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"
37 #include "wmavoice_data.h"
38 #include "celp_filters.h"
39 #include "acelp_vectors.h"
40 #include "acelp_filters.h"
46 #define MAX_BLOCKS 8 ///< maximum number of blocks per frame
47 #define MAX_LSPS 16 ///< maximum filter order
48 #define MAX_LSPS_ALIGN16 16 ///< same as #MAX_LSPS; needs to be multiple
49 ///< of 16 for ASM input buffer alignment
50 #define MAX_FRAMES 3 ///< maximum number of frames per superframe
51 #define MAX_FRAMESIZE 160 ///< maximum number of samples per frame
52 #define MAX_SIGNAL_HISTORY 416 ///< maximum excitation signal history
53 #define MAX_SFRAMESIZE (MAX_FRAMESIZE * MAX_FRAMES)
54 ///< maximum number of samples per superframe
55 #define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that
56 ///< was split over two packets
57 #define VLC_NBITS 6 ///< number of bits to read per VLC iteration
60 * Frame type VLC coding.
62 static VLC frame_type_vlc;
65 * Adaptive codebook types.
68 ACB_TYPE_NONE = 0, ///< no adaptive codebook (only hardcoded fixed)
69 ACB_TYPE_ASYMMETRIC = 1, ///< adaptive codebook with per-frame pitch, which
70 ///< we interpolate to get a per-sample pitch.
71 ///< Signal is generated using an asymmetric sinc
73 ///< @note see #wmavoice_ipol1_coeffs
74 ACB_TYPE_HAMMING = 2 ///< Per-block pitch with signal generation using
75 ///< a Hamming sinc window function
76 ///< @note see #wmavoice_ipol2_coeffs
80 * Fixed codebook types.
83 FCB_TYPE_SILENCE = 0, ///< comfort noise during silence
84 ///< generated from a hardcoded (fixed) codebook
85 ///< with per-frame (low) gain values
86 FCB_TYPE_HARDCODED = 1, ///< hardcoded (fixed) codebook with per-block
88 FCB_TYPE_AW_PULSES = 2, ///< Pitch-adaptive window (AW) pulse signals,
89 ///< used in particular for low-bitrate streams
90 FCB_TYPE_EXC_PULSES = 3, ///< Innovation (fixed) codebook pulse sets in
91 ///< combinations of either single pulses or
96 * Description of frame types.
98 static const struct frame_type_desc {
99 uint8_t n_blocks; ///< amount of blocks per frame (each block
100 ///< (contains 160/#n_blocks samples)
101 uint8_t log_n_blocks; ///< log2(#n_blocks)
102 uint8_t acb_type; ///< Adaptive codebook type (ACB_TYPE_*)
103 uint8_t fcb_type; ///< Fixed codebook type (FCB_TYPE_*)
104 uint8_t dbl_pulses; ///< how many pulse vectors have pulse pairs
105 ///< (rather than just one single pulse)
106 ///< only if #fcb_type == #FCB_TYPE_EXC_PULSES
107 uint16_t frame_size; ///< the amount of bits that make up the block
108 ///< data (per frame)
109 } frame_descs[17] = {
110 { 1, 0, ACB_TYPE_NONE, FCB_TYPE_SILENCE, 0, 0 },
111 { 2, 1, ACB_TYPE_NONE, FCB_TYPE_HARDCODED, 0, 28 },
112 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_AW_PULSES, 0, 46 },
113 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 80 },
114 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 104 },
115 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 0, 108 },
116 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 132 },
117 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 168 },
118 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 64 },
119 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 80 },
120 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 104 },
121 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 108 },
122 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 132 },
123 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 168 },
124 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 176 },
125 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 208 },
126 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 256 }
130 * WMA Voice decoding context.
132 typedef struct WMAVoiceContext {
134 * @name Global values specified in the stream header / extradata or used all over.
137 GetBitContext gb; ///< packet bitreader. During decoder init,
138 ///< it contains the extradata from the
139 ///< demuxer. During decoding, it contains
141 int8_t vbm_tree[25]; ///< converts VLC codes to frame type
143 int spillover_bitsize; ///< number of bits used to specify
144 ///< #spillover_nbits in the packet header
145 ///< = ceil(log2(ctx->block_align << 3))
146 int history_nsamples; ///< number of samples in history for signal
147 ///< prediction (through ACB)
149 /* postfilter specific values */
150 int do_apf; ///< whether to apply the averaged
151 ///< projection filter (APF)
152 int denoise_strength; ///< strength of denoising in Wiener filter
154 int denoise_tilt_corr; ///< Whether to apply tilt correction to the
155 ///< Wiener filter coefficients (postfilter)
156 int dc_level; ///< Predicted amount of DC noise, based
157 ///< on which a DC removal filter is used
159 int lsps; ///< number of LSPs per frame [10 or 16]
160 int lsp_q_mode; ///< defines quantizer defaults [0, 1]
161 int lsp_def_mode; ///< defines different sets of LSP defaults
163 int frame_lsp_bitsize; ///< size (in bits) of LSPs, when encoded
164 ///< per-frame (independent coding)
165 int sframe_lsp_bitsize; ///< size (in bits) of LSPs, when encoded
166 ///< per superframe (residual coding)
168 int min_pitch_val; ///< base value for pitch parsing code
169 int max_pitch_val; ///< max value + 1 for pitch parsing
170 int pitch_nbits; ///< number of bits used to specify the
171 ///< pitch value in the frame header
172 int block_pitch_nbits; ///< number of bits used to specify the
173 ///< first block's pitch value
174 int block_pitch_range; ///< range of the block pitch
175 int block_delta_pitch_nbits; ///< number of bits used to specify the
176 ///< delta pitch between this and the last
177 ///< block's pitch value, used in all but
179 int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is
180 ///< from -this to +this-1)
181 uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale
187 * @name Packet values specified in the packet header or related to a packet.
189 * A packet is considered to be a single unit of data provided to this
190 * decoder by the demuxer.
193 int spillover_nbits; ///< number of bits of the previous packet's
194 ///< last superframe preceding this
195 ///< packet's first full superframe (useful
196 ///< for re-synchronization also)
197 int has_residual_lsps; ///< if set, superframes contain one set of
198 ///< LSPs that cover all frames, encoded as
199 ///< independent and residual LSPs; if not
200 ///< set, each frame contains its own, fully
201 ///< independent, LSPs
202 int skip_bits_next; ///< number of bits to skip at the next call
203 ///< to #wmavoice_decode_packet() (since
204 ///< they're part of the previous superframe)
206 uint8_t sframe_cache[SFRAME_CACHE_MAXSIZE + AV_INPUT_BUFFER_PADDING_SIZE];
207 ///< cache for superframe data split over
208 ///< multiple packets
209 int sframe_cache_size; ///< set to >0 if we have data from an
210 ///< (incomplete) superframe from a previous
211 ///< packet that spilled over in the current
212 ///< packet; specifies the amount of bits in
214 PutBitContext pb; ///< bitstream writer for #sframe_cache
219 * @name Frame and superframe values
220 * Superframe and frame data - these can change from frame to frame,
221 * although some of them do in that case serve as a cache / history for
222 * the next frame or superframe.
225 double prev_lsps[MAX_LSPS]; ///< LSPs of the last frame of the previous
227 int last_pitch_val; ///< pitch value of the previous frame
228 int last_acb_type; ///< frame type [0-2] of the previous frame
229 int pitch_diff_sh16; ///< ((cur_pitch_val - #last_pitch_val)
230 ///< << 16) / #MAX_FRAMESIZE
231 float silence_gain; ///< set for use in blocks if #ACB_TYPE_NONE
233 int aw_idx_is_ext; ///< whether the AW index was encoded in
234 ///< 8 bits (instead of 6)
235 int aw_pulse_range; ///< the range over which #aw_pulse_set1()
236 ///< can apply the pulse, relative to the
237 ///< value in aw_first_pulse_off. The exact
238 ///< position of the first AW-pulse is within
239 ///< [pulse_off, pulse_off + this], and
240 ///< depends on bitstream values; [16 or 24]
241 int aw_n_pulses[2]; ///< number of AW-pulses in each block; note
242 ///< that this number can be negative (in
243 ///< which case it basically means "zero")
244 int aw_first_pulse_off[2]; ///< index of first sample to which to
245 ///< apply AW-pulses, or -0xff if unset
246 int aw_next_pulse_off_cache; ///< the position (relative to start of the
247 ///< second block) at which pulses should
248 ///< start to be positioned, serves as a
249 ///< cache for pitch-adaptive window pulses
252 int frame_cntr; ///< current frame index [0 - 0xFFFE]; is
253 ///< only used for comfort noise in #pRNG()
254 int nb_superframes; ///< number of superframes in current packet
255 float gain_pred_err[6]; ///< cache for gain prediction
256 float excitation_history[MAX_SIGNAL_HISTORY];
257 ///< cache of the signal of previous
258 ///< superframes, used as a history for
259 ///< signal generation
260 float synth_history[MAX_LSPS]; ///< see #excitation_history
264 * @name Postfilter values
266 * Variables used for postfilter implementation, mostly history for
267 * smoothing and so on, and context variables for FFT/iFFT.
