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>
30 #include "libavutil/channel_layout.h"
31 #include "libavutil/float_dsp.h"
32 #include "libavutil/mem.h"
33 #include "libavutil/thread.h"
38 #include "wmavoice_data.h"
39 #include "celp_filters.h"
40 #include "acelp_vectors.h"
41 #include "acelp_filters.h"
47 #define MAX_BLOCKS 8 ///< maximum number of blocks per frame
48 #define MAX_LSPS 16 ///< maximum filter order
49 #define MAX_LSPS_ALIGN16 16 ///< same as #MAX_LSPS; needs to be multiple
50 ///< of 16 for ASM input buffer alignment
51 #define MAX_FRAMES 3 ///< maximum number of frames per superframe
52 #define MAX_FRAMESIZE 160 ///< maximum number of samples per frame
53 #define MAX_SIGNAL_HISTORY 416 ///< maximum excitation signal history
54 #define MAX_SFRAMESIZE (MAX_FRAMESIZE * MAX_FRAMES)
55 ///< maximum number of samples per superframe
56 #define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that
57 ///< was split over two packets
58 #define VLC_NBITS 6 ///< number of bits to read per VLC iteration
61 * Frame type VLC coding.
63 static VLC frame_type_vlc;
66 * Adaptive codebook types.
69 ACB_TYPE_NONE = 0, ///< no adaptive codebook (only hardcoded fixed)
70 ACB_TYPE_ASYMMETRIC = 1, ///< adaptive codebook with per-frame pitch, which
71 ///< we interpolate to get a per-sample pitch.
72 ///< Signal is generated using an asymmetric sinc
74 ///< @note see #wmavoice_ipol1_coeffs
75 ACB_TYPE_HAMMING = 2 ///< Per-block pitch with signal generation using
76 ///< a Hamming sinc window function
77 ///< @note see #wmavoice_ipol2_coeffs
81 * Fixed codebook types.
84 FCB_TYPE_SILENCE = 0, ///< comfort noise during silence
85 ///< generated from a hardcoded (fixed) codebook
86 ///< with per-frame (low) gain values
87 FCB_TYPE_HARDCODED = 1, ///< hardcoded (fixed) codebook with per-block
89 FCB_TYPE_AW_PULSES = 2, ///< Pitch-adaptive window (AW) pulse signals,
90 ///< used in particular for low-bitrate streams
91 FCB_TYPE_EXC_PULSES = 3, ///< Innovation (fixed) codebook pulse sets in
92 ///< combinations of either single pulses or
97 * Description of frame types.
99 static const struct frame_type_desc {
100 uint8_t n_blocks; ///< amount of blocks per frame (each block
101 ///< (contains 160/#n_blocks samples)
102 uint8_t log_n_blocks; ///< log2(#n_blocks)
103 uint8_t acb_type; ///< Adaptive codebook type (ACB_TYPE_*)
104 uint8_t fcb_type; ///< Fixed codebook type (FCB_TYPE_*)
105 uint8_t dbl_pulses; ///< how many pulse vectors have pulse pairs
106 ///< (rather than just one single pulse)
107 ///< only if #fcb_type == #FCB_TYPE_EXC_PULSES
108 } frame_descs[17] = {
109 { 1, 0, ACB_TYPE_NONE, FCB_TYPE_SILENCE, 0 },
110 { 2, 1, ACB_TYPE_NONE, FCB_TYPE_HARDCODED, 0 },
111 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_AW_PULSES, 0 },
112 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2 },
113 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5 },
114 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 0 },
115 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2 },
116 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5 },
117 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0 },
118 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2 },
119 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5 },
120 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0 },
121 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2 },
122 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5 },
123 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0 },
124 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2 },
125 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5 }
129 * WMA Voice decoding context.
131 typedef struct WMAVoiceContext {
133 * @name Global values specified in the stream header / extradata or used all over.
136 GetBitContext gb; ///< packet bitreader. During decoder init,
137 ///< it contains the extradata from the
138 ///< demuxer. During decoding, it contains
140 int8_t vbm_tree[25]; ///< converts VLC codes to frame type
142 int spillover_bitsize; ///< number of bits used to specify
143 ///< #spillover_nbits in the packet header
144 ///< = ceil(log2(ctx->block_align << 3))
145 int history_nsamples; ///< number of samples in history for signal
146 ///< prediction (through ACB)
148 /* postfilter specific values */
149 int do_apf; ///< whether to apply the averaged
150 ///< projection filter (APF)
151 int denoise_strength; ///< strength of denoising in Wiener filter
153 int denoise_tilt_corr; ///< Whether to apply tilt correction to the
154 ///< Wiener filter coefficients (postfilter)
155 int dc_level; ///< Predicted amount of DC noise, based
156 ///< on which a DC removal filter is used
158 int lsps; ///< number of LSPs per frame [10 or 16]
159 int lsp_q_mode; ///< defines quantizer defaults [0, 1]
160 int lsp_def_mode; ///< defines different sets of LSP defaults
163 int min_pitch_val; ///< base value for pitch parsing code
164 int max_pitch_val; ///< max value + 1 for pitch parsing
165 int pitch_nbits; ///< number of bits used to specify the
166 ///< pitch value in the frame header
167 int block_pitch_nbits; ///< number of bits used to specify the
168 ///< first block's pitch value
169 int block_pitch_range; ///< range of the block pitch
170 int block_delta_pitch_nbits; ///< number of bits used to specify the
171 ///< delta pitch between this and the last
172 ///< block's pitch value, used in all but
174 int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is
175 ///< from -this to +this-1)
176 uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale
182 * @name Packet values specified in the packet header or related to a packet.
184 * A packet is considered to be a single unit of data provided to this
185 * decoder by the demuxer.
188 int spillover_nbits; ///< number of bits of the previous packet's
189 ///< last superframe preceding this
190 ///< packet's first full superframe (useful
191 ///< for re-synchronization also)
192 int has_residual_lsps; ///< if set, superframes contain one set of
193 ///< LSPs that cover all frames, encoded as
194 ///< independent and residual LSPs; if not
195 ///< set, each frame contains its own, fully
196 ///< independent, LSPs
197 int skip_bits_next; ///< number of bits to skip at the next call
198 ///< to #wmavoice_decode_packet() (since
199 ///< they're part of the previous superframe)
201 uint8_t sframe_cache[SFRAME_CACHE_MAXSIZE + AV_INPUT_BUFFER_PADDING_SIZE];
202 ///< cache for superframe data split over
203 ///< multiple packets
204 int sframe_cache_size; ///< set to >0 if we have data from an
205 ///< (incomplete) superframe from a previous
206 ///< packet that spilled over in the current
207 ///< packet; specifies the amount of bits in
209 PutBitContext pb; ///< bitstream writer for #sframe_cache
214 * @name Frame and superframe values
215 * Superframe and frame data - these can change from frame to frame,
216 * although some of them do in that case serve as a cache / history for
217 * the next frame or superframe.
220 double prev_lsps[MAX_LSPS]; ///< LSPs of the last frame of the previous
222 int last_pitch_val; ///< pitch value of the previous frame
223 int last_acb_type; ///< frame type [0-2] of the previous frame
224 int pitch_diff_sh16; ///< ((cur_pitch_val - #last_pitch_val)
225 ///< << 16) / #MAX_FRAMESIZE
226 float silence_gain; ///< set for use in blocks if #ACB_TYPE_NONE
228 int aw_idx_is_ext; ///< whether the AW index was encoded in
229 ///< 8 bits (instead of 6)
230 int aw_pulse_range; ///< the range over which #aw_pulse_set1()
231 ///< can apply the pulse, relative to the
232 ///< value in aw_first_pulse_off. The exact
233 ///< position of the first AW-pulse is within
234 ///< [pulse_off, pulse_off + this], and
235 ///< depends on bitstream values; [16 or 24]
236 int aw_n_pulses[2]; ///< number of AW-pulses in each block; note
237 ///< that this number can be negative (in
238 ///< which case it basically means "zero")
239 int aw_first_pulse_off[2]; ///< index of first sample to which to
240 ///< apply AW-pulses, or -0xff if unset
241 int aw_next_pulse_off_cache; ///< the position (relative to start of the
242 ///< second block) at which pulses should
243 ///< start to be positioned, serves as a
244 ///< cache for pitch-adaptive window pulses
247 int frame_cntr; ///< current frame index [0 - 0xFFFE]; is
248 ///< only used for comfort noise in #pRNG()
249 int nb_superframes; ///< number of superframes in current packet
250 float gain_pred_err[6]; ///< cache for gain prediction
251 float excitation_history[MAX_SIGNAL_HISTORY];
252 ///< cache of the signal of previous
253 ///< superframes, used as a history for
254 ///< signal generation
255 float synth_history[MAX_LSPS]; ///< see #excitation_history
259 * @name Postfilter values
261 * Variables used for postfilter implementation, mostly history for
262 * smoothing and so on, and context variables for FFT/iFFT.
