FFmpeg coverage

Directory: ../../../ffmpeg/
File: src/libavcodec/wmavoice.c
Date: 2025-07-12 11:25:58
Exec Total Coverage
Lines: 682 777 87.8%
Functions: 28 29 96.6%
Branches: 311 398 78.1%
Line Branch Exec Source
1 /*
2 * Windows Media Audio Voice decoder.
3 * Copyright (c) 2009 Ronald S. Bultje
4 *
5 * This file is part of FFmpeg.
6 *
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
11 *
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
16 *
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20 */
21
22 /**
23 * @file
24 * @brief Windows Media Audio Voice compatible decoder
25 * @author Ronald S. Bultje <rsbultje@gmail.com>
26 */
27
28 #include <math.h>
29
30 #include "libavutil/channel_layout.h"
31 #include "libavutil/float_dsp.h"
32 #include "libavutil/mem.h"
33 #include "libavutil/mem_internal.h"
34 #include "libavutil/thread.h"
35 #include "libavutil/tx.h"
36 #include "avcodec.h"
37 #include "codec_internal.h"
38 #include "decode.h"
39 #include "get_bits.h"
40 #include "put_bits.h"
41 #include "wmavoice_data.h"
42 #include "celp_filters.h"
43 #include "acelp_vectors.h"
44 #include "acelp_filters.h"
45 #include "lsp.h"
46 #include "sinewin.h"
47
48 #define MAX_BLOCKS 8 ///< maximum number of blocks per frame
49 #define MAX_LSPS 16 ///< maximum filter order
50 #define MAX_LSPS_ALIGN16 16 ///< same as #MAX_LSPS; needs to be multiple
51 ///< of 16 for ASM input buffer alignment
52 #define MAX_FRAMES 3 ///< maximum number of frames per superframe
53 #define MAX_FRAMESIZE 160 ///< maximum number of samples per frame
54 #define MAX_SIGNAL_HISTORY 416 ///< maximum excitation signal history
55 #define MAX_SFRAMESIZE (MAX_FRAMESIZE * MAX_FRAMES)
56 ///< maximum number of samples per superframe
57 #define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that
58 ///< was split over two packets
59 #define VLC_NBITS 6 ///< number of bits to read per VLC iteration
60
61 /**
62 * Frame type VLC coding.
63 */
64 static VLCElem frame_type_vlc[132];
65
66 /**
67 * Adaptive codebook types.
68 */
69 enum {
70 ACB_TYPE_NONE = 0, ///< no adaptive codebook (only hardcoded fixed)
71 ACB_TYPE_ASYMMETRIC = 1, ///< adaptive codebook with per-frame pitch, which
72 ///< we interpolate to get a per-sample pitch.
73 ///< Signal is generated using an asymmetric sinc
74 ///< window function
75 ///< @note see #wmavoice_ipol1_coeffs
76 ACB_TYPE_HAMMING = 2 ///< Per-block pitch with signal generation using
77 ///< a Hamming sinc window function
78 ///< @note see #wmavoice_ipol2_coeffs
79 };
80
81 /**
82 * Fixed codebook types.
83 */
84 enum {
85 FCB_TYPE_SILENCE = 0, ///< comfort noise during silence
86 ///< generated from a hardcoded (fixed) codebook
87 ///< with per-frame (low) gain values
88 FCB_TYPE_HARDCODED = 1, ///< hardcoded (fixed) codebook with per-block
89 ///< gain values
90 FCB_TYPE_AW_PULSES = 2, ///< Pitch-adaptive window (AW) pulse signals,
91 ///< used in particular for low-bitrate streams
92 FCB_TYPE_EXC_PULSES = 3, ///< Innovation (fixed) codebook pulse sets in
93 ///< combinations of either single pulses or
94 ///< pulse pairs
95 };
96
97 /**
98 * Description of frame types.
99 */
100 static const struct frame_type_desc {
101 uint8_t n_blocks; ///< amount of blocks per frame (each block
102 ///< (contains 160/#n_blocks samples)
103 uint8_t log_n_blocks; ///< log2(#n_blocks)
104 uint8_t acb_type; ///< Adaptive codebook type (ACB_TYPE_*)
105 uint8_t fcb_type; ///< Fixed codebook type (FCB_TYPE_*)
106 uint8_t dbl_pulses; ///< how many pulse vectors have pulse pairs
107 ///< (rather than just one single pulse)
108 ///< only if #fcb_type == #FCB_TYPE_EXC_PULSES
109 } frame_descs[17] = {
110 { 1, 0, ACB_TYPE_NONE, FCB_TYPE_SILENCE, 0 },
111 { 2, 1, ACB_TYPE_NONE, FCB_TYPE_HARDCODED, 0 },
112 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_AW_PULSES, 0 },
113 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2 },
114 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5 },
115 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 0 },
116 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2 },
117 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5 },
118 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0 },
119 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2 },
120 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5 },
121 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0 },
122 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2 },
123 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5 },
124 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0 },
125 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2 },
126 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5 }
127 };
128
129 /**
130 * WMA Voice decoding context.
131 */
132 typedef struct WMAVoiceContext {
133 /**
134 * @name Global values specified in the stream header / extradata or used all over.
135 * @{
136 */
137 GetBitContext gb; ///< packet bitreader. During decoder init,
138 ///< it contains the extradata from the
139 ///< demuxer. During decoding, it contains
140 ///< packet data.
141 int8_t vbm_tree[25]; ///< converts VLC codes to frame type
142
143 int spillover_bitsize; ///< number of bits used to specify
144 ///< #spillover_nbits in the packet header
145 ///< = ceil(log2(ctx->block_align << 3))
146 int history_nsamples; ///< number of samples in history for signal
147 ///< prediction (through ACB)
148
149 /* postfilter specific values */
150 int do_apf; ///< whether to apply the averaged
151 ///< projection filter (APF)
152 int denoise_strength; ///< strength of denoising in Wiener filter
153 ///< [0-11]
154 int denoise_tilt_corr; ///< Whether to apply tilt correction to the
155 ///< Wiener filter coefficients (postfilter)
156 int dc_level; ///< Predicted amount of DC noise, based
157 ///< on which a DC removal filter is used
158
159 int lsps; ///< number of LSPs per frame [10 or 16]
160 int lsp_q_mode; ///< defines quantizer defaults [0, 1]
161 int lsp_def_mode; ///< defines different sets of LSP defaults
162 ///< [0, 1]
163
164 int min_pitch_val; ///< base value for pitch parsing code
165 int max_pitch_val; ///< max value + 1 for pitch parsing
166 int pitch_nbits; ///< number of bits used to specify the
167 ///< pitch value in the frame header
168 int block_pitch_nbits; ///< number of bits used to specify the
169 ///< first block's pitch value
170 int block_pitch_range; ///< range of the block pitch
171 int block_delta_pitch_nbits; ///< number of bits used to specify the
172 ///< delta pitch between this and the last
173 ///< block's pitch value, used in all but
174 ///< first block
175 int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is
176 ///< from -this to +this-1)
177 uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale
178 ///< conversion
179
180 /**
181 * @}
182 *
183 * @name Packet values specified in the packet header or related to a packet.
184 *
185 * A packet is considered to be a single unit of data provided to this
186 * decoder by the demuxer.
187 * @{
188 */
189 int spillover_nbits; ///< number of bits of the previous packet's
190 ///< last superframe preceding this
191 ///< packet's first full superframe (useful
192 ///< for re-synchronization also)
193 int has_residual_lsps; ///< if set, superframes contain one set of
194 ///< LSPs that cover all frames, encoded as
195 ///< independent and residual LSPs; if not
196 ///< set, each frame contains its own, fully
197 ///< independent, LSPs
198 int skip_bits_next; ///< number of bits to skip at the next call
199 ///< to #wmavoice_decode_packet() (since
200 ///< they're part of the previous superframe)
201
202 uint8_t sframe_cache[SFRAME_CACHE_MAXSIZE + AV_INPUT_BUFFER_PADDING_SIZE]; ///<
203 ///< cache for superframe data split over
204 ///< multiple packets
205 int sframe_cache_size; ///< set to >0 if we have data from an
206 ///< (incomplete) superframe from a previous
207 ///< packet that spilled over in the current
208 ///< packet; specifies the amount of bits in
209 ///< #sframe_cache
210 PutBitContext pb; ///< bitstream writer for #sframe_cache
211
212 /**
213 * @}
214 *
215 * @name Frame and superframe values
216 * Superframe and frame data - these can change from frame to frame,
217 * although some of them do in that case serve as a cache / history for
218 * the next frame or superframe.