270 RDFTContext rdft, irdft; ///< contexts for FFT-calculation in the
271 ///< postfilter (for denoise filter)
272 DCTContext dct, dst; ///< contexts for phase shift (in Hilbert
273 ///< transform, part of postfilter)
274 float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi]
276 float postfilter_agc; ///< gain control memory, used in
277 ///< #adaptive_gain_control()
278 float dcf_mem[2]; ///< DC filter history
279 float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE];
280 ///< zero filter output (i.e. excitation)
282 float denoise_filter_cache[MAX_FRAMESIZE];
283 int denoise_filter_cache_size; ///< samples in #denoise_filter_cache
284 DECLARE_ALIGNED(32, float, tilted_lpcs_pf)[0x80];
285 ///< aligned buffer for LPC tilting
286 DECLARE_ALIGNED(32, float, denoise_coeffs_pf)[0x80];
287 ///< aligned buffer for denoise coefficients
288 DECLARE_ALIGNED(32, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16];
289 ///< aligned buffer for postfilter speech
297 * Set up the variable bit mode (VBM) tree from container extradata.
298 * @param gb bit I/O context.
299 * The bit context (s->gb) should be loaded with byte 23-46 of the
300 * container extradata (i.e. the ones containing the VBM tree).
301 * @param vbm_tree pointer to array to which the decoded VBM tree will be
303 * @return 0 on success, <0 on error.
305 static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25])
307 int cntr[8] = { 0 }, n, res;
309 memset(vbm_tree, 0xff, sizeof(vbm_tree[0]) * 25);
310 for (n = 0; n < 17; n++) {
311 res = get_bits(gb, 3);
312 if (cntr[res] > 3) // should be >= 3 + (res == 7))
314 vbm_tree[res * 3 + cntr[res]++] = n;
319 static av_cold void wmavoice_init_static_data(AVCodec *codec)
321 static const uint8_t bits[] = {
324 10, 10, 10, 12, 12, 12,
327 static const uint16_t codes[] = {
328 0x0000, 0x0001, 0x0002, // 00/01/10
329 0x000c, 0x000d, 0x000e, // 11+00/01/10
330 0x003c, 0x003d, 0x003e, // 1111+00/01/10
331 0x00fc, 0x00fd, 0x00fe, // 111111+00/01/10
332 0x03fc, 0x03fd, 0x03fe, // 11111111+00/01/10
333 0x0ffc, 0x0ffd, 0x0ffe, // 1111111111+00/01/10
334 0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx
337 INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits),
338 bits, 1, 1, codes, 2, 2, 132);
341 static av_cold void wmavoice_flush(AVCodecContext *ctx)
343 WMAVoiceContext *s = ctx->priv_data;
346 s->postfilter_agc = 0;
347 s->sframe_cache_size = 0;
348 s->skip_bits_next = 0;
349 for (n = 0; n < s->lsps; n++)
350 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
351 memset(s->excitation_history, 0,
352 sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY);
353 memset(s->synth_history, 0,
354 sizeof(*s->synth_history) * MAX_LSPS);
355 memset(s->gain_pred_err, 0,
356 sizeof(s->gain_pred_err));
359 memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0,
360 sizeof(*s->synth_filter_out_buf) * s->lsps);
361 memset(s->dcf_mem, 0,
362 sizeof(*s->dcf_mem) * 2);
363 memset(s->zero_exc_pf, 0,
364 sizeof(*s->zero_exc_pf) * s->history_nsamples);
365 memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache));
370 * Set up decoder with parameters from demuxer (extradata etc.).
372 static av_cold int wmavoice_decode_init(AVCodecContext *ctx)
374 int n, flags, pitch_range, lsp16_flag;
375 WMAVoiceContext *s = ctx->priv_data;
379 * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c),
380 * - byte 19-22: flags field (annoyingly in LE; see below for known
382 * - byte 23-46: variable bitmode tree (really just 17 * 3 bits,
385 if (ctx->extradata_size != 46) {
386 av_log(ctx, AV_LOG_ERROR,
387 "Invalid extradata size %d (should be 46)\n",
388 ctx->extradata_size);
389 return AVERROR_INVALIDDATA;
391 if (ctx->block_align <= 0) {
392 av_log(ctx, AV_LOG_ERROR, "Invalid block alignment %d.\n", ctx->block_align);
393 return AVERROR_INVALIDDATA;
396 flags = AV_RL32(ctx->extradata + 18);
397 s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align);
398 s->do_apf = flags & 0x1;
400 ff_rdft_init(&s->rdft, 7, DFT_R2C);
401 ff_rdft_init(&s->irdft, 7, IDFT_C2R);
402 ff_dct_init(&s->dct, 6, DCT_I);
403 ff_dct_init(&s->dst, 6, DST_I);
405 ff_sine_window_init(s->cos, 256);
406 memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0]));
407 for (n = 0; n < 255; n++) {
408 s->sin[n] = -s->sin[510 - n];
409 s->cos[510 - n] = s->cos[n];
412 s->denoise_strength = (flags >> 2) & 0xF;
413 if (s->denoise_strength >= 12) {
414 av_log(ctx, AV_LOG_ERROR,
415 "Invalid denoise filter strength %d (max=11)\n",
416 s->denoise_strength);
417 return AVERROR_INVALIDDATA;
419 s->denoise_tilt_corr = !!(flags & 0x40);
420 s->dc_level = (flags >> 7) & 0xF;
421 s->lsp_q_mode = !!(flags & 0x2000);
422 s->lsp_def_mode = !!(flags & 0x4000);
423 lsp16_flag = flags & 0x1000;
426 s->frame_lsp_bitsize = 34;
427 s->sframe_lsp_bitsize = 60;
430 s->frame_lsp_bitsize = 24;
431 s->sframe_lsp_bitsize = 48;
433 for (n = 0; n < s->lsps; n++)
434 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
436 init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3);
437 if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) {
438 av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n");
439 return AVERROR_INVALIDDATA;
442 s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8;
443 s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8;
444 pitch_range = s->max_pitch_val - s->min_pitch_val;
445 if (pitch_range <= 0) {
446 av_log(ctx, AV_LOG_ERROR, "Invalid pitch range; broken extradata?\n");
447 return AVERROR_INVALIDDATA;
449 s->pitch_nbits = av_ceil_log2(pitch_range);
450 s->last_pitch_val = 40;
451 s->last_acb_type = ACB_TYPE_NONE;
452 s->history_nsamples = s->max_pitch_val + 8;
454 if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) {
455 int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8,
456 max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8;
458 av_log(ctx, AV_LOG_ERROR,
459 "Unsupported samplerate %d (min=%d, max=%d)\n",
460 ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz
462 return AVERROR(ENOSYS);
465 s->block_conv_table[0] = s->min_pitch_val;
466 s->block_conv_table[1] = (pitch_range * 25) >> 6;
467 s->block_conv_table[2] = (pitch_range * 44) >> 6;
468 s->block_conv_table[3] = s->max_pitch_val - 1;
469 s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF;
470 if (s->block_delta_pitch_hrange <= 0) {
471 av_log(ctx, AV_LOG_ERROR, "Invalid delta pitch hrange; broken extradata?\n");
472 return AVERROR_INVALIDDATA;
474 s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange);
475 s->block_pitch_range = s->block_conv_table[2] +
476 s->block_conv_table[3] + 1 +
477 2 * (s->block_conv_table[1] - 2 * s->min_pitch_val);
478 s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range);
481 ctx->channel_layout = AV_CH_LAYOUT_MONO;
482 ctx->sample_fmt = AV_SAMPLE_FMT_FLT;
488 * @name Postfilter functions
489 * Postfilter functions (gain control, wiener denoise filter, DC filter,
490 * kalman smoothening, plus surrounding code to wrap it)
494 * Adaptive gain control (as used in postfilter).