265 RDFTContext rdft, irdft; ///< contexts for FFT-calculation in the
266 ///< postfilter (for denoise filter)
267 DCTContext dct, dst; ///< contexts for phase shift (in Hilbert
268 ///< transform, part of postfilter)
269 float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi]
271 float postfilter_agc; ///< gain control memory, used in
272 ///< #adaptive_gain_control()
273 float dcf_mem[2]; ///< DC filter history
274 float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE];
275 ///< zero filter output (i.e. excitation)
277 float denoise_filter_cache[MAX_FRAMESIZE];
278 int denoise_filter_cache_size; ///< samples in #denoise_filter_cache
279 DECLARE_ALIGNED(32, float, tilted_lpcs_pf)[0x80];
280 ///< aligned buffer for LPC tilting
281 DECLARE_ALIGNED(32, float, denoise_coeffs_pf)[0x80];
282 ///< aligned buffer for denoise coefficients
283 DECLARE_ALIGNED(32, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16];
284 ///< aligned buffer for postfilter speech
292 * Set up the variable bit mode (VBM) tree from container extradata.
293 * @param gb bit I/O context.
294 * The bit context (s->gb) should be loaded with byte 23-46 of the
295 * container extradata (i.e. the ones containing the VBM tree).
296 * @param vbm_tree pointer to array to which the decoded VBM tree will be
298 * @return 0 on success, <0 on error.
300 static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25])
302 int cntr[8] = { 0 }, n, res;
304 memset(vbm_tree, 0xff, sizeof(vbm_tree[0]) * 25);
305 for (n = 0; n < 17; n++) {
306 res = get_bits(gb, 3);
307 if (cntr[res] > 3) // should be >= 3 + (res == 7))
309 vbm_tree[res * 3 + cntr[res]++] = n;
314 static av_cold void wmavoice_init_static_data(void)
316 static const uint8_t bits[] = {
319 10, 10, 10, 12, 12, 12,
322 static const uint16_t codes[] = {
323 0x0000, 0x0001, 0x0002, // 00/01/10
324 0x000c, 0x000d, 0x000e, // 11+00/01/10
325 0x003c, 0x003d, 0x003e, // 1111+00/01/10
326 0x00fc, 0x00fd, 0x00fe, // 111111+00/01/10
327 0x03fc, 0x03fd, 0x03fe, // 11111111+00/01/10
328 0x0ffc, 0x0ffd, 0x0ffe, // 1111111111+00/01/10
329 0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx
332 INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits),
333 bits, 1, 1, codes, 2, 2, 132);
336 static av_cold void wmavoice_flush(AVCodecContext *ctx)
338 WMAVoiceContext *s = ctx->priv_data;
341 s->postfilter_agc = 0;
342 s->sframe_cache_size = 0;
343 s->skip_bits_next = 0;
344 for (n = 0; n < s->lsps; n++)
345 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
346 memset(s->excitation_history, 0,
347 sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY);
348 memset(s->synth_history, 0,
349 sizeof(*s->synth_history) * MAX_LSPS);
350 memset(s->gain_pred_err, 0,
351 sizeof(s->gain_pred_err));
354 memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0,
355 sizeof(*s->synth_filter_out_buf) * s->lsps);
356 memset(s->dcf_mem, 0,
357 sizeof(*s->dcf_mem) * 2);
358 memset(s->zero_exc_pf, 0,
359 sizeof(*s->zero_exc_pf) * s->history_nsamples);
360 memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache));
365 * Set up decoder with parameters from demuxer (extradata etc.).
367 static av_cold int wmavoice_decode_init(AVCodecContext *ctx)
369 static AVOnce init_static_once = AV_ONCE_INIT;
370 int n, flags, pitch_range, lsp16_flag;
371 WMAVoiceContext *s = ctx->priv_data;
373 ff_thread_once(&init_static_once, wmavoice_init_static_data);
377 * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c),
378 * - byte 19-22: flags field (annoyingly in LE; see below for known
380 * - byte 23-46: variable bitmode tree (really just 17 * 3 bits,
383 if (ctx->extradata_size != 46) {
384 av_log(ctx, AV_LOG_ERROR,
385 "Invalid extradata size %d (should be 46)\n",
386 ctx->extradata_size);
387 return AVERROR_INVALIDDATA;
389 if (ctx->block_align <= 0 || ctx->block_align > (1<<22)) {
390 av_log(ctx, AV_LOG_ERROR, "Invalid block alignment %d.\n", ctx->block_align);
391 return AVERROR_INVALIDDATA;
394 flags = AV_RL32(ctx->extradata + 18);
395 s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align);
396 s->do_apf = flags & 0x1;
398 ff_rdft_init(&s->rdft, 7, DFT_R2C);
399 ff_rdft_init(&s->irdft, 7, IDFT_C2R);
400 ff_dct_init(&s->dct, 6, DCT_I);
401 ff_dct_init(&s->dst, 6, DST_I);
403 ff_sine_window_init(s->cos, 256);
404 memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0]));
405 for (n = 0; n < 255; n++) {
406 s->sin[n] = -s->sin[510 - n];
407 s->cos[510 - n] = s->cos[n];
410 s->denoise_strength = (flags >> 2) & 0xF;
411 if (s->denoise_strength >= 12) {
412 av_log(ctx, AV_LOG_ERROR,
413 "Invalid denoise filter strength %d (max=11)\n",
414 s->denoise_strength);
415 return AVERROR_INVALIDDATA;
417 s->denoise_tilt_corr = !!(flags & 0x40);
418 s->dc_level = (flags >> 7) & 0xF;
419 s->lsp_q_mode = !!(flags & 0x2000);
420 s->lsp_def_mode = !!(flags & 0x4000);
421 lsp16_flag = flags & 0x1000;
427 for (n = 0; n < s->lsps; n++)
428 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
430 init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3);
431 if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) {
432 av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n");
433 return AVERROR_INVALIDDATA;
436 if (ctx->sample_rate >= INT_MAX / (256 * 37))
437 return AVERROR_INVALIDDATA;
439 s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8;
440 s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8;
441 pitch_range = s->max_pitch_val - s->min_pitch_val;
442 if (pitch_range <= 0) {
443 av_log(ctx, AV_LOG_ERROR, "Invalid pitch range; broken extradata?\n");
444 return AVERROR_INVALIDDATA;
446 s->pitch_nbits = av_ceil_log2(pitch_range);
447 s->last_pitch_val = 40;
448 s->last_acb_type = ACB_TYPE_NONE;
449 s->history_nsamples = s->max_pitch_val + 8;
451 if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) {
452 int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8,
453 max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8;
455 av_log(ctx, AV_LOG_ERROR,
456 "Unsupported samplerate %d (min=%d, max=%d)\n",
457 ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz
459 return AVERROR(ENOSYS);
462 s->block_conv_table[0] = s->min_pitch_val;
463 s->block_conv_table[1] = (pitch_range * 25) >> 6;
464 s->block_conv_table[2] = (pitch_range * 44) >> 6;
465 s->block_conv_table[3] = s->max_pitch_val - 1;
466 s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF;
467 if (s->block_delta_pitch_hrange <= 0) {
468 av_log(ctx, AV_LOG_ERROR, "Invalid delta pitch hrange; broken extradata?\n");
469 return AVERROR_INVALIDDATA;
471 s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange);
472 s->block_pitch_range = s->block_conv_table[2] +
473 s->block_conv_table[3] + 1 +
474 2 * (s->block_conv_table[1] - 2 * s->min_pitch_val);
475 s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range);
478 ctx->channel_layout = AV_CH_LAYOUT_MONO;
479 ctx->sample_fmt = AV_SAMPLE_FMT_FLT;
485 * @name Postfilter functions
486 * Postfilter functions (gain control, wiener denoise filter, DC filter,
487 * kalman smoothening, plus surrounding code to wrap it)
491 * Adaptive gain control (as used in postfilter).
493 * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except
494 * that the energy here is calculated using sum(abs(...)), whereas the
495 * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)).