219 * @{
220 */
221 double prev_lsps[MAX_LSPS]; ///< LSPs of the last frame of the previous
222 ///< superframe
223 int last_pitch_val; ///< pitch value of the previous frame
224 int last_acb_type; ///< frame type [0-2] of the previous frame
225 int pitch_diff_sh16; ///< ((cur_pitch_val - #last_pitch_val)
226 ///< << 16) / #MAX_FRAMESIZE
227 float silence_gain; ///< set for use in blocks if #ACB_TYPE_NONE
228
229 int aw_idx_is_ext; ///< whether the AW index was encoded in
230 ///< 8 bits (instead of 6)
231 int aw_pulse_range; ///< the range over which #aw_pulse_set1()
232 ///< can apply the pulse, relative to the
233 ///< value in aw_first_pulse_off. The exact
234 ///< position of the first AW-pulse is within
235 ///< [pulse_off, pulse_off + this], and
236 ///< depends on bitstream values; [16 or 24]
237 int aw_n_pulses[2]; ///< number of AW-pulses in each block; note
238 ///< that this number can be negative (in
239 ///< which case it basically means "zero")
240 int aw_first_pulse_off[2]; ///< index of first sample to which to
241 ///< apply AW-pulses, or -0xff if unset
242 int aw_next_pulse_off_cache; ///< the position (relative to start of the
243 ///< second block) at which pulses should
244 ///< start to be positioned, serves as a
245 ///< cache for pitch-adaptive window pulses
246 ///< between blocks
247
248 int frame_cntr; ///< current frame index [0 - 0xFFFE]; is
249 ///< only used for comfort noise in #pRNG()
250 int nb_superframes; ///< number of superframes in current packet
251 float gain_pred_err[6]; ///< cache for gain prediction
252 float excitation_history[MAX_SIGNAL_HISTORY]; ///< cache of the signal of
253 ///< previous superframes, used as a history
254 ///< for signal generation
255 float synth_history[MAX_LSPS]; ///< see #excitation_history
256 /**
257 * @}
258 *
259 * @name Postfilter values
260 *
261 * Variables used for postfilter implementation, mostly history for
262 * smoothing and so on, and context variables for FFT/iFFT.
263 * @{
264 */
265 AVTXContext *rdft, *irdft; ///< contexts for FFT-calculation in the
266 av_tx_fn rdft_fn, irdft_fn; ///< postfilter (for denoise filter)
267 AVTXContext *dct, *dst; ///< contexts for phase shift (in Hilbert
268 av_tx_fn dct_fn, dst_fn; ///< transform, part of postfilter)
269 float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi]
270 ///< range
271 float postfilter_agc; ///< gain control memory, used in
272 ///< #adaptive_gain_control()
273 float dcf_mem[2]; ///< DC filter history
274 /// zero filter output (i.e. excitation) by postfilter
275 float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE];
276 float denoise_filter_cache[MAX_FRAMESIZE];
277 int denoise_filter_cache_size; ///< samples in #denoise_filter_cache
278 /// aligned buffer for LPC tilting
279 DECLARE_ALIGNED(32, float, tilted_lpcs_pf)[0x82];
280 /// aligned buffer for denoise coefficients
281 DECLARE_ALIGNED(32, float, denoise_coeffs_pf)[0x82];
282 /// aligned buffer for postfilter speech synthesis
283 DECLARE_ALIGNED(32, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16];
284 /**
285 * @}
286 */
287 } WMAVoiceContext;
288
289 /**
290 * Set up the variable bit mode (VBM) tree from container extradata.
291 * @param gb bit I/O context.
292 * The bit context (s->gb) should be loaded with byte 23-46 of the
293 * container extradata (i.e. the ones containing the VBM tree).
294 * @param vbm_tree pointer to array to which the decoded VBM tree will be
295 * written.
296 * @return 0 on success, <0 on error.
297 */
298 8 static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25])
299 {
300 8 int cntr[8] = { 0 }, n, res;
301
302 8 memset(vbm_tree, 0xff, sizeof(vbm_tree[0]) * 25);
303
2/2
✓ Branch 0 taken 136 times.
✓ Branch 1 taken 8 times.
 144 for (n = 0; n < 17; n++) {
304 136 res = get_bits(gb, 3);
305
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 136 times.
 136 if (cntr[res] > 3) // should be >= 3 + (res == 7))
306 ✗ return -1;
307 136 vbm_tree[res * 3 + cntr[res]++] = n;
308 }
309 8 return 0;
310 }
311
312 5 static av_cold void wmavoice_init_static_data(void)
313 {
314 static const uint8_t bits[] = {
315 2, 2, 2, 4, 4, 4,
316 6, 6, 6, 8, 8, 8,
317 10, 10, 10, 12, 12, 12,
318 14, 14, 14, 14
319 };
320
321 5 VLC_INIT_STATIC_TABLE_FROM_LENGTHS(frame_type_vlc, VLC_NBITS,
322 FF_ARRAY_ELEMS(bits), bits,
323 1, NULL, 0, 0, 0, 0);
324 5 }
325
326 ✗ static av_cold void wmavoice_flush(AVCodecContext *ctx)
327 {
328 ✗ WMAVoiceContext *s = ctx->priv_data;
329 int n;
330
331 ✗ s->postfilter_agc = 0;
332 ✗ s->sframe_cache_size = 0;
333 ✗ s->skip_bits_next = 0;
334 ✗ for (n = 0; n < s->lsps; n++)
335 ✗ s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
336 ✗ memset(s->excitation_history, 0,
337 sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY);
338 ✗ memset(s->synth_history, 0,
339 sizeof(*s->synth_history) * MAX_LSPS);
340 ✗ memset(s->gain_pred_err, 0,
341 sizeof(s->gain_pred_err));
342
343 ✗ if (s->do_apf) {
344 ✗ memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0,
345 ✗ sizeof(*s->synth_filter_out_buf) * s->lsps);
346 ✗ memset(s->dcf_mem, 0,
347 sizeof(*s->dcf_mem) * 2);
348 ✗ memset(s->zero_exc_pf, 0,
349 ✗ sizeof(*s->zero_exc_pf) * s->history_nsamples);
350 ✗ memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache));
351 }
352 ✗ }
353
354 /**
355 * Set up decoder with parameters from demuxer (extradata etc.).
356 */
357 8 static av_cold int wmavoice_decode_init(AVCodecContext *ctx)
358 {
359 static AVOnce init_static_once = AV_ONCE_INIT;
360 int n, flags, pitch_range, lsp16_flag, ret;
361 8 WMAVoiceContext *s = ctx->priv_data;
362
363 8 ff_thread_once(&init_static_once, wmavoice_init_static_data);
364
365 /**
366 * Extradata layout:
367 * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c),
368 * - byte 19-22: flags field (annoyingly in LE; see below for known
369 * values),
370 * - byte 23-46: variable bitmode tree (really just 17 * 3 bits,
371 * rest is 0).
372 */
373
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
 8 if (ctx->extradata_size != 46) {
374 ✗ av_log(ctx, AV_LOG_ERROR,
375 "Invalid extradata size %d (should be 46)\n",
376 ctx->extradata_size);
377 ✗ return AVERROR_INVALIDDATA;
378 }
379
2/4
✓ Branch 0 taken 8 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 8 times.
 8 if (ctx->block_align <= 0 || ctx->block_align > (1<<22)) {
380 ✗ av_log(ctx, AV_LOG_ERROR, "Invalid block alignment %d.\n", ctx->block_align);
381 ✗ return AVERROR_INVALIDDATA;
382 }
383
384 8 flags = AV_RL32(ctx->extradata + 18);
385 8 s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align);
386 8 s->do_apf = flags & 0x1;
387
1/2
✓ Branch 0 taken 8 times.
✗ Branch 1 not taken.
 8 if (s->do_apf) {
388 8 float scale = 1.0f;
389
390 8 ret = av_tx_init(&s->rdft, &s->rdft_fn, AV_TX_FLOAT_RDFT, 0, 1 << 7, &scale, 0);
391
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
 8 if (ret < 0)
392 ✗ return ret;
393
394 8 ret = av_tx_init(&s->irdft, &s->irdft_fn, AV_TX_FLOAT_RDFT, 1, 1 << 7, &scale, 0);
395
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
 8 if (ret < 0)
396 ✗ return ret;
397
398 8 scale = 1.0 / (1 << 6);
399 8 ret = av_tx_init(&s->dct, &s->dct_fn, AV_TX_FLOAT_DCT_I, 0, 1 << 6, &scale, 0);
400
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
 8 if (ret < 0)
401 ✗ return ret;
402
403 8 scale = 1.0 / (1 << 6);
404 8 ret = av_tx_init(&s->dst, &s->dst_fn, AV_TX_FLOAT_DST_I, 0, 1 << 6, &scale, 0);
405
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
 8 if (ret < 0)
406 ✗ return ret;
407
408 8 ff_sine_window_init(s->cos, 256);
409 8 memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0]));
410
2/2
✓ Branch 0 taken 2040 times.
✓ Branch 1 taken 8 times.