496 * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except
497 * that the energy here is calculated using sum(abs(...)), whereas the
498 * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)).
500 * @param out output buffer for filtered samples
501 * @param in input buffer containing the samples as they are after the
502 * postfilter steps so far
503 * @param speech_synth input buffer containing speech synth before postfilter
504 * @param size input buffer size
505 * @param alpha exponential filter factor
506 * @param gain_mem pointer to filter memory (single float)
508 static void adaptive_gain_control(float *out, const float *in,
509 const float *speech_synth,
510 int size, float alpha, float *gain_mem)
513 float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor;
514 float mem = *gain_mem;
516 for (i = 0; i < size; i++) {
517 speech_energy += fabsf(speech_synth[i]);
518 postfilter_energy += fabsf(in[i]);
520 gain_scale_factor = postfilter_energy == 0.0 ? 0.0 :
521 (1.0 - alpha) * speech_energy / postfilter_energy;
523 for (i = 0; i < size; i++) {
524 mem = alpha * mem + gain_scale_factor;
525 out[i] = in[i] * mem;
532 * Kalman smoothing function.
534 * This function looks back pitch +/- 3 samples back into history to find
535 * the best fitting curve (that one giving the optimal gain of the two
536 * signals, i.e. the highest dot product between the two), and then
537 * uses that signal history to smoothen the output of the speech synthesis
540 * @param s WMA Voice decoding context
541 * @param pitch pitch of the speech signal
542 * @param in input speech signal
543 * @param out output pointer for smoothened signal
544 * @param size input/output buffer size
546 * @returns -1 if no smoothening took place, e.g. because no optimal
547 * fit could be found, or 0 on success.
549 static int kalman_smoothen(WMAVoiceContext *s, int pitch,
550 const float *in, float *out, int size)
553 float optimal_gain = 0, dot;
554 const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)],
555 *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)],
556 *best_hist_ptr = NULL;
558 /* find best fitting point in history */
560 dot = avpriv_scalarproduct_float_c(in, ptr, size);
561 if (dot > optimal_gain) {
565 } while (--ptr >= end);
567 if (optimal_gain <= 0)
569 dot = avpriv_scalarproduct_float_c(best_hist_ptr, best_hist_ptr, size);
570 if (dot <= 0) // would be 1.0
573 if (optimal_gain <= dot) {
574 dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000
578 /* actual smoothing */
579 for (n = 0; n < size; n++)
580 out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]);
586 * Get the tilt factor of a formant filter from its transfer function
587 * @see #tilt_factor() in amrnbdec.c, which does essentially the same,
588 * but somehow (??) it does a speech synthesis filter in the
589 * middle, which is missing here
591 * @param lpcs LPC coefficients
592 * @param n_lpcs Size of LPC buffer
593 * @returns the tilt factor
595 static float tilt_factor(const float *lpcs, int n_lpcs)
599 rh0 = 1.0 + avpriv_scalarproduct_float_c(lpcs, lpcs, n_lpcs);
600 rh1 = lpcs[0] + avpriv_scalarproduct_float_c(lpcs, &lpcs[1], n_lpcs - 1);
606 * Derive denoise filter coefficients (in real domain) from the LPCs.
608 static void calc_input_response(WMAVoiceContext *s, float *lpcs,
609 int fcb_type, float *coeffs, int remainder)
611 float last_coeff, min = 15.0, max = -15.0;
612 float irange, angle_mul, gain_mul, range, sq;
615 /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */
616 s->rdft.rdft_calc(&s->rdft, lpcs);
617 #define log_range(var, assign) do { \
618 float tmp = log10f(assign); var = tmp; \
619 max = FFMAX(max, tmp); min = FFMIN(min, tmp); \
621 log_range(last_coeff, lpcs[1] * lpcs[1]);
622 for (n = 1; n < 64; n++)
623 log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] +
624 lpcs[n * 2 + 1] * lpcs[n * 2 + 1]);
625 log_range(lpcs[0], lpcs[0] * lpcs[0]);
628 lpcs[64] = last_coeff;
630 /* Now, use this spectrum to pick out these frequencies with higher
631 * (relative) power/energy (which we then take to be "not noise"),
632 * and set up a table (still in lpc[]) of (relative) gains per frequency.
633 * These frequencies will be maintained, while others ("noise") will be
634 * decreased in the filter output. */
635 irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63]
636 gain_mul = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) :
638 angle_mul = gain_mul * (8.0 * M_LN10 / M_PI);
639 for (n = 0; n <= 64; n++) {
642 idx = FFMAX(0, lrint((max - lpcs[n]) * irange) - 1);
643 pwr = wmavoice_denoise_power_table[s->denoise_strength][idx];
644 lpcs[n] = angle_mul * pwr;
646 /* 70.57 =~ 1/log10(1.0331663) */
647 idx = (pwr * gain_mul - 0.0295) * 70.570526123;
648 if (idx > 127) { // fall back if index falls outside table range
649 coeffs[n] = wmavoice_energy_table[127] *
650 powf(1.0331663, idx - 127);
652 coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)];
655 /* calculate the Hilbert transform of the gains, which we do (since this
656 * is a sine input) by doing a phase shift (in theory, H(sin())=cos()).
657 * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the
658 * "moment" of the LPCs in this filter. */
659 s->dct.dct_calc(&s->dct, lpcs);
660 s->dst.dct_calc(&s->dst, lpcs);
662 /* Split out the coefficient indexes into phase/magnitude pairs */
663 idx = 255 + av_clip(lpcs[64], -255, 255);
664 coeffs[0] = coeffs[0] * s->cos[idx];
665 idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255);
666 last_coeff = coeffs[64] * s->cos[idx];
668 idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255);
669 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
670 coeffs[n * 2] = coeffs[n] * s->cos[idx];
674 idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255);
675 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
676 coeffs[n * 2] = coeffs[n] * s->cos[idx];
678 coeffs[1] = last_coeff;
680 /* move into real domain */
681 s->irdft.rdft_calc(&s->irdft, coeffs);
683 /* tilt correction and normalize scale */
684 memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder));
685 if (s->denoise_tilt_corr) {
688 coeffs[remainder - 1] = 0;
689 ff_tilt_compensation(&tilt_mem,
690 -1.8 * tilt_factor(coeffs, remainder - 1),
693 sq = (1.0 / 64.0) * sqrtf(1 / avpriv_scalarproduct_float_c(coeffs, coeffs,
695 for (n = 0; n < remainder; n++)
700 * This function applies a Wiener filter on the (noisy) speech signal as
701 * a means to denoise it.
703 * - take RDFT of LPCs to get the power spectrum of the noise + speech;
704 * - using this power spectrum, calculate (for each frequency) the Wiener
705 * filter gain, which depends on the frequency power and desired level
706 * of noise subtraction (when set too high, this leads to artifacts)
707 * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse
709 * - by doing a phase shift, calculate the Hilbert transform of this array
710 * of per-frequency filter-gains to get the filtering coefficients;
711 * - smoothen/normalize/de-tilt these filter coefficients as desired;
712 * - take RDFT of noisy sound, apply the coefficients and take its IRDFT
713 * to get the denoised speech signal;
714 * - the leftover (i.e. output of the IRDFT on denoised speech data beyond
715 * the frame boundary) are saved and applied to subsequent frames by an
716 * overlap-add method (otherwise you get clicking-artifacts).