497 * @param out output buffer for filtered samples
498 * @param in input buffer containing the samples as they are after the
499 * postfilter steps so far
500 * @param speech_synth input buffer containing speech synth before postfilter
501 * @param size input buffer size
502 * @param alpha exponential filter factor
503 * @param gain_mem pointer to filter memory (single float)
505 static void adaptive_gain_control(float *out, const float *in,
506 const float *speech_synth,
507 int size, float alpha, float *gain_mem)
510 float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor;
511 float mem = *gain_mem;
513 for (i = 0; i < size; i++) {
514 speech_energy += fabsf(speech_synth[i]);
515 postfilter_energy += fabsf(in[i]);
517 gain_scale_factor = postfilter_energy == 0.0 ? 0.0 :
518 (1.0 - alpha) * speech_energy / postfilter_energy;
520 for (i = 0; i < size; i++) {
521 mem = alpha * mem + gain_scale_factor;
522 out[i] = in[i] * mem;
529 * Kalman smoothing function.
531 * This function looks back pitch +/- 3 samples back into history to find
532 * the best fitting curve (that one giving the optimal gain of the two
533 * signals, i.e. the highest dot product between the two), and then
534 * uses that signal history to smoothen the output of the speech synthesis
537 * @param s WMA Voice decoding context
538 * @param pitch pitch of the speech signal
539 * @param in input speech signal
540 * @param out output pointer for smoothened signal
541 * @param size input/output buffer size
543 * @returns -1 if no smoothening took place, e.g. because no optimal
544 * fit could be found, or 0 on success.
546 static int kalman_smoothen(WMAVoiceContext *s, int pitch,
547 const float *in, float *out, int size)
550 float optimal_gain = 0, dot;
551 const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)],
552 *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)],
553 *best_hist_ptr = NULL;
555 /* find best fitting point in history */
557 dot = avpriv_scalarproduct_float_c(in, ptr, size);
558 if (dot > optimal_gain) {
562 } while (--ptr >= end);
564 if (optimal_gain <= 0)
566 dot = avpriv_scalarproduct_float_c(best_hist_ptr, best_hist_ptr, size);
567 if (dot <= 0) // would be 1.0
570 if (optimal_gain <= dot) {
571 dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000
575 /* actual smoothing */
576 for (n = 0; n < size; n++)
577 out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]);
583 * Get the tilt factor of a formant filter from its transfer function
584 * @see #tilt_factor() in amrnbdec.c, which does essentially the same,
585 * but somehow (??) it does a speech synthesis filter in the
586 * middle, which is missing here
588 * @param lpcs LPC coefficients
589 * @param n_lpcs Size of LPC buffer
590 * @returns the tilt factor
592 static float tilt_factor(const float *lpcs, int n_lpcs)
596 rh0 = 1.0 + avpriv_scalarproduct_float_c(lpcs, lpcs, n_lpcs);
597 rh1 = lpcs[0] + avpriv_scalarproduct_float_c(lpcs, &lpcs[1], n_lpcs - 1);
603 * Derive denoise filter coefficients (in real domain) from the LPCs.
605 static void calc_input_response(WMAVoiceContext *s, float *lpcs,
606 int fcb_type, float *coeffs, int remainder)
608 float last_coeff, min = 15.0, max = -15.0;
609 float irange, angle_mul, gain_mul, range, sq;
612 /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */
613 s->rdft.rdft_calc(&s->rdft, lpcs);
614 #define log_range(var, assign) do { \
615 float tmp = log10f(assign); var = tmp; \
616 max = FFMAX(max, tmp); min = FFMIN(min, tmp); \
618 log_range(last_coeff, lpcs[1] * lpcs[1]);
619 for (n = 1; n < 64; n++)
620 log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] +
621 lpcs[n * 2 + 1] * lpcs[n * 2 + 1]);
622 log_range(lpcs[0], lpcs[0] * lpcs[0]);
625 lpcs[64] = last_coeff;
627 /* Now, use this spectrum to pick out these frequencies with higher
628 * (relative) power/energy (which we then take to be "not noise"),
629 * and set up a table (still in lpc[]) of (relative) gains per frequency.
630 * These frequencies will be maintained, while others ("noise") will be
631 * decreased in the filter output. */
632 irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63]
633 gain_mul = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) :
635 angle_mul = gain_mul * (8.0 * M_LN10 / M_PI);
636 for (n = 0; n <= 64; n++) {
639 idx = FFMAX(0, lrint((max - lpcs[n]) * irange) - 1);
640 pwr = wmavoice_denoise_power_table[s->denoise_strength][idx];
641 lpcs[n] = angle_mul * pwr;
643 /* 70.57 =~ 1/log10(1.0331663) */
644 idx = (pwr * gain_mul - 0.0295) * 70.570526123;
645 if (idx > 127) { // fall back if index falls outside table range
646 coeffs[n] = wmavoice_energy_table[127] *
647 powf(1.0331663, idx - 127);
649 coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)];
652 /* calculate the Hilbert transform of the gains, which we do (since this
653 * is a sine input) by doing a phase shift (in theory, H(sin())=cos()).
654 * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the
655 * "moment" of the LPCs in this filter. */
656 s->dct.dct_calc(&s->dct, lpcs);
657 s->dst.dct_calc(&s->dst, lpcs);
659 /* Split out the coefficient indexes into phase/magnitude pairs */
660 idx = 255 + av_clip(lpcs[64], -255, 255);
661 coeffs[0] = coeffs[0] * s->cos[idx];
662 idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255);
663 last_coeff = coeffs[64] * s->cos[idx];
665 idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255);
666 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
667 coeffs[n * 2] = coeffs[n] * s->cos[idx];
671 idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255);
672 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
673 coeffs[n * 2] = coeffs[n] * s->cos[idx];
675 coeffs[1] = last_coeff;
677 /* move into real domain */
678 s->irdft.rdft_calc(&s->irdft, coeffs);
680 /* tilt correction and normalize scale */
681 memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder));
682 if (s->denoise_tilt_corr) {
685 coeffs[remainder - 1] = 0;
686 ff_tilt_compensation(&tilt_mem,
687 -1.8 * tilt_factor(coeffs, remainder - 1),
690 sq = (1.0 / 64.0) * sqrtf(1 / avpriv_scalarproduct_float_c(coeffs, coeffs,
692 for (n = 0; n < remainder; n++)
697 * This function applies a Wiener filter on the (noisy) speech signal as
698 * a means to denoise it.
700 * - take RDFT of LPCs to get the power spectrum of the noise + speech;
701 * - using this power spectrum, calculate (for each frequency) the Wiener
702 * filter gain, which depends on the frequency power and desired level
703 * of noise subtraction (when set too high, this leads to artifacts)
704 * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse
706 * - by doing a phase shift, calculate the Hilbert transform of this array
707 * of per-frequency filter-gains to get the filtering coefficients;
708 * - smoothen/normalize/de-tilt these filter coefficients as desired;
709 * - take RDFT of noisy sound, apply the coefficients and take its IRDFT
710 * to get the denoised speech signal;
711 * - the leftover (i.e. output of the IRDFT on denoised speech data beyond
712 * the frame boundary) are saved and applied to subsequent frames by an
713 * overlap-add method (otherwise you get clicking-artifacts).