 2048 for (n = 0; n < 255; n++) {
411 2040 s->sin[n] = -s->sin[510 - n];
412 2040 s->cos[510 - n] = s->cos[n];
413 }
414 }
415 8 s->denoise_strength = (flags >> 2) & 0xF;
416
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
 8 if (s->denoise_strength >= 12) {
417 ✗ av_log(ctx, AV_LOG_ERROR,
418 "Invalid denoise filter strength %d (max=11)\n",
419 s->denoise_strength);
420 ✗ return AVERROR_INVALIDDATA;
421 }
422 8 s->denoise_tilt_corr = !!(flags & 0x40);
423 8 s->dc_level = (flags >> 7) & 0xF;
424 8 s->lsp_q_mode = !!(flags & 0x2000);
425 8 s->lsp_def_mode = !!(flags & 0x4000);
426 8 lsp16_flag = flags & 0x1000;
427
2/2
✓ Branch 0 taken 4 times.
✓ Branch 1 taken 4 times.
 8 if (lsp16_flag) {
428 4 s->lsps = 16;
429 } else {
430 4 s->lsps = 10;
431 }
432
2/2
✓ Branch 0 taken 104 times.
✓ Branch 1 taken 8 times.
 112 for (n = 0; n < s->lsps; n++)
433 104 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
434
435 8 init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3);
436
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 8 times.
 8 if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) {
437 ✗ av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n");
438 ✗ return AVERROR_INVALIDDATA;
439 }
440
441
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
 8 if (ctx->sample_rate >= INT_MAX / (256 * 37))
442 ✗ return AVERROR_INVALIDDATA;
443
444 8 s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8;
445 8 s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8;
446 8 pitch_range = s->max_pitch_val - s->min_pitch_val;
447
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
 8 if (pitch_range <= 0) {
448 ✗ av_log(ctx, AV_LOG_ERROR, "Invalid pitch range; broken extradata?\n");
449 ✗ return AVERROR_INVALIDDATA;
450 }
451 8 s->pitch_nbits = av_ceil_log2(pitch_range);
452 8 s->last_pitch_val = 40;
453 8 s->last_acb_type = ACB_TYPE_NONE;
454 8 s->history_nsamples = s->max_pitch_val + 8;
455
456
2/4
✓ Branch 0 taken 8 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 8 times.
 8 if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) {
457 ✗ int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8,
458 ✗ max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8;
459
460 ✗ av_log(ctx, AV_LOG_ERROR,
461 "Unsupported samplerate %d (min=%d, max=%d)\n",
462 ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz
463
464 ✗ return AVERROR(ENOSYS);
465 }
466
467 8 s->block_conv_table[0] = s->min_pitch_val;
468 8 s->block_conv_table[1] = (pitch_range * 25) >> 6;
469 8 s->block_conv_table[2] = (pitch_range * 44) >> 6;
470 8 s->block_conv_table[3] = s->max_pitch_val - 1;
471 8 s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF;
472
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
 8 if (s->block_delta_pitch_hrange <= 0) {
473 ✗ av_log(ctx, AV_LOG_ERROR, "Invalid delta pitch hrange; broken extradata?\n");
474 ✗ return AVERROR_INVALIDDATA;
475 }
476 8 s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange);
477 8 s->block_pitch_range = s->block_conv_table[2] +
478 8 s->block_conv_table[3] + 1 +
479 8 2 * (s->block_conv_table[1] - 2 * s->min_pitch_val);
480 8 s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range);
481
482 8 av_channel_layout_uninit(&ctx->ch_layout);
483 8 ctx->ch_layout = (AVChannelLayout)AV_CHANNEL_LAYOUT_MONO;
484 8 ctx->sample_fmt = AV_SAMPLE_FMT_FLT;
485
486 8 return 0;
487 }
488
489 /**
490 * @name Postfilter functions
491 * Postfilter functions (gain control, wiener denoise filter, DC filter,
492 * kalman smoothening, plus surrounding code to wrap it)
493 * @{
494 */
495 /**
496 * Adaptive gain control (as used in postfilter).
497 *
498 * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except
499 * that the energy here is calculated using sum(abs(...)), whereas the
500 * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)).
501 *
502 * @param out output buffer for filtered samples
503 * @param in input buffer containing the samples as they are after the
504 * postfilter steps so far
505 * @param speech_synth input buffer containing speech synth before postfilter
506 * @param size input buffer size
507 * @param alpha exponential filter factor
508 * @param gain_mem pointer to filter memory (single float)
509 */
510 6612 static void adaptive_gain_control(float *out, const float *in,
511 const float *speech_synth,
512 int size, float alpha, float *gain_mem)
513 {
514 int i;
515 6612 float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor;
516 6612 float mem = *gain_mem;
517
518
2/2
✓ Branch 0 taken 528960 times.
✓ Branch 1 taken 6612 times.
 535572 for (i = 0; i < size; i++) {
519 528960 speech_energy += fabsf(speech_synth[i]);
520 528960 postfilter_energy += fabsf(in[i]);
521 }
522
1/2
✓ Branch 0 taken 6612 times.
✗ Branch 1 not taken.
 6612 gain_scale_factor = postfilter_energy == 0.0 ? 0.0 :
523 6612 (1.0 - alpha) * speech_energy / postfilter_energy;
524
525
2/2
✓ Branch 0 taken 528960 times.
✓ Branch 1 taken 6612 times.
 535572 for (i = 0; i < size; i++) {
526 528960 mem = alpha * mem + gain_scale_factor;
527 528960 out[i] = in[i] * mem;
528 }
529
530 6612 *gain_mem = mem;
531 6612 }
532
533 /**
534 * Kalman smoothing function.
535 *
536 * This function looks back pitch +/- 3 samples back into history to find
537 * the best fitting curve (that one giving the optimal gain of the two
538 * signals, i.e. the highest dot product between the two), and then
539 * uses that signal history to smoothen the output of the speech synthesis
540 * filter.
541 *
542 * @param s WMA Voice decoding context
543 * @param pitch pitch of the speech signal
544 * @param in input speech signal
545 * @param out output pointer for smoothened signal
546 * @param size input/output buffer size
547 *
548 * @returns -1 if no smoothening took place, e.g. because no optimal
549 * fit could be found, or 0 on success.
550 */
551 5070 static int kalman_smoothen(WMAVoiceContext *s, int pitch,
552 const float *in, float *out, int size)
553 {
554 int n;
555 5070 float optimal_gain = 0, dot;
556
2/2
✓ Branch 0 taken 44 times.
✓ Branch 1 taken 5026 times.
 5070 const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)],
557 5070 *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)],
558 5070 *best_hist_ptr = NULL;
559
560 /* find best fitting point in history */
561 do {
562 35388 dot = ff_scalarproduct_float_c(in, ptr, size);
563
2/2
✓ Branch 0 taken 12328 times.
✓ Branch 1 taken 23060 times.
 35388 if (dot > optimal_gain) {
564 12328 optimal_gain = dot;
565 12328 best_hist_ptr = ptr;
566 }
567
2/2
✓ Branch 0 taken 30318 times.
✓ Branch 1 taken 5070 times.
 35388 } while (--ptr >= end);
568
569
2/2
✓ Branch 0 taken 26 times.
✓ Branch 1 taken 5044 times.
 5070 if (optimal_gain <= 0)
570 26 return -1;
571 5044 dot = ff_scalarproduct_float_c(best_hist_ptr, best_hist_ptr, size);
572
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 5044 times.
 5044 if (dot <= 0) // would be 1.0
573 ✗ return -1;
574
575
2/2
✓ Branch 0 taken 4872 times.
✓ Branch 1 taken 172 times.
 5044 if (optimal_gain <= dot) {
576 4872 dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000
577 } else
578 172 dot = 0.625;
579
580 /* actual smoothing */
581
2/2
✓ Branch 0 taken 403520 times.
✓ Branch 1 taken 5044 times.
 408564 for (n = 0; n < size; n++)
582 403520 out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]);
583
584 5044 return 0;
585 }
586
587 /**
588 * Get the tilt factor of a formant filter from its transfer function
589 * @see #tilt_factor() in amrnbdec.c, which does essentially the same,
590 * but somehow (??) it does a speech synthesis filter in the
591 * middle, which is missing here
592 *
593 * @param lpcs LPC coefficients
594 * @param n_lpcs Size of LPC buffer
595 * @returns the tilt factor
596 */
597 7098 static float tilt_factor(const float *lpcs, int n_lpcs)
598 {
599 float rh0, rh1;
600
601 7098 rh0 = 1.0 + ff_scalarproduct_float_c(lpcs, lpcs, n_lpcs);
602 7098 rh1 = lpcs[0] + ff_scalarproduct_float_c(lpcs, &lpcs[1], n_lpcs - 1);
603
604 7098 return rh1 / rh0;
605 }
606
607 /**
608 * Derive denoise filter coefficients (in real domain) from the LPCs.