718 * @param s WMA Voice decoding context
719 * @param fcb_type Frame (codebook) type
720 * @param synth_pf input: the noisy speech signal, output: denoised speech
721 * data; should be 16-byte aligned (for ASM purposes)
722 * @param size size of the speech data
723 * @param lpcs LPCs used to synthesize this frame's speech data
725 static void wiener_denoise(WMAVoiceContext *s, int fcb_type,
726 float *synth_pf, int size,
729 int remainder, lim, n;
731 if (fcb_type != FCB_TYPE_SILENCE) {
732 float *tilted_lpcs = s->tilted_lpcs_pf,
733 *coeffs = s->denoise_coeffs_pf, tilt_mem = 0;
735 tilted_lpcs[0] = 1.0;
736 memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps);
737 memset(&tilted_lpcs[s->lsps + 1], 0,
738 sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1));
739 ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps),
740 tilted_lpcs, s->lsps + 2);
742 /* The IRDFT output (127 samples for 7-bit filter) beyond the frame
743 * size is applied to the next frame. All input beyond this is zero,
744 * and thus all output beyond this will go towards zero, hence we can
745 * limit to min(size-1, 127-size) as a performance consideration. */
746 remainder = FFMIN(127 - size, size - 1);
747 calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder);
749 /* apply coefficients (in frequency spectrum domain), i.e. complex
750 * number multiplication */
751 memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size));
752 s->rdft.rdft_calc(&s->rdft, synth_pf);
753 s->rdft.rdft_calc(&s->rdft, coeffs);
754 synth_pf[0] *= coeffs[0];
755 synth_pf[1] *= coeffs[1];
756 for (n = 1; n < 64; n++) {
757 float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1];
758 synth_pf[n * 2] = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1];
759 synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1];
761 s->irdft.rdft_calc(&s->irdft, synth_pf);
764 /* merge filter output with the history of previous runs */
765 if (s->denoise_filter_cache_size) {
766 lim = FFMIN(s->denoise_filter_cache_size, size);
767 for (n = 0; n < lim; n++)
768 synth_pf[n] += s->denoise_filter_cache[n];
769 s->denoise_filter_cache_size -= lim;
770 memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size],
771 sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size);
774 /* move remainder of filter output into a cache for future runs */
775 if (fcb_type != FCB_TYPE_SILENCE) {
776 lim = FFMIN(remainder, s->denoise_filter_cache_size);
777 for (n = 0; n < lim; n++)
778 s->denoise_filter_cache[n] += synth_pf[size + n];
779 if (lim < remainder) {
780 memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim],
781 sizeof(s->denoise_filter_cache[0]) * (remainder - lim));
782 s->denoise_filter_cache_size = remainder;
788 * Averaging projection filter, the postfilter used in WMAVoice.
790 * This uses the following steps:
791 * - A zero-synthesis filter (generate excitation from synth signal)
792 * - Kalman smoothing on excitation, based on pitch
793 * - Re-synthesized smoothened output
794 * - Iterative Wiener denoise filter
795 * - Adaptive gain filter
798 * @param s WMAVoice decoding context
799 * @param synth Speech synthesis output (before postfilter)
800 * @param samples Output buffer for filtered samples
801 * @param size Buffer size of synth & samples
802 * @param lpcs Generated LPCs used for speech synthesis
803 * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned)
804 * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses)
805 * @param pitch Pitch of the input signal
807 static void postfilter(WMAVoiceContext *s, const float *synth,
808 float *samples, int size,
809 const float *lpcs, float *zero_exc_pf,
810 int fcb_type, int pitch)
812 float synth_filter_in_buf[MAX_FRAMESIZE / 2],
813 *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16],
814 *synth_filter_in = zero_exc_pf;
816 av_assert0(size <= MAX_FRAMESIZE / 2);
818 /* generate excitation from input signal */
819 ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps);
821 if (fcb_type >= FCB_TYPE_AW_PULSES &&
822 !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size))
823 synth_filter_in = synth_filter_in_buf;
825 /* re-synthesize speech after smoothening, and keep history */
826 ff_celp_lp_synthesis_filterf(synth_pf, lpcs,
827 synth_filter_in, size, s->lsps);
828 memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps],
829 sizeof(synth_pf[0]) * s->lsps);
831 wiener_denoise(s, fcb_type, synth_pf, size, lpcs);
833 adaptive_gain_control(samples, synth_pf, synth, size, 0.99,
836 if (s->dc_level > 8) {
837 /* remove ultra-low frequency DC noise / highpass filter;
838 * coefficients are identical to those used in SIPR decoding,
839 * and very closely resemble those used in AMR-NB decoding. */
840 ff_acelp_apply_order_2_transfer_function(samples, samples,
841 (const float[2]) { -1.99997, 1.0 },
842 (const float[2]) { -1.9330735188, 0.93589198496 },
843 0.93980580475, s->dcf_mem, size);
852 * @param lsps output pointer to the array that will hold the LSPs
853 * @param num number of LSPs to be dequantized
854 * @param values quantized values, contains n_stages values
855 * @param sizes range (i.e. max value) of each quantized value
856 * @param n_stages number of dequantization runs
857 * @param table dequantization table to be used
858 * @param mul_q LSF multiplier
859 * @param base_q base (lowest) LSF values
861 static void dequant_lsps(double *lsps, int num,
862 const uint16_t *values,
863 const uint16_t *sizes,
864 int n_stages, const uint8_t *table,
866 const double *base_q)
870 memset(lsps, 0, num * sizeof(*lsps));
871 for (n = 0; n < n_stages; n++) {
872 const uint8_t *t_off = &table[values[n] * num];
873 double base = base_q[n], mul = mul_q[n];
875 for (m = 0; m < num; m++)
876 lsps[m] += base + mul * t_off[m];
878 table += sizes[n] * num;
883 * @name LSP dequantization routines
884 * LSP dequantization routines, for 10/16LSPs and independent/residual coding.
885 * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;
886 * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits.
890 * Parse 10 independently-coded LSPs.
892 static void dequant_lsp10i(GetBitContext *gb, double *lsps)
894 static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 };
895 static const double mul_lsf[4] = {
896 5.2187144800e-3, 1.4626986422e-3,
897 9.6179549166e-4, 1.1325736225e-3
899 static const double base_lsf[4] = {
900 M_PI * -2.15522e-1, M_PI * -6.1646e-2,
901 M_PI * -3.3486e-2, M_PI * -5.7408e-2
905 v[0] = get_bits(gb, 8);
906 v[1] = get_bits(gb, 6);
907 v[2] = get_bits(gb, 5);
908 v[3] = get_bits(gb, 5);
910 dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i,
915 * Parse 10 independently-coded LSPs, and then derive the tables to
916 * generate LSPs for the other frames from them (residual coding).
918 static void dequant_lsp10r(GetBitContext *gb,
919 double *i_lsps, const double *old,
920 double *a1, double *a2, int q_mode)
922 static const uint16_t vec_sizes[3] = { 128, 64, 64 };
923 static const double mul_lsf[3] = {
924 2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3
926 static const double base_lsf[3] = {
927 M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2
929 const float (*ipol_tab)[2][10] = q_mode ?
930 wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a;
931 uint16_t interpol, v[3];
934 dequant_lsp10i(gb, i_lsps);
936 interpol = get_bits(gb, 5);
937 v[0] = get_bits(gb, 7);
938 v[1] = get_bits(gb, 6);
939 v[2] = get_bits(gb, 6);
941 for (n = 0; n < 10; n++) {
942 double delta = old[n] - i_lsps[n];
943 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
944 a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
947 dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r,
952 * Parse 16 independently-coded LSPs.
954 static void dequant_lsp16i(GetBitContext *gb, double *lsps)
956 static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 };
957 static const double mul_lsf[5] = {
958 3.3439586280e-3, 6.9908173703e-4,
959 3.3216608306e-3, 1.0334960326e-3,
962 static const double base_lsf[5] = {
963 M_PI * -1.27576e-1, M_PI * -2.4292e-2,
964 M_PI * -1.28094e-1, M_PI * -3.2128e-2,
969 v[0] = get_bits(gb, 8);
970 v[1] = get_bits(gb, 6);
971 v[2] = get_bits(gb, 7);
972 v[3] = get_bits(gb, 6);
973 v[4] = get_bits(gb, 7);
975 dequant_lsps( lsps, 5, v, vec_sizes, 2,
976 wmavoice_dq_lsp16i1, mul_lsf, base_lsf);
977 dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2,
978 wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]);
979 dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1,
980 wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]);
984 * Parse 16 independently-coded LSPs, and then derive the tables to
985 * generate LSPs for the other frames from them (residual coding).
987 static void dequant_lsp16r(GetBitContext *gb,
988 double *i_lsps, const double *old,
989 double *a1, double *a2, int q_mode)
991 static const uint16_t vec_sizes[3] = { 128, 128, 128 };
992 static const double mul_lsf[3] = {
993 1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3
995 static const double base_lsf[3] = {
996 M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2
998 const float (*ipol_tab)[2][16] = q_mode ?