715 * @param s WMA Voice decoding context
716 * @param fcb_type Frame (codebook) type
717 * @param synth_pf input: the noisy speech signal, output: denoised speech
718 * data; should be 16-byte aligned (for ASM purposes)
719 * @param size size of the speech data
720 * @param lpcs LPCs used to synthesize this frame's speech data
722 static void wiener_denoise(WMAVoiceContext *s, int fcb_type,
723 float *synth_pf, int size,
726 int remainder, lim, n;
728 if (fcb_type != FCB_TYPE_SILENCE) {
729 float *tilted_lpcs = s->tilted_lpcs_pf,
730 *coeffs = s->denoise_coeffs_pf, tilt_mem = 0;
732 tilted_lpcs[0] = 1.0;
733 memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps);
734 memset(&tilted_lpcs[s->lsps + 1], 0,
735 sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1));
736 ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps),
737 tilted_lpcs, s->lsps + 2);
739 /* The IRDFT output (127 samples for 7-bit filter) beyond the frame
740 * size is applied to the next frame. All input beyond this is zero,
741 * and thus all output beyond this will go towards zero, hence we can
742 * limit to min(size-1, 127-size) as a performance consideration. */
743 remainder = FFMIN(127 - size, size - 1);
744 calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder);
746 /* apply coefficients (in frequency spectrum domain), i.e. complex
747 * number multiplication */
748 memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size));
749 s->rdft.rdft_calc(&s->rdft, synth_pf);
750 s->rdft.rdft_calc(&s->rdft, coeffs);
751 synth_pf[0] *= coeffs[0];
752 synth_pf[1] *= coeffs[1];
753 for (n = 1; n < 64; n++) {
754 float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1];
755 synth_pf[n * 2] = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1];
756 synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1];
758 s->irdft.rdft_calc(&s->irdft, synth_pf);
761 /* merge filter output with the history of previous runs */
762 if (s->denoise_filter_cache_size) {
763 lim = FFMIN(s->denoise_filter_cache_size, size);
764 for (n = 0; n < lim; n++)
765 synth_pf[n] += s->denoise_filter_cache[n];
766 s->denoise_filter_cache_size -= lim;
767 memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size],
768 sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size);
771 /* move remainder of filter output into a cache for future runs */
772 if (fcb_type != FCB_TYPE_SILENCE) {
773 lim = FFMIN(remainder, s->denoise_filter_cache_size);
774 for (n = 0; n < lim; n++)
775 s->denoise_filter_cache[n] += synth_pf[size + n];
776 if (lim < remainder) {
777 memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim],
778 sizeof(s->denoise_filter_cache[0]) * (remainder - lim));
779 s->denoise_filter_cache_size = remainder;
785 * Averaging projection filter, the postfilter used in WMAVoice.
787 * This uses the following steps:
788 * - A zero-synthesis filter (generate excitation from synth signal)
789 * - Kalman smoothing on excitation, based on pitch
790 * - Re-synthesized smoothened output
791 * - Iterative Wiener denoise filter
792 * - Adaptive gain filter
795 * @param s WMAVoice decoding context
796 * @param synth Speech synthesis output (before postfilter)
797 * @param samples Output buffer for filtered samples
798 * @param size Buffer size of synth & samples
799 * @param lpcs Generated LPCs used for speech synthesis
800 * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned)
801 * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses)
802 * @param pitch Pitch of the input signal
804 static void postfilter(WMAVoiceContext *s, const float *synth,
805 float *samples, int size,
806 const float *lpcs, float *zero_exc_pf,
807 int fcb_type, int pitch)
809 float synth_filter_in_buf[MAX_FRAMESIZE / 2],
810 *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16],
811 *synth_filter_in = zero_exc_pf;
813 av_assert0(size <= MAX_FRAMESIZE / 2);
815 /* generate excitation from input signal */
816 ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps);
818 if (fcb_type >= FCB_TYPE_AW_PULSES &&
819 !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size))
820 synth_filter_in = synth_filter_in_buf;
822 /* re-synthesize speech after smoothening, and keep history */
823 ff_celp_lp_synthesis_filterf(synth_pf, lpcs,
824 synth_filter_in, size, s->lsps);
825 memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps],
826 sizeof(synth_pf[0]) * s->lsps);
828 wiener_denoise(s, fcb_type, synth_pf, size, lpcs);
830 adaptive_gain_control(samples, synth_pf, synth, size, 0.99,
833 if (s->dc_level > 8) {
834 /* remove ultra-low frequency DC noise / highpass filter;
835 * coefficients are identical to those used in SIPR decoding,
836 * and very closely resemble those used in AMR-NB decoding. */
837 ff_acelp_apply_order_2_transfer_function(samples, samples,
838 (const float[2]) { -1.99997, 1.0 },
839 (const float[2]) { -1.9330735188, 0.93589198496 },
840 0.93980580475, s->dcf_mem, size);
849 * @param lsps output pointer to the array that will hold the LSPs
850 * @param num number of LSPs to be dequantized
851 * @param values quantized values, contains n_stages values
852 * @param sizes range (i.e. max value) of each quantized value
853 * @param n_stages number of dequantization runs
854 * @param table dequantization table to be used
855 * @param mul_q LSF multiplier
856 * @param base_q base (lowest) LSF values
858 static void dequant_lsps(double *lsps, int num,
859 const uint16_t *values,
860 const uint16_t *sizes,
861 int n_stages, const uint8_t *table,
863 const double *base_q)
867 memset(lsps, 0, num * sizeof(*lsps));
868 for (n = 0; n < n_stages; n++) {
869 const uint8_t *t_off = &table[values[n] * num];
870 double base = base_q[n], mul = mul_q[n];
872 for (m = 0; m < num; m++)
873 lsps[m] += base + mul * t_off[m];
875 table += sizes[n] * num;
880 * @name LSP dequantization routines
881 * LSP dequantization routines, for 10/16LSPs and independent/residual coding.
882 * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;
883 * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits.
887 * Parse 10 independently-coded LSPs.
889 static void dequant_lsp10i(GetBitContext *gb, double *lsps)
891 static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 };
892 static const double mul_lsf[4] = {
893 5.2187144800e-3, 1.4626986422e-3,
894 9.6179549166e-4, 1.1325736225e-3
896 static const double base_lsf[4] = {
897 M_PI * -2.15522e-1, M_PI * -6.1646e-2,
898 M_PI * -3.3486e-2, M_PI * -5.7408e-2
902 v[0] = get_bits(gb, 8);
903 v[1] = get_bits(gb, 6);
904 v[2] = get_bits(gb, 5);
905 v[3] = get_bits(gb, 5);
907 dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i,
912 * Parse 10 independently-coded LSPs, and then derive the tables to
913 * generate LSPs for the other frames from them (residual coding).
915 static void dequant_lsp10r(GetBitContext *gb,
916 double *i_lsps, const double *old,
917 double *a1, double *a2, int q_mode)
919 static const uint16_t vec_sizes[3] = { 128, 64, 64 };
920 static const double mul_lsf[3] = {
921 2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3
923 static const double base_lsf[3] = {
924 M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2
926 const float (*ipol_tab)[2][10] = q_mode ?
927 wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a;
928 uint16_t interpol, v[3];
931 dequant_lsp10i(gb, i_lsps);
933 interpol = get_bits(gb, 5);
934 v[0] = get_bits(gb, 7);
935 v[1] = get_bits(gb, 6);
936 v[2] = get_bits(gb, 6);
938 for (n = 0; n < 10; n++) {
939 double delta = old[n] - i_lsps[n];
940 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
941 a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
944 dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r,
949 * Parse 16 independently-coded LSPs.
951 static void dequant_lsp16i(GetBitContext *gb, double *lsps)
953 static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 };
954 static const double mul_lsf[5] = {
955 3.3439586280e-3, 6.9908173703e-4,
956 3.3216608306e-3, 1.0334960326e-3,
959 static const double base_lsf[5] = {
960 M_PI * -1.27576e-1, M_PI * -2.4292e-2,
961 M_PI * -1.28094e-1, M_PI * -3.2128e-2,
966 v[0] = get_bits(gb, 8);
967 v[1] = get_bits(gb, 6);
968 v[2] = get_bits(gb, 7);
969 v[3] = get_bits(gb, 6);
970 v[4] = get_bits(gb, 7);
972 dequant_lsps( lsps, 5, v, vec_sizes, 2,
973 wmavoice_dq_lsp16i1, mul_lsf, base_lsf);
974 dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2,
975 wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]);
976 dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1,
977 wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]);
981 * Parse 16 independently-coded LSPs, and then derive the tables to
982 * generate LSPs for the other frames from them (residual coding).
984 static void dequant_lsp16r(GetBitContext *gb,
985 double *i_lsps, const double *old,
986 double *a1, double *a2, int q_mode)
988 static const uint16_t vec_sizes[3] = { 128, 128, 128 };
989 static const double mul_lsf[3] = {
990 1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3
992 static const double base_lsf[3] = {
993 M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2
995 const float (*ipol_tab)[2][16] = q_mode ?
996 wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a;
997 uint16_t interpol, v[3];
1000 dequant_lsp16i(gb, i_lsps);
1002 interpol = get_bits(gb, 5);
1003 v[0] = get_bits(gb, 7);
1004 v[1] = get_bits(gb, 7);
1005 v[2] = get_bits(gb, 7);
1007 for (n = 0; n < 16; n++) {
1008 double delta = old[n] - i_lsps[n];
1009 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
1010 a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
1013 dequant_lsps( a2, 10, v, vec_sizes, 1,
1014 wmavoice_dq_lsp16r1, mul_lsf, base_lsf);
1015 dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1,
1016 wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]);
1017 dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1,
1018 wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]);
1023 * @name Pitch-adaptive window coding functions
1024 * The next few functions are for pitch-adaptive window coding.