609 */
610 5614 static void calc_input_response(WMAVoiceContext *s, float *lpcs_src,
611 int fcb_type, float *coeffs_dst, int remainder)
612 {
613 5614 float last_coeff, min = 15.0, max = -15.0;
614 float irange, angle_mul, gain_mul, range, sq;
615 5614 LOCAL_ALIGNED_32(float, coeffs, [0x82]);
616 5614 LOCAL_ALIGNED_32(float, lpcs, [0x82]);
617 5614 LOCAL_ALIGNED_32(float, lpcs_dct, [0x82]);
618 int n, idx;
619
620 5614 memcpy(coeffs, coeffs_dst, 0x82*sizeof(float));
621
622 /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */
623 5614 s->rdft_fn(s->rdft, lpcs, lpcs_src, sizeof(float));
624 #define log_range(var, assign) do { \
625 float tmp = log10f(assign); var = tmp; \
626 max = FFMAX(max, tmp); min = FFMIN(min, tmp); \
627 } while (0)
628
2/4
✗ Branch 0 not taken.
✓ Branch 1 taken 5614 times.
✓ Branch 2 taken 5614 times.
✗ Branch 3 not taken.
 5614 log_range(last_coeff, lpcs[64] * lpcs[64]);
629
2/2
✓ Branch 0 taken 353682 times.
✓ Branch 1 taken 5614 times.
 359296 for (n = 1; n < 64; n++)
630
4/4
✓ Branch 0 taken 286139 times.
✓ Branch 1 taken 67543 times.
✓ Branch 2 taken 16276 times.
✓ Branch 3 taken 337406 times.
 353682 log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] +
631 lpcs[n * 2 + 1] * lpcs[n * 2 + 1]);
632
4/4
✓ Branch 0 taken 5604 times.
✓ Branch 1 taken 10 times.
✓ Branch 2 taken 1233 times.
✓ Branch 3 taken 4381 times.
 5614 log_range(lpcs[0], lpcs[0] * lpcs[0]);
633 #undef log_range
634 5614 range = max - min;
635 5614 lpcs[64] = last_coeff;
636
637 /* Now, use this spectrum to pick out these frequencies with higher
638 * (relative) power/energy (which we then take to be "not noise"),
639 * and set up a table (still in lpc[]) of (relative) gains per frequency.
640 * These frequencies will be maintained, while others ("noise") will be
641 * decreased in the filter output. */
642 5614 irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63]
643
2/2
✓ Branch 0 taken 544 times.
✓ Branch 1 taken 5070 times.
 5614 gain_mul = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) :
644 (5.0 / 14.7));
645 5614 angle_mul = gain_mul * (8.0 * M_LN10 / M_PI);
646
2/2
✓ Branch 0 taken 364910 times.
✓ Branch 1 taken 5614 times.
 370524 for (n = 0; n <= 64; n++) {
647 float pwr;
648
649 364910 idx = lrint((max - lpcs[n]) * irange - 1);
650 364910 idx = FFMAX(0, idx);
651 364910 pwr = wmavoice_denoise_power_table[s->denoise_strength][idx];
652 364910 lpcs[n] = angle_mul * pwr;
653
654 /* 70.57 =~ 1/log10(1.0331663) */
655 364910 idx = av_clipd((pwr * gain_mul - 0.0295) * 70.570526123, 0, INT_MAX / 2);
656
657
2/2
✓ Branch 0 taken 9151 times.
✓ Branch 1 taken 355759 times.
 364910 if (idx > 127) { // fall back if index falls outside table range
658 9151 coeffs[n] = wmavoice_energy_table[127] *
659 9151 powf(1.0331663, idx - 127);
660 } else
661 355759 coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)];
662 }
663
664 /* calculate the Hilbert transform of the gains, which we do (since this
665 * is a sine input) by doing a phase shift (in theory, H(sin())=cos()).
666 * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the
667 * "moment" of the LPCs in this filter. */
668 5614 s->dct_fn(s->dct, lpcs_dct, lpcs, sizeof(float));
669 5614 s->dst_fn(s->dst, lpcs, lpcs_dct, sizeof(float));
670
671 /* Split out the coefficient indexes into phase/magnitude pairs */
672 5614 idx = 255 + av_clip(lpcs[64], -255, 255);
673 5614 coeffs[0] = coeffs[0] * s->cos[idx];
674 5614 idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255);
675 5614 last_coeff = coeffs[64] * s->cos[idx];
676 5614 for (n = 63;; n--) {
677 353682 idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255);
678 179648 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
679 179648 coeffs[n * 2] = coeffs[n] * s->cos[idx];
680
681
2/2
✓ Branch 0 taken 5614 times.
✓ Branch 1 taken 174034 times.
 179648 if (!--n) break;
682
683 174034 idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255);
684 174034 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
685 174034 coeffs[n * 2] = coeffs[n] * s->cos[idx];
686 }
687 5614 coeffs[64] = last_coeff;
688
689 /* move into real domain */
690 5614 s->irdft_fn(s->irdft, coeffs_dst, coeffs, sizeof(AVComplexFloat));
691
692 /* tilt correction and normalize scale */
693 5614 memset(&coeffs_dst[remainder], 0, sizeof(coeffs_dst[0]) * (128 - remainder));
694
2/2
✓ Branch 0 taken 1484 times.
✓ Branch 1 taken 4130 times.
 5614 if (s->denoise_tilt_corr) {
695 1484 float tilt_mem = 0;
696
697 1484 coeffs_dst[remainder - 1] = 0;
698 1484 ff_tilt_compensation(&tilt_mem,
699 1484 -1.8 * tilt_factor(coeffs_dst, remainder - 1),
700 coeffs_dst, remainder);
701 }
702 5614 sq = (1.0 / 64.0) * sqrtf(1 / ff_scalarproduct_float_c(coeffs_dst, coeffs_dst,
703 remainder));
704
2/2
✓ Branch 0 taken 263858 times.
✓ Branch 1 taken 5614 times.
 269472 for (n = 0; n < remainder; n++)
705 263858 coeffs_dst[n] *= sq;
706 5614 }
707
708 /**
709 * This function applies a Wiener filter on the (noisy) speech signal as
710 * a means to denoise it.
711 *
712 * - take RDFT of LPCs to get the power spectrum of the noise + speech;
713 * - using this power spectrum, calculate (for each frequency) the Wiener
714 * filter gain, which depends on the frequency power and desired level
715 * of noise subtraction (when set too high, this leads to artifacts)
716 * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse
717 * of 4-8kHz);
718 * - by doing a phase shift, calculate the Hilbert transform of this array
719 * of per-frequency filter-gains to get the filtering coefficients;
720 * - smoothen/normalize/de-tilt these filter coefficients as desired;
721 * - take RDFT of noisy sound, apply the coefficients and take its IRDFT
722 * to get the denoised speech signal;
723 * - the leftover (i.e. output of the IRDFT on denoised speech data beyond
724 * the frame boundary) are saved and applied to subsequent frames by an
725 * overlap-add method (otherwise you get clicking-artifacts).
726 *
727 * @param s WMA Voice decoding context
728 * @param fcb_type Frame (codebook) type
729 * @param synth_pf input: the noisy speech signal, output: denoised speech
730 * data; should be 16-byte aligned (for ASM purposes)
731 * @param size size of the speech data
732 * @param lpcs LPCs used to synthesize this frame's speech data
733 */
734 6612 static void wiener_denoise(WMAVoiceContext *s, int fcb_type,
735 float *synth_pf, int size,
736 const float *lpcs)
737 {
738 int remainder, lim, n;
739
740
2/2
✓ Branch 0 taken 5614 times.
✓ Branch 1 taken 998 times.
 6612 if (fcb_type != FCB_TYPE_SILENCE) {
741 5614 LOCAL_ALIGNED_32(float, coeffs_f, [0x82]);
742 5614 LOCAL_ALIGNED_32(float, synth_f, [0x82]);
743 5614 float *tilted_lpcs = s->tilted_lpcs_pf,
744 5614 *coeffs = s->denoise_coeffs_pf, tilt_mem = 0;
745
746 5614 tilted_lpcs[0] = 1.0;
747 5614 memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps);
748 5614 memset(&tilted_lpcs[s->lsps + 1], 0,
749 5614 sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1));
750 5614 ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps),
751 5614 tilted_lpcs, s->lsps + 2);
752
753 /* The IRDFT output (127 samples for 7-bit filter) beyond the frame
754 * size is applied to the next frame. All input beyond this is zero,
755 * and thus all output beyond this will go towards zero, hence we can
756 * limit to min(size-1, 127-size) as a performance consideration. */
757
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 5614 times.
 5614 remainder = FFMIN(127 - size, size - 1);
758 5614 calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder);
759
760 /* apply coefficients (in frequency spectrum domain), i.e. complex
761 * number multiplication */
762 5614 memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size));
763 5614 s->rdft_fn(s->rdft, synth_f, synth_pf, sizeof(float));
764 5614 s->rdft_fn(s->rdft, coeffs_f, coeffs, sizeof(float));
765 5614 synth_f[0] *= coeffs_f[0];
766 5614 synth_f[1] *= coeffs_f[1];
767
2/2
✓ Branch 0 taken 359296 times.
✓ Branch 1 taken 5614 times.