999 wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a;
1000 uint16_t interpol, v[3];
1003 dequant_lsp16i(gb, i_lsps);
1005 interpol = get_bits(gb, 5);
1006 v[0] = get_bits(gb, 7);
1007 v[1] = get_bits(gb, 7);
1008 v[2] = get_bits(gb, 7);
1010 for (n = 0; n < 16; n++) {
1011 double delta = old[n] - i_lsps[n];
1012 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
1013 a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
1016 dequant_lsps( a2, 10, v, vec_sizes, 1,
1017 wmavoice_dq_lsp16r1, mul_lsf, base_lsf);
1018 dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1,
1019 wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]);
1020 dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1,
1021 wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]);
1026 * @name Pitch-adaptive window coding functions
1027 * The next few functions are for pitch-adaptive window coding.
1031 * Parse the offset of the first pitch-adaptive window pulses, and
1032 * the distribution of pulses between the two blocks in this frame.
1033 * @param s WMA Voice decoding context private data
1034 * @param gb bit I/O context
1035 * @param pitch pitch for each block in this frame
1037 static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb,
1040 static const int16_t start_offset[94] = {
1041 -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11,
1042 13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26,
1043 27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43,
1044 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67,
1045 69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91,
1046 93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115,
1047 117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139,
1048 141, 143, 145, 147, 149, 151, 153, 155, 157, 159
1052 /* position of pulse */
1053 s->aw_idx_is_ext = 0;
1054 if ((bits = get_bits(gb, 6)) >= 54) {
1055 s->aw_idx_is_ext = 1;
1056 bits += (bits - 54) * 3 + get_bits(gb, 2);
1059 /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count
1060 * the distribution of the pulses in each block contained in this frame. */
1061 s->aw_pulse_range = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16;
1062 for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ;
1063 s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0];
1064 s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2;
1065 offset += s->aw_n_pulses[0] * pitch[0];
1066 s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1];
1067 s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2;
1069 /* if continuing from a position before the block, reset position to
1070 * start of block (when corrected for the range over which it can be
1071 * spread in aw_pulse_set1()). */
1072 if (start_offset[bits] < MAX_FRAMESIZE / 2) {
1073 while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0)
1074 s->aw_first_pulse_off[1] -= pitch[1];
1075 if (start_offset[bits] < 0)
1076 while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0)
1077 s->aw_first_pulse_off[0] -= pitch[0];
1082 * Apply second set of pitch-adaptive window pulses.
1083 * @param s WMA Voice decoding context private data
1084 * @param gb bit I/O context
1085 * @param block_idx block index in frame [0, 1]
1086 * @param fcb structure containing fixed codebook vector info
1087 * @return -1 on error, 0 otherwise
1089 static int aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb,
1090 int block_idx, AMRFixed *fcb)
1092 uint16_t use_mask_mem[9]; // only 5 are used, rest is padding
1093 uint16_t *use_mask = use_mask_mem + 2;
1094 /* in this function, idx is the index in the 80-bit (+ padding) use_mask
1095 * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits
1096 * of idx are the position of the bit within a particular item in the
1097 * array (0 being the most significant bit, and 15 being the least
1098 * significant bit), and the remainder (>> 4) is the index in the
1099 * use_mask[]-array. This is faster and uses less memory than using a
1100 * 80-byte/80-int array. */
1101 int pulse_off = s->aw_first_pulse_off[block_idx],
1102 pulse_start, n, idx, range, aidx, start_off = 0;
1104 /* set offset of first pulse to within this block */
1105 if (s->aw_n_pulses[block_idx] > 0)
1106 while (pulse_off + s->aw_pulse_range < 1)
1107 pulse_off += fcb->pitch_lag;
1109 /* find range per pulse */
1110 if (s->aw_n_pulses[0] > 0) {
1111 if (block_idx == 0) {
1113 } else /* block_idx = 1 */ {
1115 if (s->aw_n_pulses[block_idx] > 0)
1116 pulse_off = s->aw_next_pulse_off_cache;
1120 pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0;
1122 /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,
1123 * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus
1124 * we exclude that range from being pulsed again in this function. */
1125 memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0]));
1126 memset( use_mask, -1, 5 * sizeof(use_mask[0]));
1127 memset(&use_mask[5], 0, 2 * sizeof(use_mask[0]));
1128 if (s->aw_n_pulses[block_idx] > 0)
1129 for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) {
1130 int excl_range = s->aw_pulse_range; // always 16 or 24
1131 uint16_t *use_mask_ptr = &use_mask[idx >> 4];
1132 int first_sh = 16 - (idx & 15);
1133 *use_mask_ptr++ &= 0xFFFFu << first_sh;
1134 excl_range -= first_sh;
1135 if (excl_range >= 16) {
1136 *use_mask_ptr++ = 0;
1137 *use_mask_ptr &= 0xFFFF >> (excl_range - 16);
1139 *use_mask_ptr &= 0xFFFF >> excl_range;
1142 /* find the 'aidx'th offset that is not excluded */
1143 aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4);
1144 for (n = 0; n <= aidx; pulse_start++) {
1145 for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ;
1146 if (idx >= MAX_FRAMESIZE / 2) { // find from zero
1147 if (use_mask[0]) idx = 0x0F;
1148 else if (use_mask[1]) idx = 0x1F;
1149 else if (use_mask[2]) idx = 0x2F;
1150 else if (use_mask[3]) idx = 0x3F;
1151 else if (use_mask[4]) idx = 0x4F;
1153 idx -= av_log2_16bit(use_mask[idx >> 4]);
1155 if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) {
1156 use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15));
1162 fcb->x[fcb->n] = start_off;
1163 fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0;
1166 /* set offset for next block, relative to start of that block */
1167 n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag;
1168 s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0;
1173 * Apply first set of pitch-adaptive window pulses.
1174 * @param s WMA Voice decoding context private data
1175 * @param gb bit I/O context
1176 * @param block_idx block index in frame [0, 1]
1177 * @param fcb storage location for fixed codebook pulse info
1179 static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb,
1180 int block_idx, AMRFixed *fcb)
1182 int val = get_bits(gb, 12 - 2 * (s->aw_idx_is_ext && !block_idx));
1185 if (s->aw_n_pulses[block_idx] > 0) {
1186 int n, v_mask, i_mask, sh, n_pulses;
1188 if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each
1193 } else { // 4 pulses, 1:sign + 2:index each
1200 for (n = n_pulses - 1; n >= 0; n--, val >>= sh) {
1201 fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0;
1202 fcb->x[fcb->n] = (val & i_mask) * n_pulses + n +
1203 s->aw_first_pulse_off[block_idx];
1204 while (fcb->x[fcb->n] < 0)
1205 fcb->x[fcb->n] += fcb->pitch_lag;
1206 if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2)
1210 int num2 = (val & 0x1FF) >> 1, delta, idx;
1212 if (num2 < 1 * 79) { delta = 1; idx = num2 + 1; }
1213 else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; }
1214 else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; }
1215 else { delta = 7; idx = num2 + 1 - 3 * 75; }
1216 v = (val & 0x200) ? -1.0 : 1.0;
1218 fcb->no_repeat_mask |= 3 << fcb->n;
1219 fcb->x[fcb->n] = idx - delta;
1221 fcb->x[fcb->n + 1] = idx;
1222 fcb->y[fcb->n + 1] = (val & 1) ? -v : v;
1230 * Generate a random number from frame_cntr and block_idx, which will live
1231 * in the range [0, 1000 - block_size] (so it can be used as an index in a
1232 * table of size 1000 of which you want to read block_size entries).
1234 * @param frame_cntr current frame number
1235 * @param block_num current block index
1236 * @param block_size amount of entries we want to read from a table
1237 * that has 1000 entries
1238 * @return a (non-)random number in the [0, 1000 - block_size] range.