1028 * Parse the offset of the first pitch-adaptive window pulses, and
1029 * the distribution of pulses between the two blocks in this frame.
1030 * @param s WMA Voice decoding context private data
1031 * @param gb bit I/O context
1032 * @param pitch pitch for each block in this frame
1034 static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb,
1037 static const int16_t start_offset[94] = {
1038 -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11,
1039 13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26,
1040 27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43,
1041 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67,
1042 69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91,
1043 93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115,
1044 117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139,
1045 141, 143, 145, 147, 149, 151, 153, 155, 157, 159
1049 /* position of pulse */
1050 s->aw_idx_is_ext = 0;
1051 if ((bits = get_bits(gb, 6)) >= 54) {
1052 s->aw_idx_is_ext = 1;
1053 bits += (bits - 54) * 3 + get_bits(gb, 2);
1056 /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count
1057 * the distribution of the pulses in each block contained in this frame. */
1058 s->aw_pulse_range = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16;
1059 for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ;
1060 s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0];
1061 s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2;
1062 offset += s->aw_n_pulses[0] * pitch[0];
1063 s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1];
1064 s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2;
1066 /* if continuing from a position before the block, reset position to
1067 * start of block (when corrected for the range over which it can be
1068 * spread in aw_pulse_set1()). */
1069 if (start_offset[bits] < MAX_FRAMESIZE / 2) {
1070 while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0)
1071 s->aw_first_pulse_off[1] -= pitch[1];
1072 if (start_offset[bits] < 0)
1073 while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0)
1074 s->aw_first_pulse_off[0] -= pitch[0];
1079 * Apply second set of pitch-adaptive window pulses.
1080 * @param s WMA Voice decoding context private data
1081 * @param gb bit I/O context
1082 * @param block_idx block index in frame [0, 1]
1083 * @param fcb structure containing fixed codebook vector info
1084 * @return -1 on error, 0 otherwise
1086 static int aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb,
1087 int block_idx, AMRFixed *fcb)
1089 uint16_t use_mask_mem[9]; // only 5 are used, rest is padding
1090 uint16_t *use_mask = use_mask_mem + 2;
1091 /* in this function, idx is the index in the 80-bit (+ padding) use_mask
1092 * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits
1093 * of idx are the position of the bit within a particular item in the
1094 * array (0 being the most significant bit, and 15 being the least
1095 * significant bit), and the remainder (>> 4) is the index in the
1096 * use_mask[]-array. This is faster and uses less memory than using a
1097 * 80-byte/80-int array. */
1098 int pulse_off = s->aw_first_pulse_off[block_idx],
1099 pulse_start, n, idx, range, aidx, start_off = 0;
1101 /* set offset of first pulse to within this block */
1102 if (s->aw_n_pulses[block_idx] > 0)
1103 while (pulse_off + s->aw_pulse_range < 1)
1104 pulse_off += fcb->pitch_lag;
1106 /* find range per pulse */
1107 if (s->aw_n_pulses[0] > 0) {
1108 if (block_idx == 0) {
1110 } else /* block_idx = 1 */ {
1112 if (s->aw_n_pulses[block_idx] > 0)
1113 pulse_off = s->aw_next_pulse_off_cache;
1117 pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0;
1119 /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,
1120 * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus
1121 * we exclude that range from being pulsed again in this function. */
1122 memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0]));
1123 memset( use_mask, -1, 5 * sizeof(use_mask[0]));
1124 memset(&use_mask[5], 0, 2 * sizeof(use_mask[0]));
1125 if (s->aw_n_pulses[block_idx] > 0)
1126 for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) {
1127 int excl_range = s->aw_pulse_range; // always 16 or 24
1128 uint16_t *use_mask_ptr = &use_mask[idx >> 4];
1129 int first_sh = 16 - (idx & 15);
1130 *use_mask_ptr++ &= 0xFFFFu << first_sh;
1131 excl_range -= first_sh;
1132 if (excl_range >= 16) {
1133 *use_mask_ptr++ = 0;
1134 *use_mask_ptr &= 0xFFFF >> (excl_range - 16);
1136 *use_mask_ptr &= 0xFFFF >> excl_range;
1139 /* find the 'aidx'th offset that is not excluded */
1140 aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4);
1141 for (n = 0; n <= aidx; pulse_start++) {
1142 for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ;
1143 if (idx >= MAX_FRAMESIZE / 2) { // find from zero
1144 if (use_mask[0]) idx = 0x0F;
1145 else if (use_mask[1]) idx = 0x1F;
1146 else if (use_mask[2]) idx = 0x2F;
1147 else if (use_mask[3]) idx = 0x3F;
1148 else if (use_mask[4]) idx = 0x4F;
1150 idx -= av_log2_16bit(use_mask[idx >> 4]);
1152 if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) {
1153 use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15));
1159 fcb->x[fcb->n] = start_off;
1160 fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0;
1163 /* set offset for next block, relative to start of that block */
1164 n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag;
1165 s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0;
1170 * Apply first set of pitch-adaptive window pulses.
1171 * @param s WMA Voice decoding context private data
1172 * @param gb bit I/O context
1173 * @param block_idx block index in frame [0, 1]
1174 * @param fcb storage location for fixed codebook pulse info
1176 static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb,
1177 int block_idx, AMRFixed *fcb)
1179 int val = get_bits(gb, 12 - 2 * (s->aw_idx_is_ext && !block_idx));
1182 if (s->aw_n_pulses[block_idx] > 0) {
1183 int n, v_mask, i_mask, sh, n_pulses;
1185 if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each
1190 } else { // 4 pulses, 1:sign + 2:index each
1197 for (n = n_pulses - 1; n >= 0; n--, val >>= sh) {
1198 fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0;
1199 fcb->x[fcb->n] = (val & i_mask) * n_pulses + n +
1200 s->aw_first_pulse_off[block_idx];
1201 while (fcb->x[fcb->n] < 0)
1202 fcb->x[fcb->n] += fcb->pitch_lag;
1203 if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2)
1207 int num2 = (val & 0x1FF) >> 1, delta, idx;
1209 if (num2 < 1 * 79) { delta = 1; idx = num2 + 1; }
1210 else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; }
1211 else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; }
1212 else { delta = 7; idx = num2 + 1 - 3 * 75; }
1213 v = (val & 0x200) ? -1.0 : 1.0;
1215 fcb->no_repeat_mask |= 3 << fcb->n;
1216 fcb->x[fcb->n] = idx - delta;
1218 fcb->x[fcb->n + 1] = idx;
1219 fcb->y[fcb->n + 1] = (val & 1) ? -v : v;
1227 * Generate a random number from frame_cntr and block_idx, which will live
1228 * in the range [0, 1000 - block_size] (so it can be used as an index in a
1229 * table of size 1000 of which you want to read block_size entries).
1231 * @param frame_cntr current frame number
1232 * @param block_num current block index
1233 * @param block_size amount of entries we want to read from a table
1234 * that has 1000 entries
1235 * @return a (non-)random number in the [0, 1000 - block_size] range.
1237 static int pRNG(int frame_cntr, int block_num, int block_size)
1239 /* array to simplify the calculation of z:
1240 * y = (x % 9) * 5 + 6;
1241 * z = (49995 * x) / y;
1242 * Since y only has 9 values, we can remove the division by using a
1243 * LUT and using FASTDIV-style divisions. For each of the 9 values
1244 * of y, we can rewrite z as:
1245 * z = x * (49995 / y) + x * ((49995 % y) / y)
1246 * In this table, each col represents one possible value of y, the
1247 * first number is 49995 / y, and the second is the FASTDIV variant
1248 * of 49995 % y / y. */
1249 static const unsigned int div_tbl[9][2] = {
1250 { 8332, 3 * 715827883U }, // y = 6
1251 { 4545, 0 * 390451573U }, // y = 11
1252 { 3124, 11 * 268435456U }, // y = 16
1253 { 2380, 15 * 204522253U }, // y = 21
1254 { 1922, 23 * 165191050U }, // y = 26
1255 { 1612, 23 * 138547333U }, // y = 31
1256 { 1388, 27 * 119304648U }, // y = 36
1257 { 1219, 16 * 104755300U }, // y = 41
1258 { 1086, 39 * 93368855U } // y = 46
1260 unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr;
1261 if (x >= 0xFFFF) x -= 0xFFFF; // max value of x is 8*1877+0xFFFE=0x13AA6,
1262 // so this is effectively a modulo (%)
1263 y = x - 9 * MULH(477218589, x); // x % 9
1264 z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1]));
1265 // z = x * 49995 / (y * 5 + 6)
1266 return z % (1000 - block_size);
1270 * Parse hardcoded signal for a single block.
