nsamples);
2015 memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache));
2019 AVCodec wmavoice_decoder = {
2023 sizeof(WMAVoiceContext),
2024 wmavoice_decode_init,
2026 wmavoice_decode_end,
2027 wmavoice_decode_packet,
2028 CODEC_CAP_SUBFRAMES,
2029 .flush = wmavoice_flush,
2030 .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"),