1240 static int pRNG(int frame_cntr, int block_num, int block_size)
1242 /* array to simplify the calculation of z:
1243 * y = (x % 9) * 5 + 6;
1244 * z = (49995 * x) / y;
1245 * Since y only has 9 values, we can remove the division by using a
1246 * LUT and using FASTDIV-style divisions. For each of the 9 values
1247 * of y, we can rewrite z as:
1248 * z = x * (49995 / y) + x * ((49995 % y) / y)
1249 * In this table, each col represents one possible value of y, the
1250 * first number is 49995 / y, and the second is the FASTDIV variant
1251 * of 49995 % y / y. */
1252 static const unsigned int div_tbl[9][2] = {
1253 { 8332, 3 * 715827883U }, // y = 6
1254 { 4545, 0 * 390451573U }, // y = 11
1255 { 3124, 11 * 268435456U }, // y = 16
1256 { 2380, 15 * 204522253U }, // y = 21
1257 { 1922, 23 * 165191050U }, // y = 26
1258 { 1612, 23 * 138547333U }, // y = 31
1259 { 1388, 27 * 119304648U }, // y = 36
1260 { 1219, 16 * 104755300U }, // y = 41
1261 { 1086, 39 * 93368855U } // y = 46
1263 unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr;
1264 if (x >= 0xFFFF) x -= 0xFFFF; // max value of x is 8*1877+0xFFFE=0x13AA6,
1265 // so this is effectively a modulo (%)
1266 y = x - 9 * MULH(477218589, x); // x % 9
1267 z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1]));
1268 // z = x * 49995 / (y * 5 + 6)
1269 return z % (1000 - block_size);
1273 * Parse hardcoded signal for a single block.
1274 * @note see #synth_block().
1276 static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb,
1277 int block_idx, int size,
1278 const struct frame_type_desc *frame_desc,
1284 av_assert0(size <= MAX_FRAMESIZE);
1286 /* Set the offset from which we start reading wmavoice_std_codebook */
1287 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1288 r_idx = pRNG(s->frame_cntr, block_idx, size);
1289 gain = s->silence_gain;
1290 } else /* FCB_TYPE_HARDCODED */ {
1291 r_idx = get_bits(gb, 8);
1292 gain = wmavoice_gain_universal[get_bits(gb, 6)];
1295 /* Clear gain prediction parameters */
1296 memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err));
1298 /* Apply gain to hardcoded codebook and use that as excitation signal */
1299 for (n = 0; n < size; n++)
1300 excitation[n] = wmavoice_std_codebook[r_idx + n] * gain;
1304 * Parse FCB/ACB signal for a single block.
1305 * @note see #synth_block().
1307 static void synth_block_fcb_acb(WMAVoiceContext *s, GetBitContext *gb,
1308 int block_idx, int size,
1309 int block_pitch_sh2,
1310 const struct frame_type_desc *frame_desc,
1313 static const float gain_coeff[6] = {
1314 0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458
1316 float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain;
1317 int n, idx, gain_weight;
1320 av_assert0(size <= MAX_FRAMESIZE / 2);
1321 memset(pulses, 0, sizeof(*pulses) * size);
1323 fcb.pitch_lag = block_pitch_sh2 >> 2;
1324 fcb.pitch_fac = 1.0;
1325 fcb.no_repeat_mask = 0;
1328 /* For the other frame types, this is where we apply the innovation
1329 * (fixed) codebook pulses of the speech signal. */
1330 if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1331 aw_pulse_set1(s, gb, block_idx, &fcb);
1332 if (aw_pulse_set2(s, gb, block_idx, &fcb)) {
1333 /* Conceal the block with silence and return.
1334 * Skip the correct amount of bits to read the next
1335 * block from the correct offset. */
1336 int r_idx = pRNG(s->frame_cntr, block_idx, size);
1338 for (n = 0; n < size; n++)
1340 wmavoice_std_codebook[r_idx + n] * s->silence_gain;
1341 skip_bits(gb, 7 + 1);
1344 } else /* FCB_TYPE_EXC_PULSES */ {
1345 int offset_nbits = 5 - frame_desc->log_n_blocks;
1347 fcb.no_repeat_mask = -1;
1348 /* similar to ff_decode_10_pulses_35bits(), but with single pulses
1349 * (instead of double) for a subset of pulses */
1350 for (n = 0; n < 5; n++) {
1354 sign = get_bits1(gb) ? 1.0 : -1.0;
1355 pos1 = get_bits(gb, offset_nbits);
1356 fcb.x[fcb.n] = n + 5 * pos1;
1357 fcb.y[fcb.n++] = sign;
1358 if (n < frame_desc->dbl_pulses) {
1359 pos2 = get_bits(gb, offset_nbits);
1360 fcb.x[fcb.n] = n + 5 * pos2;
1361 fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign;
1365 ff_set_fixed_vector(pulses, &fcb, 1.0, size);
1367 /* Calculate gain for adaptive & fixed codebook signal.
1368 * see ff_amr_set_fixed_gain(). */
1369 idx = get_bits(gb, 7);
1370 fcb_gain = expf(avpriv_scalarproduct_float_c(s->gain_pred_err,
1372 5.2409161640 + wmavoice_gain_codebook_fcb[idx]);
1373 acb_gain = wmavoice_gain_codebook_acb[idx];
1374 pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx],
1375 -2.9957322736 /* log(0.05) */,
1376 1.6094379124 /* log(5.0) */);
1378 gain_weight = 8 >> frame_desc->log_n_blocks;
1379 memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err,
1380 sizeof(*s->gain_pred_err) * (6 - gain_weight));
1381 for (n = 0; n < gain_weight; n++)
1382 s->gain_pred_err[n] = pred_err;
1384 /* Calculation of adaptive codebook */
1385 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1387 for (n = 0; n < size; n += len) {
1389 int abs_idx = block_idx * size + n;
1390 int pitch_sh16 = (s->last_pitch_val << 16) +
1391 s->pitch_diff_sh16 * abs_idx;
1392 int pitch = (pitch_sh16 + 0x6FFF) >> 16;
1393 int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000;
1394 idx = idx_sh16 >> 16;
1395 if (s->pitch_diff_sh16) {
1396 if (s->pitch_diff_sh16 > 0) {
1397 next_idx_sh16 = (idx_sh16) &~ 0xFFFF;
1399 next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF;
1400 len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8,
1405 ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch],
1406 wmavoice_ipol1_coeffs, 17,
1409 } else /* ACB_TYPE_HAMMING */ {
1410 int block_pitch = block_pitch_sh2 >> 2;
1411 idx = block_pitch_sh2 & 3;
1413 ff_acelp_interpolatef(excitation, &excitation[-block_pitch],
1414 wmavoice_ipol2_coeffs, 4,
1417 av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch,
1418 sizeof(float) * size);
1421 /* Interpolate ACB/FCB and use as excitation signal */
1422 ff_weighted_vector_sumf(excitation, excitation, pulses,
1423 acb_gain, fcb_gain, size);
1427 * Parse data in a single block.
1429 * @param s WMA Voice decoding context private data
1430 * @param gb bit I/O context
1431 * @param block_idx index of the to-be-read block
1432 * @param size amount of samples to be read in this block
1433 * @param block_pitch_sh2 pitch for this block << 2
1434 * @param lsps LSPs for (the end of) this frame
1435 * @param prev_lsps LSPs for the last frame
1436 * @param frame_desc frame type descriptor
1437 * @param excitation target memory for the ACB+FCB interpolated signal
1438 * @param synth target memory for the speech synthesis filter output
1439 * @return 0 on success, <0 on error.
1441 static void synth_block(WMAVoiceContext *s, GetBitContext *gb,
1442 int block_idx, int size,
1443 int block_pitch_sh2,
1444 const double *lsps, const double *prev_lsps,
1445 const struct frame_type_desc *frame_desc,
1446 float *excitation, float *synth)
1448 double i_lsps[MAX_LSPS];
1449 float lpcs[MAX_LSPS];
1453 if (frame_desc->acb_type == ACB_TYPE_NONE)
1454 synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation);
1456 synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2,
1457 frame_desc, excitation);
1459 /* convert interpolated LSPs to LPCs */
1460 fac = (block_idx + 0.5) / frame_desc->n_blocks;
1461 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1462 i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n]));
1463 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1465 /* Speech synthesis */
1466 ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps);
1470 * Synthesize output samples for a single frame.
1472 * @param ctx WMA Voice decoder context
1473 * @param gb bit I/O context (s->gb or one for cross-packet superframes)
1474 * @param frame_idx Frame number within superframe [0-2]
1475 * @param samples pointer to output sample buffer, has space for at least 160
1477 * @param lsps LSP array
1478 * @param prev_lsps array of previous frame's LSPs
1479 * @param excitation target buffer for excitation signal
1480 * @param synth target buffer for synthesized speech data
1481 * @return 0 on success, <0 on error.