1271 * @note see #synth_block().
1273 static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb,
1274 int block_idx, int size,
1275 const struct frame_type_desc *frame_desc,
1281 av_assert0(size <= MAX_FRAMESIZE);
1283 /* Set the offset from which we start reading wmavoice_std_codebook */
1284 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1285 r_idx = pRNG(s->frame_cntr, block_idx, size);
1286 gain = s->silence_gain;
1287 } else /* FCB_TYPE_HARDCODED */ {
1288 r_idx = get_bits(gb, 8);
1289 gain = wmavoice_gain_universal[get_bits(gb, 6)];
1292 /* Clear gain prediction parameters */
1293 memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err));
1295 /* Apply gain to hardcoded codebook and use that as excitation signal */
1296 for (n = 0; n < size; n++)
1297 excitation[n] = wmavoice_std_codebook[r_idx + n] * gain;
1301 * Parse FCB/ACB signal for a single block.
1302 * @note see #synth_block().
1304 static void synth_block_fcb_acb(WMAVoiceContext *s, GetBitContext *gb,
1305 int block_idx, int size,
1306 int block_pitch_sh2,
1307 const struct frame_type_desc *frame_desc,
1310 static const float gain_coeff[6] = {
1311 0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458
1313 float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain;
1314 int n, idx, gain_weight;
1317 av_assert0(size <= MAX_FRAMESIZE / 2);
1318 memset(pulses, 0, sizeof(*pulses) * size);
1320 fcb.pitch_lag = block_pitch_sh2 >> 2;
1321 fcb.pitch_fac = 1.0;
1322 fcb.no_repeat_mask = 0;
1325 /* For the other frame types, this is where we apply the innovation
1326 * (fixed) codebook pulses of the speech signal. */
1327 if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1328 aw_pulse_set1(s, gb, block_idx, &fcb);
1329 if (aw_pulse_set2(s, gb, block_idx, &fcb)) {
1330 /* Conceal the block with silence and return.
1331 * Skip the correct amount of bits to read the next
1332 * block from the correct offset. */
1333 int r_idx = pRNG(s->frame_cntr, block_idx, size);
1335 for (n = 0; n < size; n++)
1337 wmavoice_std_codebook[r_idx + n] * s->silence_gain;
1338 skip_bits(gb, 7 + 1);
1341 } else /* FCB_TYPE_EXC_PULSES */ {
1342 int offset_nbits = 5 - frame_desc->log_n_blocks;
1344 fcb.no_repeat_mask = -1;
1345 /* similar to ff_decode_10_pulses_35bits(), but with single pulses
1346 * (instead of double) for a subset of pulses */
1347 for (n = 0; n < 5; n++) {
1351 sign = get_bits1(gb) ? 1.0 : -1.0;
1352 pos1 = get_bits(gb, offset_nbits);
1353 fcb.x[fcb.n] = n + 5 * pos1;
1354 fcb.y[fcb.n++] = sign;
1355 if (n < frame_desc->dbl_pulses) {
1356 pos2 = get_bits(gb, offset_nbits);
1357 fcb.x[fcb.n] = n + 5 * pos2;
1358 fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign;
1362 ff_set_fixed_vector(pulses, &fcb, 1.0, size);
1364 /* Calculate gain for adaptive & fixed codebook signal.
1365 * see ff_amr_set_fixed_gain(). */
1366 idx = get_bits(gb, 7);
1367 fcb_gain = expf(avpriv_scalarproduct_float_c(s->gain_pred_err,
1369 5.2409161640 + wmavoice_gain_codebook_fcb[idx]);
1370 acb_gain = wmavoice_gain_codebook_acb[idx];
1371 pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx],
1372 -2.9957322736 /* log(0.05) */,
1373 1.6094379124 /* log(5.0) */);
1375 gain_weight = 8 >> frame_desc->log_n_blocks;
1376 memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err,
1377 sizeof(*s->gain_pred_err) * (6 - gain_weight));
1378 for (n = 0; n < gain_weight; n++)
1379 s->gain_pred_err[n] = pred_err;
1381 /* Calculation of adaptive codebook */
1382 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1384 for (n = 0; n < size; n += len) {
1386 int abs_idx = block_idx * size + n;
1387 int pitch_sh16 = (s->last_pitch_val << 16) +
1388 s->pitch_diff_sh16 * abs_idx;
1389 int pitch = (pitch_sh16 + 0x6FFF) >> 16;
1390 int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000;
1391 idx = idx_sh16 >> 16;
1392 if (s->pitch_diff_sh16) {
1393 if (s->pitch_diff_sh16 > 0) {
1394 next_idx_sh16 = (idx_sh16) &~ 0xFFFF;
1396 next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF;
1397 len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8,
1402 ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch],
1403 wmavoice_ipol1_coeffs, 17,
1406 } else /* ACB_TYPE_HAMMING */ {
1407 int block_pitch = block_pitch_sh2 >> 2;
1408 idx = block_pitch_sh2 & 3;
1410 ff_acelp_interpolatef(excitation, &excitation[-block_pitch],
1411 wmavoice_ipol2_coeffs, 4,
1414 av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch,
1415 sizeof(float) * size);
1418 /* Interpolate ACB/FCB and use as excitation signal */
1419 ff_weighted_vector_sumf(excitation, excitation, pulses,
1420 acb_gain, fcb_gain, size);
1424 * Parse data in a single block.
1426 * @param s WMA Voice decoding context private data
1427 * @param gb bit I/O context
1428 * @param block_idx index of the to-be-read block
1429 * @param size amount of samples to be read in this block
1430 * @param block_pitch_sh2 pitch for this block << 2
1431 * @param lsps LSPs for (the end of) this frame
1432 * @param prev_lsps LSPs for the last frame
1433 * @param frame_desc frame type descriptor
1434 * @param excitation target memory for the ACB+FCB interpolated signal
1435 * @param synth target memory for the speech synthesis filter output
1436 * @return 0 on success, <0 on error.
1438 static void synth_block(WMAVoiceContext *s, GetBitContext *gb,
1439 int block_idx, int size,
1440 int block_pitch_sh2,
1441 const double *lsps, const double *prev_lsps,
1442 const struct frame_type_desc *frame_desc,
1443 float *excitation, float *synth)
1445 double i_lsps[MAX_LSPS];
1446 float lpcs[MAX_LSPS];
1450 if (frame_desc->acb_type == ACB_TYPE_NONE)
1451 synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation);
1453 synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2,
1454 frame_desc, excitation);
1456 /* convert interpolated LSPs to LPCs */
1457 fac = (block_idx + 0.5) / frame_desc->n_blocks;
1458 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1459 i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n]));
1460 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1462 /* Speech synthesis */
1463 ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps);
1467 * Synthesize output samples for a single frame.
1469 * @param ctx WMA Voice decoder context
1470 * @param gb bit I/O context (s->gb or one for cross-packet superframes)
1471 * @param frame_idx Frame number within superframe [0-2]
1472 * @param samples pointer to output sample buffer, has space for at least 160
1474 * @param lsps LSP array
1475 * @param prev_lsps array of previous frame's LSPs
1476 * @param excitation target buffer for excitation signal
1477 * @param synth target buffer for synthesized speech data
1478 * @return 0 on success, <0 on error.