 364910 for (n = 1; n <= 64; n++) {
768 359296 float v1 = synth_f[n * 2], v2 = synth_f[n * 2 + 1];
769 359296 synth_f[n * 2] = v1 * coeffs_f[n * 2] - v2 * coeffs_f[n * 2 + 1];
770 359296 synth_f[n * 2 + 1] = v2 * coeffs_f[n * 2] + v1 * coeffs_f[n * 2 + 1];
771 }
772 5614 s->irdft_fn(s->irdft, synth_pf, synth_f, sizeof(AVComplexFloat));
773 }
774
775 /* merge filter output with the history of previous runs */
776
2/2
✓ Branch 0 taken 5612 times.
✓ Branch 1 taken 1000 times.
 6612 if (s->denoise_filter_cache_size) {
777 5612 lim = FFMIN(s->denoise_filter_cache_size, size);
778
2/2
✓ Branch 0 taken 263764 times.
✓ Branch 1 taken 5612 times.
 269376 for (n = 0; n < lim; n++)
779 263764 synth_pf[n] += s->denoise_filter_cache[n];
780 5612 s->denoise_filter_cache_size -= lim;
781 5612 memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size],
782 5612 sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size);
783 }
784
785 /* move remainder of filter output into a cache for future runs */
786
2/2
✓ Branch 0 taken 5614 times.
✓ Branch 1 taken 998 times.
 6612 if (fcb_type != FCB_TYPE_SILENCE) {
787 5614 lim = FFMIN(remainder, s->denoise_filter_cache_size);
788
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 5614 times.
 5614 for (n = 0; n < lim; n++)
789 ✗ s->denoise_filter_cache[n] += synth_pf[size + n];
790
1/2
✓ Branch 0 taken 5614 times.
✗ Branch 1 not taken.
 5614 if (lim < remainder) {
791 5614 memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim],
792 5614 sizeof(s->denoise_filter_cache[0]) * (remainder - lim));
793 5614 s->denoise_filter_cache_size = remainder;
794 }
795 }
796 6612 }
797
798 /**
799 * Averaging projection filter, the postfilter used in WMAVoice.
800 *
801 * This uses the following steps:
802 * - A zero-synthesis filter (generate excitation from synth signal)
803 * - Kalman smoothing on excitation, based on pitch
804 * - Re-synthesized smoothened output
805 * - Iterative Wiener denoise filter
806 * - Adaptive gain filter
807 * - DC filter
808 *
809 * @param s WMAVoice decoding context
810 * @param synth Speech synthesis output (before postfilter)
811 * @param samples Output buffer for filtered samples
812 * @param size Buffer size of synth & samples
813 * @param lpcs Generated LPCs used for speech synthesis
814 * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned)
815 * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses)
816 * @param pitch Pitch of the input signal
817 */
818 6612 static void postfilter(WMAVoiceContext *s, const float *synth,
819 float *samples, int size,
820 const float *lpcs, float *zero_exc_pf,
821 int fcb_type, int pitch)
822 {
823 float synth_filter_in_buf[MAX_FRAMESIZE / 2],
824 6612 *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16],
825 6612 *synth_filter_in = zero_exc_pf;
826
827
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 6612 times.
 6612 av_assert0(size <= MAX_FRAMESIZE / 2);
828
829 /* generate excitation from input signal */
830 6612 ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps);
831
832
4/4
✓ Branch 0 taken 5070 times.
✓ Branch 1 taken 1542 times.
✓ Branch 2 taken 5044 times.
✓ Branch 3 taken 26 times.
 11682 if (fcb_type >= FCB_TYPE_AW_PULSES &&
833 5070 !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size))
834 5044 synth_filter_in = synth_filter_in_buf;
835
836 /* re-synthesize speech after smoothening, and keep history */
837 6612 ff_celp_lp_synthesis_filterf(synth_pf, lpcs,
838 synth_filter_in, size, s->lsps);
839 6612 memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps],
840 6612 sizeof(synth_pf[0]) * s->lsps);
841
842 6612 wiener_denoise(s, fcb_type, synth_pf, size, lpcs);
843
844 6612 adaptive_gain_control(samples, synth_pf, synth, size, 0.99,
845 &s->postfilter_agc);
846
847
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 6612 times.
 6612 if (s->dc_level > 8) {
848 /* remove ultra-low frequency DC noise / highpass filter;
849 * coefficients are identical to those used in SIPR decoding,
850 * and very closely resemble those used in AMR-NB decoding. */
851 ✗ ff_acelp_apply_order_2_transfer_function(samples, samples,
852 ✗ (const float[2]) { -1.99997, 1.0 },
853 ✗ (const float[2]) { -1.9330735188, 0.93589198496 },
854 ✗ 0.93980580475, s->dcf_mem, size);
855 }
856 6612 }
857 /**
858 * @}
859 */
860
861 /**
862 * Dequantize LSPs
863 * @param lsps output pointer to the array that will hold the LSPs
864 * @param num number of LSPs to be dequantized
865 * @param values quantized values, contains n_stages values
866 * @param sizes range (i.e. max value) of each quantized value
867 * @param n_stages number of dequantization runs
868 * @param table dequantization table to be used
869 * @param mul_q LSF multiplier
870 * @param base_q base (lowest) LSF values
871 */
872 4404 static void dequant_lsps(double *lsps, int num,
873 const uint16_t *values,
874 const uint16_t *sizes,
875 int n_stages, const uint8_t *table,
876 const double *mul_q,
877 const double *base_q)
878 {
879 int n, m;
880
881 4404 memset(lsps, 0, num * sizeof(*lsps));
882
2/2
✓ Branch 0 taken 8264 times.
✓ Branch 1 taken 4404 times.
 12668 for (n = 0; n < n_stages; n++) {
883 8264 const uint8_t *t_off = &table[values[n] * num];
884 8264 double base = base_q[n], mul = mul_q[n];
885
886
2/2
✓ Branch 0 taken 87100 times.
✓ Branch 1 taken 8264 times.
 95364 for (m = 0; m < num; m++)
887 87100 lsps[m] += base + mul * t_off[m];
888
889 8264 table += sizes[n] * num;
890 }
891 4404 }
892
893 /**
894 * @name LSP dequantization routines
895 * LSP dequantization routines, for 10/16LSPs and independent/residual coding.
896 * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;
897 * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits.
898 * @{
899 */
900 /**
901 * Parse 10 independently-coded LSPs.
902 */
903 552 static void dequant_lsp10i(GetBitContext *gb, double *lsps)
904 {
905 static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 };
906 static const double mul_lsf[4] = {
907 5.2187144800e-3, 1.4626986422e-3,
908 9.6179549166e-4, 1.1325736225e-3
909 };
910 static const double base_lsf[4] = {
911 M_PI * -2.15522e-1, M_PI * -6.1646e-2,
912 M_PI * -3.3486e-2, M_PI * -5.7408e-2
913 };
914 uint16_t v[4];
915
916 552 v[0] = get_bits(gb, 8);
917 552 v[1] = get_bits(gb, 6);
918 552 v[2] = get_bits(gb, 5);
919 552 v[3] = get_bits(gb, 5);
920
921 552 dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i,
922 mul_lsf, base_lsf);
923 552 }
924
925 /**
926 * Parse 10 independently-coded LSPs, and then derive the tables to
927 * generate LSPs for the other frames from them (residual coding).
928 */
929 552 static void dequant_lsp10r(GetBitContext *gb,
930 double *i_lsps, const double *old,
931 double *a1, double *a2, int q_mode)
932 {
933 static const uint16_t vec_sizes[3] = { 128, 64, 64 };
934 static const double mul_lsf[3] = {
935 2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3
936 };
937 static const double base_lsf[3] = {
938 M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2
939 };
940 552 const float (*ipol_tab)[2][10] = q_mode ?
941
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 552 times.
 552 wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a;
942 uint16_t interpol, v[3];
943 int n;
944
945 552 dequant_lsp10i(gb, i_lsps);
946
947 552 interpol = get_bits(gb, 5);
948 552 v[0] = get_bits(gb, 7);
949 552 v[1] = get_bits(gb, 6);
950 552 v[2] = get_bits(gb, 6);
951
952
2/2
✓ Branch 0 taken 5520 times.
✓ Branch 1 taken 552 times.
 6072 for (n = 0; n < 10; n++) {
953 5520 double delta = old[n] - i_lsps[n];
954 5520 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
955 5520 a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
956 }
957
958 552 dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r,
959 mul_lsf, base_lsf);
960 552 }
961
962 /**
963 * Parse 16 independently-coded LSPs.
964 */
965 550 static void dequant_lsp16i(GetBitContext *gb, double *lsps)
966 {
967 static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 };
968 static const double mul_lsf[5] = {
969 3.3439586280e-3, 6.9908173703e-4,
970 3.3216608306e-3, 1.0334960326e-3,
971 3.1899104283e-3
972 };
973 static const double base_lsf[5] = {
974 M_PI * -1.27576e-1, M_PI * -2.4292e-2,
975 M_PI * -1.28094e-1, M_PI * -3.2128e-2,
976 M_PI * -1.29816e-1
977 };
978 uint16_t v[5];
979
980 550 v[0] = get_bits(gb, 8);
981 550 v[1] = get_bits(gb, 6);
982 550 v[2] = get_bits(gb, 7);
983 550 v[3] = get_bits(gb, 6);
984 550 v[4] = get_bits(gb, 7);
985
986 550 dequant_lsps( lsps, 5, v, vec_sizes, 2,
987 wmavoice_dq_lsp16i1, mul_lsf, base_lsf);
988 550 dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2,
989 wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]);
990 550 dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1,
991 wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]);
992 550 }
993
994 /**
995 * Parse 16 independently-coded LSPs, and then derive the tables to
996 * generate LSPs for the other frames from them (residual coding).