1483 static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx,
1485 const double *lsps, const double *prev_lsps,
1486 float *excitation, float *synth)
1488 WMAVoiceContext *s = ctx->priv_data;
1489 int n, n_blocks_x2, log_n_blocks_x2, av_uninit(cur_pitch_val);
1490 int pitch[MAX_BLOCKS], av_uninit(last_block_pitch);
1492 /* Parse frame type ("frame header"), see frame_descs */
1493 int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)], block_nsamples;
1496 av_log(ctx, AV_LOG_ERROR,
1497 "Invalid frame type VLC code, skipping\n");
1498 return AVERROR_INVALIDDATA;
1501 block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks;
1503 /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */
1504 if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) {
1505 /* Pitch is provided per frame, which is interpreted as the pitch of
1506 * the last sample of the last block of this frame. We can interpolate
1507 * the pitch of other blocks (and even pitch-per-sample) by gradually
1508 * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */
1509 n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1;
1510 log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1;
1511 cur_pitch_val = s->min_pitch_val + get_bits(gb, s->pitch_nbits);
1512 cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1);
1513 if (s->last_acb_type == ACB_TYPE_NONE ||
1514 20 * abs(cur_pitch_val - s->last_pitch_val) >
1515 (cur_pitch_val + s->last_pitch_val))
1516 s->last_pitch_val = cur_pitch_val;
1518 /* pitch per block */
1519 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1520 int fac = n * 2 + 1;
1522 pitch[n] = (MUL16(fac, cur_pitch_val) +
1523 MUL16((n_blocks_x2 - fac), s->last_pitch_val) +
1524 frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2;
1527 /* "pitch-diff-per-sample" for calculation of pitch per sample */
1528 s->pitch_diff_sh16 =
1529 ((cur_pitch_val - s->last_pitch_val) << 16) / MAX_FRAMESIZE;
1532 /* Global gain (if silence) and pitch-adaptive window coordinates */
1533 switch (frame_descs[bd_idx].fcb_type) {
1534 case FCB_TYPE_SILENCE:
1535 s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)];
1537 case FCB_TYPE_AW_PULSES:
1538 aw_parse_coords(s, gb, pitch);
1542 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1545 /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */
1546 switch (frame_descs[bd_idx].acb_type) {
1547 case ACB_TYPE_HAMMING: {
1548 /* Pitch is given per block. Per-block pitches are encoded as an
1549 * absolute value for the first block, and then delta values
1550 * relative to this value) for all subsequent blocks. The scale of
1551 * this pitch value is semi-logarithmic compared to its use in the
1552 * decoder, so we convert it to normal scale also. */
1554 t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2,
1555 t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1,
1556 t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1;
1559 block_pitch = get_bits(gb, s->block_pitch_nbits);
1561 block_pitch = last_block_pitch - s->block_delta_pitch_hrange +
1562 get_bits(gb, s->block_delta_pitch_nbits);
1563 /* Convert last_ so that any next delta is within _range */
1564 last_block_pitch = av_clip(block_pitch,
1565 s->block_delta_pitch_hrange,
1566 s->block_pitch_range -
1567 s->block_delta_pitch_hrange);
1569 /* Convert semi-log-style scale back to normal scale */
1570 if (block_pitch < t1) {
1571 bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch;
1574 if (block_pitch < t2) {
1576 (s->block_conv_table[1] << 2) + (block_pitch << 1);
1579 if (block_pitch < t3) {
1581 (s->block_conv_table[2] + block_pitch) << 2;
1583 bl_pitch_sh2 = s->block_conv_table[3] << 2;
1586 pitch[n] = bl_pitch_sh2 >> 2;
1590 case ACB_TYPE_ASYMMETRIC: {
1591 bl_pitch_sh2 = pitch[n] << 2;
1595 default: // ACB_TYPE_NONE has no pitch
1600 synth_block(s, gb, n, block_nsamples, bl_pitch_sh2,
1601 lsps, prev_lsps, &frame_descs[bd_idx],
1602 &excitation[n * block_nsamples],
1603 &synth[n * block_nsamples]);
1606 /* Averaging projection filter, if applicable. Else, just copy samples
1607 * from synthesis buffer */
1609 double i_lsps[MAX_LSPS];
1610 float lpcs[MAX_LSPS];
1612 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1613 i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n]));
1614 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1615 postfilter(s, synth, samples, 80, lpcs,
1616 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx],
1617 frame_descs[bd_idx].fcb_type, pitch[0]);
1619 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1620 i_lsps[n] = cos(lsps[n]);
1621 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1622 postfilter(s, &synth[80], &samples[80], 80, lpcs,
1623 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80],
1624 frame_descs[bd_idx].fcb_type, pitch[0]);
1626 memcpy(samples, synth, 160 * sizeof(synth[0]));
1628 /* Cache values for next frame */
1630 if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%)
1631 s->last_acb_type = frame_descs[bd_idx].acb_type;
1632 switch (frame_descs[bd_idx].acb_type) {
1634 s->last_pitch_val = 0;
1636 case ACB_TYPE_ASYMMETRIC:
1637 s->last_pitch_val = cur_pitch_val;
1639 case ACB_TYPE_HAMMING:
1640 s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1];
1648 * Ensure minimum value for first item, maximum value for last value,
1649 * proper spacing between each value and proper ordering.
1651 * @param lsps array of LSPs
1652 * @param num size of LSP array
1654 * @note basically a double version of #ff_acelp_reorder_lsf(), might be
1655 * useful to put in a generic location later on. Parts are also
1656 * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(),
1657 * which is in float.
1659 static void stabilize_lsps(double *lsps, int num)
1663 /* set minimum value for first, maximum value for last and minimum
1664 * spacing between LSF values.
1665 * Very similar to ff_set_min_dist_lsf(), but in double. */
1666 lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI);
1667 for (n = 1; n < num; n++)
1668 lsps[n] = FFMAX(lsps[n], lsps[n - 1] + 0.0125 * M_PI);
1669 lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI);
1671 /* reorder (looks like one-time / non-recursed bubblesort).
1672 * Very similar to ff_sort_nearly_sorted_floats(), but in double. */
1673 for (n = 1; n < num; n++) {
1674 if (lsps[n] < lsps[n - 1]) {
1675 for (m = 1; m < num; m++) {
1676 double tmp = lsps[m];
1677 for (l = m - 1; l >= 0; l--) {
1678 if (lsps[l] <= tmp) break;
1679 lsps[l + 1] = lsps[l];
1689 * Synthesize output samples for a single superframe. If we have any data
1690 * cached in s->sframe_cache, that will be used instead of whatever is loaded
1693 * WMA Voice superframes contain 3 frames, each containing 160 audio samples,
1694 * to give a total of 480 samples per frame. See #synth_frame() for frame
1695 * parsing. In addition to 3 frames, superframes can also contain the LSPs
1696 * (if these are globally specified for all frames (residually); they can
1697 * also be specified individually per-frame. See the s->has_residual_lsps
1698 * option), and can specify the number of samples encoded in this superframe
1699 * (if less than 480), usually used to prevent blanks at track boundaries.
1701 * @param ctx WMA Voice decoder context
1702 * @return 0 on success, <0 on error or 1 if there was not enough data to
1703 * fully parse the superframe
1705 static int synth_superframe(AVCodecContext *ctx, AVFrame *frame,
1708 WMAVoiceContext *s = ctx->priv_data;
1709 GetBitContext *gb = &s->gb, s_gb;
1710 int n, res, n_samples = MAX_SFRAMESIZE;
1711 double lsps[MAX_FRAMES][MAX_LSPS];
1712 const double *mean_lsf = s->lsps == 16 ?