1480 static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx,
1482 const double *lsps, const double *prev_lsps,
1483 float *excitation, float *synth)
1485 WMAVoiceContext *s = ctx->priv_data;
1486 int n, n_blocks_x2, log_n_blocks_x2, av_uninit(cur_pitch_val);
1487 int pitch[MAX_BLOCKS], av_uninit(last_block_pitch);
1489 /* Parse frame type ("frame header"), see frame_descs */
1490 int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)], block_nsamples;
1493 av_log(ctx, AV_LOG_ERROR,
1494 "Invalid frame type VLC code, skipping\n");
1495 return AVERROR_INVALIDDATA;
1498 block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks;
1500 /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */
1501 if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) {
1502 /* Pitch is provided per frame, which is interpreted as the pitch of
1503 * the last sample of the last block of this frame. We can interpolate
1504 * the pitch of other blocks (and even pitch-per-sample) by gradually
1505 * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */
1506 n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1;
1507 log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1;
1508 cur_pitch_val = s->min_pitch_val + get_bits(gb, s->pitch_nbits);
1509 cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1);
1510 if (s->last_acb_type == ACB_TYPE_NONE ||
1511 20 * abs(cur_pitch_val - s->last_pitch_val) >
1512 (cur_pitch_val + s->last_pitch_val))
1513 s->last_pitch_val = cur_pitch_val;
1515 /* pitch per block */
1516 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1517 int fac = n * 2 + 1;
1519 pitch[n] = (MUL16(fac, cur_pitch_val) +
1520 MUL16((n_blocks_x2 - fac), s->last_pitch_val) +
1521 frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2;
1524 /* "pitch-diff-per-sample" for calculation of pitch per sample */
1525 s->pitch_diff_sh16 =
1526 (cur_pitch_val - s->last_pitch_val) * (1 << 16) / MAX_FRAMESIZE;
1529 /* Global gain (if silence) and pitch-adaptive window coordinates */
1530 switch (frame_descs[bd_idx].fcb_type) {
1531 case FCB_TYPE_SILENCE:
1532 s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)];
1534 case FCB_TYPE_AW_PULSES:
1535 aw_parse_coords(s, gb, pitch);
1539 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1542 /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */
1543 switch (frame_descs[bd_idx].acb_type) {
1544 case ACB_TYPE_HAMMING: {
1545 /* Pitch is given per block. Per-block pitches are encoded as an
1546 * absolute value for the first block, and then delta values
1547 * relative to this value) for all subsequent blocks. The scale of
1548 * this pitch value is semi-logarithmic compared to its use in the
1549 * decoder, so we convert it to normal scale also. */
1551 t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2,
1552 t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1,
1553 t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1;
1556 block_pitch = get_bits(gb, s->block_pitch_nbits);
1558 block_pitch = last_block_pitch - s->block_delta_pitch_hrange +
1559 get_bits(gb, s->block_delta_pitch_nbits);
1560 /* Convert last_ so that any next delta is within _range */
1561 last_block_pitch = av_clip(block_pitch,
1562 s->block_delta_pitch_hrange,
1563 s->block_pitch_range -
1564 s->block_delta_pitch_hrange);
1566 /* Convert semi-log-style scale back to normal scale */
1567 if (block_pitch < t1) {
1568 bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch;
1571 if (block_pitch < t2) {
1573 (s->block_conv_table[1] << 2) + (block_pitch << 1);
1576 if (block_pitch < t3) {
1578 (s->block_conv_table[2] + block_pitch) << 2;
1580 bl_pitch_sh2 = s->block_conv_table[3] << 2;
1583 pitch[n] = bl_pitch_sh2 >> 2;
1587 case ACB_TYPE_ASYMMETRIC: {
1588 bl_pitch_sh2 = pitch[n] << 2;
1592 default: // ACB_TYPE_NONE has no pitch
1597 synth_block(s, gb, n, block_nsamples, bl_pitch_sh2,
1598 lsps, prev_lsps, &frame_descs[bd_idx],
1599 &excitation[n * block_nsamples],
1600 &synth[n * block_nsamples]);
1603 /* Averaging projection filter, if applicable. Else, just copy samples
1604 * from synthesis buffer */
1606 double i_lsps[MAX_LSPS];
1607 float lpcs[MAX_LSPS];
1609 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1610 i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n]));
1611 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1612 postfilter(s, synth, samples, 80, lpcs,
1613 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx],
1614 frame_descs[bd_idx].fcb_type, pitch[0]);
1616 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1617 i_lsps[n] = cos(lsps[n]);
1618 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1619 postfilter(s, &synth[80], &samples[80], 80, lpcs,
1620 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80],
1621 frame_descs[bd_idx].fcb_type, pitch[0]);
1623 memcpy(samples, synth, 160 * sizeof(synth[0]));
1625 /* Cache values for next frame */
1627 if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%)
1628 s->last_acb_type = frame_descs[bd_idx].acb_type;
1629 switch (frame_descs[bd_idx].acb_type) {
1631 s->last_pitch_val = 0;
1633 case ACB_TYPE_ASYMMETRIC:
1634 s->last_pitch_val = cur_pitch_val;
1636 case ACB_TYPE_HAMMING:
1637 s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1];
1645 * Ensure minimum value for first item, maximum value for last value,
1646 * proper spacing between each value and proper ordering.
1648 * @param lsps array of LSPs
1649 * @param num size of LSP array
1651 * @note basically a double version of #ff_acelp_reorder_lsf(), might be
1652 * useful to put in a generic location later on. Parts are also
1653 * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(),
1654 * which is in float.
1656 static void stabilize_lsps(double *lsps, int num)
1660 /* set minimum value for first, maximum value for last and minimum
1661 * spacing between LSF values.
1662 * Very similar to ff_set_min_dist_lsf(), but in double. */
1663 lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI);
1664 for (n = 1; n < num; n++)
1665 lsps[n] = FFMAX(lsps[n], lsps[n - 1] + 0.0125 * M_PI);
1666 lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI);
1668 /* reorder (looks like one-time / non-recursed bubblesort).
1669 * Very similar to ff_sort_nearly_sorted_floats(), but in double. */
1670 for (n = 1; n < num; n++) {
1671 if (lsps[n] < lsps[n - 1]) {
1672 for (m = 1; m < num; m++) {
1673 double tmp = lsps[m];
1674 for (l = m - 1; l >= 0; l--) {
1675 if (lsps[l] <= tmp) break;
1676 lsps[l + 1] = lsps[l];
1686 * Synthesize output samples for a single superframe. If we have any data
1687 * cached in s->sframe_cache, that will be used instead of whatever is loaded
1690 * WMA Voice superframes contain 3 frames, each containing 160 audio samples,
1691 * to give a total of 480 samples per frame. See #synth_frame() for frame
1692 * parsing. In addition to 3 frames, superframes can also contain the LSPs
1693 * (if these are globally specified for all frames (residually); they can
1694 * also be specified individually per-frame. See the s->has_residual_lsps
1695 * option), and can specify the number of samples encoded in this superframe
1696 * (if less than 480), usually used to prevent blanks at track boundaries.
1698 * @param ctx WMA Voice decoder context
1699 * @return 0 on success, <0 on error or 1 if there was not enough data to
1700 * fully parse the superframe
1702 static int synth_superframe(AVCodecContext *ctx, AVFrame *frame,
1705 WMAVoiceContext *s = ctx->priv_data;
1706 GetBitContext *gb = &s->gb, s_gb;
1707 int n, res, n_samples = MAX_SFRAMESIZE;
1708 double lsps[MAX_FRAMES][MAX_LSPS];
1709 const double *mean_lsf = s->lsps == 16 ?
1710 wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode];
1711 float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12];
1712 float synth[MAX_LSPS + MAX_SFRAMESIZE];
1715 memcpy(synth, s->synth_history,
1716 s->lsps * sizeof(*synth));
1717 memcpy(excitation, s->excitation_history,
1718 s->history_nsamples * sizeof(*excitation));
1720 if (s->sframe_cache_size > 0) {
1722 init_get_bits(gb, s->sframe_cache, s->sframe_cache_size);
1723 s->sframe_cache_size = 0;
1726 /* First bit is speech/music bit, it differentiates between WMAVoice
1727 * speech samples (the actual codec) and WMAVoice music samples, which
1728 * are really WMAPro-in-WMAVoice-superframes. I've never seen those in
1730 if (!get_bits1(gb)) {
1731 avpriv_request_sample(ctx, "WMAPro-in-WMAVoice");
1732 return AVERROR_PATCHWELCOME;
1735 /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */
1736 if (get_bits1(gb)) {
1737 if ((n_samples = get_bits(gb, 12)) > MAX_SFRAMESIZE) {
1738 av_log(ctx, AV_LOG_ERROR,
1739 "Superframe encodes > %d samples (%d), not allowed\n",
1740 MAX_SFRAMESIZE, n_samples);
1741 return AVERROR_INVALIDDATA;
1745 /* Parse LSPs, if global for the superframe (can also be per-frame). */
1746 if (s->has_residual_lsps) {
1747 double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2];
1749 for (n = 0; n < s->lsps; n++)
1750 prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n];
1752 if (s->lsps == 10) {
1753 dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1754 } else /* s->lsps == 16 */
1755 dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1757 for (n = 0; n < s->lsps; n++) {
1758 lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]);
1759 lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]);
1760 lsps[2][n] += mean_lsf[n];
1762 for (n = 0; n < 3; n++)
1763 stabilize_lsps(lsps[n], s->lsps);
1766 /* synth_superframe can run multiple times per packet
1767 * free potential previous frame */
1768 av_frame_unref(frame);
1770 /* get output buffer */
1771 frame->nb_samples = MAX_SFRAMESIZE;
1772 if ((res = ff_get_buffer(ctx, frame, 0)) < 0)
1774 frame->nb_samples = n_samples;
1775 samples = (float *)frame->data[0];
1777 /* Parse frames, optionally preceded by per-frame (independent) LSPs. */
1778 for (n = 0; n < 3; n++) {
1779 if (!s->has_residual_lsps) {
1782 if (s->lsps == 10) {
1783 dequant_lsp10i(gb, lsps[n]);
1784 } else /* s->lsps == 16 */
1785 dequant_lsp16i(gb, lsps[n]);
1787 for (m = 0; m < s->lsps; m++)
1788 lsps[n][m] += mean_lsf[m];
1789 stabilize_lsps(lsps[n], s->lsps);
1792 if ((res = synth_frame(ctx, gb, n,
1793 &samples[n * MAX_FRAMESIZE],
1794 lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1],
1795 &excitation[s->history_nsamples + n * MAX_FRAMESIZE],
1796 &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 if (get_bits_left(gb) < 0) {
1811 wmavoice_flush(ctx);
1812 return AVERROR_INVALIDDATA;
1817 /* Update history */
1818 memcpy(s->prev_lsps, lsps[2],
1819 s->lsps * sizeof(*s->prev_lsps));
1820 memcpy(s->synth_history, &synth[MAX_SFRAMESIZE],
1821 s->lsps * sizeof(*synth));
1822 memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE],
1823 s->history_nsamples * sizeof(*excitation));
1825 memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE],
1826 s->history_nsamples * sizeof(*s->zero_exc_pf));
1832 * Parse the packet header at the start of each packet (input data to this
1835 * @param s WMA Voice decoding context private data
1836 * @return <0 on error, nb_superframes on success.