997 */
998 550 static void dequant_lsp16r(GetBitContext *gb,
999 double *i_lsps, const double *old,
1000 double *a1, double *a2, int q_mode)
1001 {
1002 static const uint16_t vec_sizes[3] = { 128, 128, 128 };
1003 static const double mul_lsf[3] = {
1004 1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3
1005 };
1006 static const double base_lsf[3] = {
1007 M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2
1008 };
1009 550 const float (*ipol_tab)[2][16] = q_mode ?
1010
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 550 times.
 550 wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a;
1011 uint16_t interpol, v[3];
1012 int n;
1013
1014 550 dequant_lsp16i(gb, i_lsps);
1015
1016 550 interpol = get_bits(gb, 5);
1017 550 v[0] = get_bits(gb, 7);
1018 550 v[1] = get_bits(gb, 7);
1019 550 v[2] = get_bits(gb, 7);
1020
1021
2/2
✓ Branch 0 taken 8800 times.
✓ Branch 1 taken 550 times.
 9350 for (n = 0; n < 16; n++) {
1022 8800 double delta = old[n] - i_lsps[n];
1023 8800 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
1024 8800 a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
1025 }
1026
1027 550 dequant_lsps( a2, 10, v, vec_sizes, 1,
1028 wmavoice_dq_lsp16r1, mul_lsf, base_lsf);
1029 550 dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1,
1030 wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]);
1031 550 dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1,
1032 wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]);
1033 550 }
1034
1035 /**
1036 * @}
1037 * @name Pitch-adaptive window coding functions
1038 * The next few functions are for pitch-adaptive window coding.
1039 * @{
1040 */
1041 /**
1042 * Parse the offset of the first pitch-adaptive window pulses, and
1043 * the distribution of pulses between the two blocks in this frame.
1044 * @param s WMA Voice decoding context private data
1045 * @param gb bit I/O context
1046 * @param pitch pitch for each block in this frame
1047 */
1048 341 static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb,
1049 const int *pitch)
1050 {
1051 static const int16_t start_offset[94] = {
1052 -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11,
1053 13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26,
1054 27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43,
1055 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67,
1056 69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91,
1057 93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115,
1058 117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139,
1059 141, 143, 145, 147, 149, 151, 153, 155, 157, 159
1060 };
1061 int bits, offset;
1062
1063 /* position of pulse */
1064 341 s->aw_idx_is_ext = 0;
1065
2/2
✓ Branch 1 taken 10 times.
✓ Branch 2 taken 331 times.
 341 if ((bits = get_bits(gb, 6)) >= 54) {
1066 10 s->aw_idx_is_ext = 1;
1067 10 bits += (bits - 54) * 3 + get_bits(gb, 2);
1068 }
1069
1070 /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count
1071 * the distribution of the pulses in each block contained in this frame. */
1072
2/2
✓ Branch 0 taken 338 times.
✓ Branch 1 taken 3 times.
 341 s->aw_pulse_range = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16;
1073
2/2
✓ Branch 0 taken 50 times.
✓ Branch 1 taken 341 times.
 391 for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ;
1074 341 s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0];
1075 341 s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2;
1076 341 offset += s->aw_n_pulses[0] * pitch[0];
1077 341 s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1];
1078 341 s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2;
1079
1080 /* if continuing from a position before the block, reset position to
1081 * start of block (when corrected for the range over which it can be
1082 * spread in aw_pulse_set1()). */
1083
2/2
✓ Branch 0 taken 331 times.
✓ Branch 1 taken 10 times.
 341 if (start_offset[bits] < MAX_FRAMESIZE / 2) {
1084
2/2
✓ Branch 0 taken 56 times.
✓ Branch 1 taken 331 times.
 387 while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0)
1085 56 s->aw_first_pulse_off[1] -= pitch[1];
1086
2/2
✓ Branch 0 taken 50 times.
✓ Branch 1 taken 281 times.
 331 if (start_offset[bits] < 0)
1087
2/2
✓ Branch 0 taken 50 times.
✓ Branch 1 taken 50 times.
 100 while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0)
1088 50 s->aw_first_pulse_off[0] -= pitch[0];
1089 }
1090 341 }
1091
1092 /**
1093 * Apply second set of pitch-adaptive window pulses.
1094 * @param s WMA Voice decoding context private data
1095 * @param gb bit I/O context
1096 * @param block_idx block index in frame [0, 1]
1097 * @param fcb structure containing fixed codebook vector info
1098 * @return -1 on error, 0 otherwise
1099 */
1100 682 static int aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb,
1101 int block_idx, AMRFixed *fcb)
1102 {
1103 uint16_t use_mask_mem[9]; // only 5 are used, rest is padding
1104 682 uint16_t *use_mask = use_mask_mem + 2;
1105 /* in this function, idx is the index in the 80-bit (+ padding) use_mask
1106 * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits
1107 * of idx are the position of the bit within a particular item in the
1108 * array (0 being the most significant bit, and 15 being the least
1109 * significant bit), and the remainder (>> 4) is the index in the
1110 * use_mask[]-array. This is faster and uses less memory than using a
1111 * 80-byte/80-int array. */
1112 682 int pulse_off = s->aw_first_pulse_off[block_idx],
1113 682 pulse_start, n, idx, range, aidx, start_off = 0;
1114
1115 /* set offset of first pulse to within this block */
1116
2/2
✓ Branch 0 taken 657 times.
✓ Branch 1 taken 25 times.
 682 if (s->aw_n_pulses[block_idx] > 0)
1117
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 657 times.
 657 while (pulse_off + s->aw_pulse_range < 1)
1118 ✗ pulse_off += fcb->pitch_lag;
1119
1120 /* find range per pulse */
1121
2/2
✓ Branch 0 taken 646 times.
✓ Branch 1 taken 36 times.
 682 if (s->aw_n_pulses[0] > 0) {
1122
2/2
✓ Branch 0 taken 323 times.
✓ Branch 1 taken 323 times.
 646 if (block_idx == 0) {
1123 323 range = 32;
1124 } else /* block_idx = 1 */ {
1125 323 range = 8;
1126
2/2
✓ Branch 0 taken 316 times.
✓ Branch 1 taken 7 times.
 323 if (s->aw_n_pulses[block_idx] > 0)
1127 316 pulse_off = s->aw_next_pulse_off_cache;
1128 }
1129 } else
1130 36 range = 16;
1131
2/2
✓ Branch 0 taken 657 times.
✓ Branch 1 taken 25 times.
 682 pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0;
1132
1133 /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,
1134 * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus
1135 * we exclude that range from being pulsed again in this function. */
1136 682 memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0]));
1137 682 memset( use_mask, -1, 5 * sizeof(use_mask[0]));
1138 682 memset(&use_mask[5], 0, 2 * sizeof(use_mask[0]));
1139
2/2
✓ Branch 0 taken 657 times.
✓ Branch 1 taken 25 times.
 682 if (s->aw_n_pulses[block_idx] > 0)
1140
2/2
✓ Branch 0 taken 911 times.
✓ Branch 1 taken 657 times.
 1568 for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) {
1141 911 int excl_range = s->aw_pulse_range; // always 16 or 24
1142 911 uint16_t *use_mask_ptr = &use_mask[idx >> 4];
1143 911 int first_sh = 16 - (idx & 15);
1144 911 *use_mask_ptr++ &= 0xFFFFu << first_sh;
1145 911 excl_range -= first_sh;
1146
2/2
✓ Branch 0 taken 468 times.
✓ Branch 1 taken 443 times.
 911 if (excl_range >= 16) {
1147 468 *use_mask_ptr++ = 0;
1148 468 *use_mask_ptr &= 0xFFFF >> (excl_range - 16);
1149 } else
1150 443 *use_mask_ptr &= 0xFFFF >> excl_range;
1151 }
1152
1153 /* find the 'aidx'th offset that is not excluded */
1154
2/2
✓ Branch 0 taken 646 times.
✓ Branch 1 taken 36 times.
 682 aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4);
1155
2/2
✓ Branch 0 taken 16143 times.
✓ Branch 1 taken 682 times.
 16825 for (n = 0; n <= aidx; pulse_start++) {
1156
2/2
✓ Branch 0 taken 2315 times.
✓ Branch 1 taken 16143 times.
 18458 for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ;
1157
2/2
✓ Branch 0 taken 538 times.
✓ Branch 1 taken 15605 times.
 16143 if (idx >= MAX_FRAMESIZE / 2) { // find from zero
1158
2/2
✓ Branch 0 taken 415 times.