1713 wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode];
1714 float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12];
1715 float synth[MAX_LSPS + MAX_SFRAMESIZE];
1718 memcpy(synth, s->synth_history,
1719 s->lsps * sizeof(*synth));
1720 memcpy(excitation, s->excitation_history,
1721 s->history_nsamples * sizeof(*excitation));
1723 if (s->sframe_cache_size > 0) {
1725 init_get_bits(gb, s->sframe_cache, s->sframe_cache_size);
1726 s->sframe_cache_size = 0;
1729 /* First bit is speech/music bit, it differentiates between WMAVoice
1730 * speech samples (the actual codec) and WMAVoice music samples, which
1731 * are really WMAPro-in-WMAVoice-superframes. I've never seen those in
1733 if (!get_bits1(gb)) {
1734 avpriv_request_sample(ctx, "WMAPro-in-WMAVoice");
1735 return AVERROR_PATCHWELCOME;
1738 /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */
1739 if (get_bits1(gb)) {
1740 if ((n_samples = get_bits(gb, 12)) > MAX_SFRAMESIZE) {
1741 av_log(ctx, AV_LOG_ERROR,
1742 "Superframe encodes > %d samples (%d), not allowed\n",
1743 MAX_SFRAMESIZE, n_samples);
1744 return AVERROR_INVALIDDATA;
1748 /* Parse LSPs, if global for the superframe (can also be per-frame). */
1749 if (s->has_residual_lsps) {
1750 double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2];
1752 for (n = 0; n < s->lsps; n++)
1753 prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n];
1755 if (s->lsps == 10) {
1756 dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1757 } else /* s->lsps == 16 */
1758 dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1760 for (n = 0; n < s->lsps; n++) {
1761 lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]);
1762 lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]);
1763 lsps[2][n] += mean_lsf[n];
1765 for (n = 0; n < 3; n++)
1766 stabilize_lsps(lsps[n], s->lsps);
1769 /* get output buffer */
1770 frame->nb_samples = MAX_SFRAMESIZE;
1771 if ((res = ff_get_buffer(ctx, frame, 0)) < 0)
1773 frame->nb_samples = n_samples;
1774 samples = (float *)frame->data[0];
1776 /* Parse frames, optionally preceded by per-frame (independent) LSPs. */
1777 for (n = 0; n < 3; n++) {
1778 if (!s->has_residual_lsps) {
1781 if (s->lsps == 10) {
1782 dequant_lsp10i(gb, lsps[n]);
1783 } else /* s->lsps == 16 */
1784 dequant_lsp16i(gb, lsps[n]);
1786 for (m = 0; m < s->lsps; m++)
1787 lsps[n][m] += mean_lsf[m];
1788 stabilize_lsps(lsps[n], s->lsps);
1791 if ((res = synth_frame(ctx, gb, n,
1792 &samples[n * MAX_FRAMESIZE],
1793 lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1],
1794 &excitation[s->history_nsamples + n * MAX_FRAMESIZE],
1795 &synth[s->lsps + n * MAX_FRAMESIZE]))) {
1801 /* Statistics? FIXME - we don't check for length, a slight overrun
1802 * will be caught by internal buffer padding, and anything else
1803 * will be skipped, not read. */
1804 if (get_bits1(gb)) {
1805 res = get_bits(gb, 4);
1806 skip_bits(gb, 10 * (res + 1));
1809 if (get_bits_left(gb) < 0) {
1810 wmavoice_flush(ctx);
1811 return AVERROR_INVALIDDATA;
1816 /* Update history */
1817 memcpy(s->prev_lsps, lsps[2],
1818 s->lsps * sizeof(*s->prev_lsps));
1819 memcpy(s->synth_history, &synth[MAX_SFRAMESIZE],
1820 s->lsps * sizeof(*synth));
1821 memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE],
1822 s->history_nsamples * sizeof(*excitation));
1824 memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE],
1825 s->history_nsamples * sizeof(*s->zero_exc_pf));
1831 * Parse the packet header at the start of each packet (input data to this
1834 * @param s WMA Voice decoding context private data
1835 * @return <0 on error, nb_superframes on success.
1837 static int parse_packet_header(WMAVoiceContext *s)
1839 GetBitContext *gb = &s->gb;
1840 unsigned int res, n_superframes = 0;
1842 skip_bits(gb, 4); // packet sequence number
1843 s->has_residual_lsps = get_bits1(gb);
1845 res = get_bits(gb, 6); // number of superframes per packet
1846 // (minus first one if there is spillover)
1847 n_superframes += res;
1848 } while (res == 0x3F);
1849 s->spillover_nbits = get_bits(gb, s->spillover_bitsize);
1851 return get_bits_left(gb) >= 0 ? n_superframes : AVERROR_INVALIDDATA;
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 #avpriv_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 if (nbits > pb->size_in_bits - put_bits_count(pb))
1880 rmn_bits &= 7; rmn_bytes >>= 3;
1881 if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0)
1882 put_bits(pb, rmn_bits, get_bits(gb, rmn_bits));
1883 avpriv_copy_bits(pb, data + size - rmn_bytes,
1884 FFMIN(nbits - rmn_bits, rmn_bytes << 3));
1888 * Packet decoding: a packet is anything that the (ASF) demuxer contains,
1889 * and we expect that the demuxer / application provides it to us as such
1890 * (else you'll probably get garbage as output). Every packet has a size of
1891 * ctx->block_align bytes, starts with a packet header (see
1892 * #parse_packet_header()), and then a series of superframes. Superframe
1893 * boundaries may exceed packets, i.e. superframes can split data over
1894 * multiple (two) packets.
1896 * For more information about frames, see #synth_superframe().
1898 static int wmavoice_decode_packet(AVCodecContext *ctx, void *data,
1899 int *got_frame_ptr, AVPacket *avpkt)
1901 WMAVoiceContext *s = ctx->priv_data;
1902 GetBitContext *gb = &s->gb;
1905 /* Packets are sometimes a multiple of ctx->block_align, with a packet
1906 * header at each ctx->block_align bytes. However, FFmpeg's ASF demuxer
1907 * feeds us ASF packets, which may concatenate multiple "codec" packets
1908 * in a single "muxer" packet, so we artificially emulate that by
1909 * capping the packet size at ctx->block_align. */
1910 for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align);
1911 init_get_bits(&s->gb, avpkt->data, size << 3);
1913 /* size == ctx->block_align is used to indicate whether we are dealing with
1914 * a new packet or a packet of which we already read the packet header
1916 if (!(size % ctx->block_align)) { // new packet header
1918 s->spillover_nbits = 0;
1919 s->nb_superframes = 0;
1921 if ((res = parse_packet_header(s)) < 0)
1923 s->nb_superframes = res;
1926 /* If the packet header specifies a s->spillover_nbits, then we want
1927 * to push out all data of the previous packet (+ spillover) before
1928 * continuing to parse new superframes in the current packet. */
1929 if (s->sframe_cache_size > 0) {
1930 int cnt = get_bits_count(gb);
1931 if (cnt + s->spillover_nbits > avpkt->size * 8) {
1932 s->spillover_nbits = avpkt->size * 8 - cnt;
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, got_frame_ptr)) == 0 &&
1939 cnt += s->spillover_nbits;
1940 s->skip_bits_next = cnt & 7;
1944 skip_bits_long (gb, s->spillover_nbits - cnt +
1945 get_bits_count(gb)); // resync
1946 } else if (s->spillover_nbits) {
1947 skip_bits_long(gb, s->spillover_nbits); // resync
1949 } else if (s->skip_bits_next)
1950 skip_bits(gb, s->skip_bits_next);
1952 /* Try parsing superframes in current packet */
1953 s->sframe_cache_size = 0;
1954 s->skip_bits_next = 0;
1955 pos = get_bits_left(gb);
1956 if (s->nb_superframes-- == 0) {
1959 } else if (s->nb_superframes > 0) {
1960 if ((res = synth_superframe(ctx, data, got_frame_ptr)) < 0) {
1962 } else if (*got_frame_ptr) {
1963 int cnt = get_bits_count(gb);
1964 s->skip_bits_next = cnt & 7;
1968 } else if ((s->sframe_cache_size = pos) > 0) {
1969 /* ... cache it for spillover in next packet */
1970 init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE);
1971 copy_bits(&s->pb, avpkt->data, size, gb, s->sframe_cache_size);
1972 // FIXME bad - just copy bytes as whole and add use the
1973 // skip_bits_next field
1979 static av_cold int wmavoice_decode_end(AVCodecContext *ctx)
1981 WMAVoiceContext *s = ctx->priv_data;
1984 ff_rdft_end(&s->rdft);
1985 ff_rdft_end(&s->irdft);
1986 ff_dct_end(&s->dct);
1987 ff_dct_end(&s->dst);
1993 AVCodec ff_wmavoice_decoder = {
1995 .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"),
1996 .type = AVMEDIA_TYPE_AUDIO,
1997 .id = AV_CODEC_ID_WMAVOICE,
1998 .priv_data_size = sizeof(WMAVoiceContext),
1999 .init = wmavoice_decode_init,
2000 .init_static_data = wmavoice_init_static_data,
2001 .close = wmavoice_decode_end,
2002 .decode = wmavoice_decode_packet,
2003 .capabilities = AV_CODEC_CAP_SUBFRAMES | AV_CODEC_CAP_DR1 | AV_CODEC_CAP_DELAY,
2004 .flush = wmavoice_flush,