1838 static int parse_packet_header(WMAVoiceContext *s)
1840 GetBitContext *gb = &s->gb;
1841 unsigned int res, n_superframes = 0;
1843 skip_bits(gb, 4); // packet sequence number
1844 s->has_residual_lsps = get_bits1(gb);
1846 if (get_bits_left(gb) < 6 + s->spillover_bitsize)
1847 return AVERROR_INVALIDDATA;
1849 res = get_bits(gb, 6); // number of superframes per packet
1850 // (minus first one if there is spillover)
1851 n_superframes += res;
1852 } while (res == 0x3F);
1853 s->spillover_nbits = get_bits(gb, s->spillover_bitsize);
1855 return get_bits_left(gb) >= 0 ? n_superframes : AVERROR_INVALIDDATA;
1859 * Copy (unaligned) bits from gb/data/size to pb.
1861 * @param pb target buffer to copy bits into
1862 * @param data source buffer to copy bits from
1863 * @param size size of the source data, in bytes
1864 * @param gb bit I/O context specifying the current position in the source.
1865 * data. This function might use this to align the bit position to
1866 * a whole-byte boundary before calling #avpriv_copy_bits() on aligned
1868 * @param nbits the amount of bits to copy from source to target
1870 * @note after calling this function, the current position in the input bit
1871 * I/O context is undefined.
1873 static void copy_bits(PutBitContext *pb,
1874 const uint8_t *data, int size,
1875 GetBitContext *gb, int nbits)
1877 int rmn_bytes, rmn_bits;
1879 rmn_bits = rmn_bytes = get_bits_left(gb);
1880 if (rmn_bits < nbits)
1882 if (nbits > pb->size_in_bits - put_bits_count(pb))
1884 rmn_bits &= 7; rmn_bytes >>= 3;
1885 if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0)
1886 put_bits(pb, rmn_bits, get_bits(gb, rmn_bits));
1887 avpriv_copy_bits(pb, data + size - rmn_bytes,
1888 FFMIN(nbits - rmn_bits, rmn_bytes << 3));
1892 * Packet decoding: a packet is anything that the (ASF) demuxer contains,
1893 * and we expect that the demuxer / application provides it to us as such
1894 * (else you'll probably get garbage as output). Every packet has a size of
1895 * ctx->block_align bytes, starts with a packet header (see
1896 * #parse_packet_header()), and then a series of superframes. Superframe
1897 * boundaries may exceed packets, i.e. superframes can split data over
1898 * multiple (two) packets.
1900 * For more information about frames, see #synth_superframe().
1902 static int wmavoice_decode_packet(AVCodecContext *ctx, void *data,
1903 int *got_frame_ptr, AVPacket *avpkt)
1905 WMAVoiceContext *s = ctx->priv_data;
1906 GetBitContext *gb = &s->gb;
1909 /* Packets are sometimes a multiple of ctx->block_align, with a packet
1910 * header at each ctx->block_align bytes. However, FFmpeg's ASF demuxer
1911 * feeds us ASF packets, which may concatenate multiple "codec" packets
1912 * in a single "muxer" packet, so we artificially emulate that by
1913 * capping the packet size at ctx->block_align. */
1914 for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align);
1915 init_get_bits8(&s->gb, avpkt->data, size);
1917 /* size == ctx->block_align is used to indicate whether we are dealing with
1918 * a new packet or a packet of which we already read the packet header
1920 if (!(size % ctx->block_align)) { // new packet header
1922 s->spillover_nbits = 0;
1923 s->nb_superframes = 0;
1925 if ((res = parse_packet_header(s)) < 0)
1927 s->nb_superframes = res;
1930 /* If the packet header specifies a s->spillover_nbits, then we want
1931 * to push out all data of the previous packet (+ spillover) before
1932 * continuing to parse new superframes in the current packet. */
1933 if (s->sframe_cache_size > 0) {
1934 int cnt = get_bits_count(gb);
1935 if (cnt + s->spillover_nbits > avpkt->size * 8) {
1936 s->spillover_nbits = avpkt->size * 8 - cnt;
1938 copy_bits(&s->pb, avpkt->data, size, gb, s->spillover_nbits);
1939 flush_put_bits(&s->pb);
1940 s->sframe_cache_size += s->spillover_nbits;
1941 if ((res = synth_superframe(ctx, data, got_frame_ptr)) == 0 &&
1943 cnt += s->spillover_nbits;
1944 s->skip_bits_next = cnt & 7;
1948 skip_bits_long (gb, s->spillover_nbits - cnt +
1949 get_bits_count(gb)); // resync
1950 } else if (s->spillover_nbits) {
1951 skip_bits_long(gb, s->spillover_nbits); // resync
1953 } else if (s->skip_bits_next)
1954 skip_bits(gb, s->skip_bits_next);
1956 /* Try parsing superframes in current packet */
1957 s->sframe_cache_size = 0;
1958 s->skip_bits_next = 0;
1959 pos = get_bits_left(gb);
1960 if (s->nb_superframes-- == 0) {
1963 } else if (s->nb_superframes > 0) {
1964 if ((res = synth_superframe(ctx, data, got_frame_ptr)) < 0) {
1966 } else if (*got_frame_ptr) {
1967 int cnt = get_bits_count(gb);
1968 s->skip_bits_next = cnt & 7;
1972 } else if ((s->sframe_cache_size = pos) > 0) {
1973 /* ... cache it for spillover in next packet */
1974 init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE);
1975 copy_bits(&s->pb, avpkt->data, size, gb, s->sframe_cache_size);
1976 // FIXME bad - just copy bytes as whole and add use the
1977 // skip_bits_next field
1983 static av_cold int wmavoice_decode_end(AVCodecContext *ctx)
1985 WMAVoiceContext *s = ctx->priv_data;
1988 ff_rdft_end(&s->rdft);
1989 ff_rdft_end(&s->irdft);
1990 ff_dct_end(&s->dct);
1991 ff_dct_end(&s->dst);
1997 AVCodec ff_wmavoice_decoder = {
1999 .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"),
2000 .type = AVMEDIA_TYPE_AUDIO,
2001 .id = AV_CODEC_ID_WMAVOICE,
2002 .priv_data_size = sizeof(WMAVoiceContext),
2003 .init = wmavoice_decode_init,
2004 .close = wmavoice_decode_end,
2005 .decode = wmavoice_decode_packet,
2006 .capabilities = AV_CODEC_CAP_SUBFRAMES | AV_CODEC_CAP_DR1 | AV_CODEC_CAP_DELAY,
2007 .caps_internal = FF_CODEC_CAP_INIT_CLEANUP,
2008 .flush = wmavoice_flush,