✓ Branch 1 taken 123 times.
 538 if (use_mask[0]) idx = 0x0F;
1159
2/2
✓ Branch 0 taken 105 times.
✓ Branch 1 taken 18 times.
 123 else if (use_mask[1]) idx = 0x1F;
1160
1/2
✓ Branch 0 taken 18 times.
✗ Branch 1 not taken.
 18 else if (use_mask[2]) idx = 0x2F;
1161 ✗ else if (use_mask[3]) idx = 0x3F;
1162 ✗ else if (use_mask[4]) idx = 0x4F;
1163 ✗ else return -1;
1164 538 idx -= av_log2_16bit(use_mask[idx >> 4]);
1165 }
1166
2/2
✓ Branch 0 taken 7465 times.
✓ Branch 1 taken 8678 times.
 16143 if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) {
1167 7465 use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15));
1168 7465 n++;
1169 7465 start_off = idx;
1170 }
1171 }
1172
1173 682 fcb->x[fcb->n] = start_off;
1174
2/2
✓ Branch 1 taken 365 times.
✓ Branch 2 taken 317 times.
 682 fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0;
1175 682 fcb->n++;
1176
1177 /* set offset for next block, relative to start of that block */
1178 682 n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag;
1179
2/2
✓ Branch 0 taken 676 times.
✓ Branch 1 taken 6 times.
 682 s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0;
1180 682 return 0;
1181 }
1182
1183 /**
1184 * Apply first set of pitch-adaptive window pulses.
1185 * @param s WMA Voice decoding context private data
1186 * @param gb bit I/O context
1187 * @param block_idx block index in frame [0, 1]
1188 * @param fcb storage location for fixed codebook pulse info
1189 */
1190 682 static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb,
1191 int block_idx, AMRFixed *fcb)
1192 {
1193
4/4
✓ Branch 0 taken 20 times.
✓ Branch 1 taken 662 times.
✓ Branch 2 taken 10 times.
✓ Branch 3 taken 10 times.
 682 int val = get_bits(gb, 12 - 2 * (s->aw_idx_is_ext && !block_idx));
1194 float v;
1195
1196
2/2
✓ Branch 0 taken 657 times.
✓ Branch 1 taken 25 times.
 682 if (s->aw_n_pulses[block_idx] > 0) {
1197 int n, v_mask, i_mask, sh, n_pulses;
1198
1199
2/2
✓ Branch 0 taken 652 times.
✓ Branch 1 taken 5 times.
 657 if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each
1200 652 n_pulses = 3;
1201 652 v_mask = 8;
1202 652 i_mask = 7;
1203 652 sh = 4;
1204 } else { // 4 pulses, 1:sign + 2:index each
1205 5 n_pulses = 4;
1206 5 v_mask = 4;
1207 5 i_mask = 3;
1208 5 sh = 3;
1209 }
1210
1211
2/2
✓ Branch 0 taken 1976 times.
✓ Branch 1 taken 657 times.
 2633 for (n = n_pulses - 1; n >= 0; n--, val >>= sh) {
1212
2/2
✓ Branch 0 taken 950 times.
✓ Branch 1 taken 1026 times.
 1976 fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0;
1213 1976 fcb->x[fcb->n] = (val & i_mask) * n_pulses + n +
1214 1976 s->aw_first_pulse_off[block_idx];
1215
2/2
✓ Branch 0 taken 241 times.
✓ Branch 1 taken 1976 times.
 2217 while (fcb->x[fcb->n] < 0)
1216 241 fcb->x[fcb->n] += fcb->pitch_lag;
1217
2/2
✓ Branch 0 taken 1959 times.
✓ Branch 1 taken 17 times.
 1976 if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2)
1218 1959 fcb->n++;
1219 }
1220 } else {
1221 25 int num2 = (val & 0x1FF) >> 1, delta, idx;
1222
1223
2/2
✓ Branch 0 taken 4 times.
✓ Branch 1 taken 21 times.
 25 if (num2 < 1 * 79) { delta = 1; idx = num2 + 1; }
1224
2/2
✓ Branch 0 taken 6 times.
✓ Branch 1 taken 15 times.
 21 else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; }
1225
2/2
✓ Branch 0 taken 10 times.
✓ Branch 1 taken 5 times.
 15 else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; }
1226 5 else { delta = 7; idx = num2 + 1 - 3 * 75; }
1227
2/2
✓ Branch 0 taken 11 times.
✓ Branch 1 taken 14 times.
 25 v = (val & 0x200) ? -1.0 : 1.0;
1228
1229 25 fcb->no_repeat_mask |= 3 << fcb->n;
1230 25 fcb->x[fcb->n] = idx - delta;
1231 25 fcb->y[fcb->n] = v;
1232 25 fcb->x[fcb->n + 1] = idx;
1233
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 18 times.
 25 fcb->y[fcb->n + 1] = (val & 1) ? -v : v;
1234 25 fcb->n += 2;
1235 }
1236 682 }
1237
1238 /**
1239 * @}
1240 *
1241 * Generate a random number from frame_cntr and block_idx, which will live
1242 * in the range [0, 1000 - block_size] (so it can be used as an index in a
1243 * table of size 1000 of which you want to read block_size entries).
1244 *
1245 * @param frame_cntr current frame number
1246 * @param block_num current block index
1247 * @param block_size amount of entries we want to read from a table
1248 * that has 1000 entries
1249 * @return a (non-)random number in the [0, 1000 - block_size] range.
1250 */
1251 499 static int pRNG(int frame_cntr, int block_num, int block_size)
1252 {
1253 /* array to simplify the calculation of z:
1254 * y = (x % 9) * 5 + 6;
1255 * z = (49995 * x) / y;
1256 * Since y only has 9 values, we can remove the division by using a
1257 * LUT and using FASTDIV-style divisions. For each of the 9 values
1258 * of y, we can rewrite z as:
1259 * z = x * (49995 / y) + x * ((49995 % y) / y)
1260 * In this table, each col represents one possible value of y, the
1261 * first number is 49995 / y, and the second is the FASTDIV variant
1262 * of 49995 % y / y. */
1263 static const unsigned int div_tbl[9][2] = {
1264 { 8332, 3 * 715827883U }, // y = 6
1265 { 4545, 0 * 390451573U }, // y = 11
1266 { 3124, 11 * 268435456U }, // y = 16
1267 { 2380, 15 * 204522253U }, // y = 21
1268 { 1922, 23 * 165191050U }, // y = 26
1269 { 1612, 23 * 138547333U }, // y = 31
1270 { 1388, 27 * 119304648U }, // y = 36
1271 { 1219, 16 * 104755300U }, // y = 41
1272 { 1086, 39 * 93368855U } // y = 46
1273 };
1274 499 unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr;
1275
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 499 times.
 499 if (x >= 0xFFFF) x -= 0xFFFF; // max value of x is 8*1877+0xFFFE=0x13AA6,
1276 // so this is effectively a modulo (%)
1277 499 y = x - 9 * MULH(477218589, x); // x % 9
1278 499 z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1]));
1279 // z = x * 49995 / (y * 5 + 6)
1280 499 return z % (1000 - block_size);
1281 }
1282
1283 /**
1284 * Parse hardcoded signal for a single block.
1285 * @note see #synth_block().
1286 */
1287 1043 static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb,
1288 int block_idx, int size,
1289 const struct frame_type_desc *frame_desc,
1290 float *excitation)
1291 {
1292 float gain;
1293 int n, r_idx;
1294
1295
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 1043 times.
 1043 av_assert0(size <= MAX_FRAMESIZE);
1296
1297 /* Set the offset from which we start reading wmavoice_std_codebook */
1298
2/2
✓ Branch 0 taken 499 times.
✓ Branch 1 taken 544 times.
 1043 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1299 499 r_idx = pRNG(s->frame_cntr, block_idx, size);
1300 499 gain = s->silence_gain;
1301 } else /* FCB_TYPE_HARDCODED */ {
1302 544 r_idx = get_bits(gb, 8);
1303 544 gain = wmavoice_gain_universal[get_bits(gb, 6)];
1304 }
1305
1306 /* Clear gain prediction parameters */
1307 1043 memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err));
1308
1309 /* Apply gain to hardcoded codebook and use that as excitation signal */
1310
2/2
✓ Branch 0 taken 123360 times.
✓ Branch 1 taken 1043 times.
 124403 for (n = 0; n < size; n++)
1311 123360 excitation[n] = wmavoice_std_codebook[r_idx + n] * gain;
1312 1043 }
1313
1314 /**
1315 * Parse FCB/ACB signal for a single block.
1316 * @note see #synth_block().
1317 */
1318 9740 static void synth_block_fcb_acb(WMAVoiceContext *s, GetBitContext *gb,
1319 int block_idx, int size,
1320 int block_pitch_sh2,
1321 const struct frame_type_desc *frame_desc,
1322 float *excitation)
1323 {
1324 static const float gain_coeff[6] = {
1325 0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458
1326 };
1327 float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain;
1328 int n, idx, gain_weight;
1329 AMRFixed fcb;
1330
1331
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 9740 times.
 9740 av_assert0(size <= MAX_FRAMESIZE / 2);
1332 9740 memset(pulses, 0, sizeof(*pulses) * size);
1333
1334 9740 fcb.pitch_lag = block_pitch_sh2 >> 2;
1335 9740 fcb.pitch_fac = 1.0;
1336 9740 fcb.no_repeat_mask = 0;
1337 9740 fcb.n = 0;
1338
1339 /* For the other frame types, this is where we apply the innovation
1340 * (fixed) codebook pulses of the speech signal. */
1341
2/2
✓ Branch 0 taken 682 times.
✓ Branch 1 taken 9058 times.
 9740 if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1342 682 aw_pulse_set1(s, gb, block_idx, &fcb);
1343
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 682 times.
 682 if (aw_pulse_set2(s, gb, block_idx, &fcb)) {
1344 /* Conceal the block with silence and return.
1345 * Skip the correct amount of bits to read the next
1346 * block from the correct offset. */
1347 ✗ int r_idx = pRNG(s->frame_cntr, block_idx, size);
1348
1349 ✗ for (n = 0; n < size; n++)
1350 ✗ excitation[n] =
1351 ✗ wmavoice_std_codebook[r_idx + n] * s->silence_gain;
1352 ✗ skip_bits(gb, 7 + 1);
1353 ✗ return;
1354 }
1355 } else /* FCB_TYPE_EXC_PULSES */ {
1356 9058 int offset_nbits = 5 - frame_desc->log_n_blocks;
1357
1358 9058 fcb.no_repeat_mask = -1;
1359 /* similar to ff_decode_10_pulses_35bits(), but with single pulses
1360 * (instead of double) for a subset of pulses */
1361
2/2
✓ Branch 0 taken 45290 times.
✓ Branch 1 taken 9058 times.
 54348 for (n = 0; n < 5; n++) {
1362 float sign;
1363 int pos1, pos2;
1364
1365
2/2
✓ Branch 1 taken 22708 times.
✓ Branch 2 taken 22582 times.
 45290 sign = get_bits1(gb) ? 1.0 : -1.0;
1366 45290 pos1 = get_bits(gb, offset_nbits);
1367 45290 fcb.x[fcb.n] = n + 5 * pos1;
1368 45290 fcb.y[fcb.n++] = sign;
1369
2/2
✓ Branch 0 taken 36270 times.
✓ Branch 1 taken 9020 times.
 45290 if (n < frame_desc->dbl_pulses) {
1370 36270 pos2 = get_bits(gb, offset_nbits);
1371 36270 fcb.x[fcb.n] = n + 5 * pos2;
1372
2/2
✓ Branch 0 taken 21490 times.
✓ Branch 1 taken 14780 times.
 36270 fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign;
1373 }
1374 }
1375 }
1376 9740 ff_set_fixed_vector(pulses, &fcb, 1.0, size);
1377
1378 /* Calculate gain for adaptive & fixed codebook signal.
1379 * see ff_amr_set_fixed_gain(). */
1380 9740 idx = get_bits(gb, 7);
1381 9740 fcb_gain = expf(ff_scalarproduct_float_c(s->gain_pred_err,
1382 9740 gain_coeff, 6) -
1383 9740 5.2409161640 + wmavoice_gain_codebook_fcb[idx]);
1384 9740 acb_gain = wmavoice_gain_codebook_acb[idx];
1385 9740 pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx],
1386 -2.9957322736 /* log(0.05) */,
1387 1.6094379124 /* log(5.0) */);
1388
1389 9740 gain_weight = 8 >> frame_desc->log_n_blocks;
1390 9740 memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err,
1391 9740 sizeof(*s->gain_pred_err) * (6 - gain_weight));
1392
2/2
✓ Branch 0 taken 20280 times.
✓ Branch 1 taken 9740 times.
 30020 for (n = 0; n < gain_weight; n++)
1393 20280 s->gain_pred_err[n] = pred_err;
1394
1395 /* Calculation of adaptive codebook */
1396
2/2
✓ Branch 0 taken 1276 times.
✓ Branch 1 taken 8464 times.
 9740 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1397 int len;
1398
2/2
✓ Branch 0 taken 17876 times.
✓ Branch 1 taken 1276 times.
 19152 for (n = 0; n < size; n += len) {
1399 int next_idx_sh16;
1400 17876 int abs_idx = block_idx * size + n;
1401 17876 int pitch_sh16 = (s->last_pitch_val << 16) +
1402 17876 s->pitch_diff_sh16 * abs_idx;
1403 17876 int pitch = (pitch_sh16 + 0x6FFF) >> 16;
1404 17876 int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000;
1405 17876 idx = idx_sh16 >> 16;
1406
2/2
✓ Branch 0 taken 17442 times.
✓ Branch 1 taken 434 times.
 17876 if (s->pitch_diff_sh16) {
1407
2/2
✓ Branch 0 taken 10526 times.
✓ Branch 1 taken 6916 times.
 17442 if (s->pitch_diff_sh16 > 0) {
1408 10526 next_idx_sh16 = (idx_sh16) &~ 0xFFFF;
1409 } else
1410 6916 next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF;
1411 17442 len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8,
1412 1, size - n);
1413 } else
1414 434 len = size;
1415
1416 17876 ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch],
1417 wmavoice_ipol1_coeffs, 17,
1418 idx, 9, len);
1419 }
1420 } else /* ACB_TYPE_HAMMING */ {
1421 8464 int block_pitch = block_pitch_sh2 >> 2;
1422 8464 idx = block_pitch_sh2 & 3;
1423
2/2
✓ Branch 0 taken 3652 times.
✓ Branch 1 taken 4812 times.
 8464 if (idx) {
1424 3652 ff_acelp_interpolatef(excitation, &excitation[-block_pitch],
1425 wmavoice_ipol2_coeffs, 4,
1426 idx, 8, size);
1427 } else
1428 4812 av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch,
1429 sizeof(float) * size);
1430 }
1431
1432 /* Interpolate ACB/FCB and use as excitation signal */
1433 9740 ff_weighted_vector_sumf(excitation, excitation, pulses,
1434 acb_gain, fcb_gain, size);
1435 }
1436
1437 /**
1438 * Parse data in a single block.
1439 *
1440 * @param s WMA Voice decoding context private data
1441 * @param gb bit I/O context
1442 * @param block_idx index of the to-be-read block
1443 * @param size amount of samples to be read in this block
1444 * @param block_pitch_sh2 pitch for this block << 2
1445 * @param lsps LSPs for (the end of) this frame
1446 * @param prev_lsps LSPs for the last frame
1447 * @param frame_desc frame type descriptor
1448 * @param excitation target memory for the ACB+FCB interpolated signal
1449 * @param synth target memory for the speech synthesis filter output
1450 * @return 0 on success, <0 on error.
1451 */
1452 10783 static void synth_block(WMAVoiceContext *s, GetBitContext *gb,
1453 int block_idx, int size,
1454 int block_pitch_sh2,
1455 const double *lsps, const double *prev_lsps,
1456 const struct frame_type_desc *frame_desc,
1457 float *excitation, float *synth)
1458 {
1459 double i_lsps[MAX_LSPS];
1460 float lpcs[MAX_LSPS];
1461 float fac;
1462 int n;
1463
1464
2/2
✓ Branch 0 taken 1043 times.
✓ Branch 1 taken 9740 times.
 10783 if (frame_desc->acb_type == ACB_TYPE_NONE)
1465 1043 synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation);
1466 else
1467 9740 synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2,
1468 frame_desc, excitation);
1469
1470 /* convert interpolated LSPs to LPCs */
1471 10783 fac = (block_idx + 0.5) / frame_desc->n_blocks;
1472
2/2
✓ Branch 0 taken 140776 times.
✓ Branch 1 taken 10783 times.
 151559 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1473 140776 i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n]));
1474 10783 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1475
1476 /* Speech synthesis */
1477 10783 ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps);
1478 10783 }
1479
1480 /**
1481 * Synthesize output samples for a single frame.
1482 *
1483 * @param ctx WMA Voice decoder context
1484 * @param gb bit I/O context (s->gb or one for cross-packet superframes)
1485 * @param frame_idx Frame number within superframe [0-2]
1486 * @param samples pointer to output sample buffer, has space for at least 160
1487 * samples
1488 * @param lsps LSP array
1489 * @param prev_lsps array of previous frame's LSPs
1490 * @param excitation target buffer for excitation signal
1491 * @param synth target buffer for synthesized speech data
1492 * @return 0 on success, <0 on error.
1493 */
1494 3306 static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx,
1495 float *samples,
1496 const double *lsps, const double *prev_lsps,
1497 float *excitation, float *synth)
1498 {
1499 3306 WMAVoiceContext *s = ctx->priv_data;
1500 3306 int n, n_blocks_x2, log_n_blocks_x2, av_uninit(cur_pitch_val);
1501 3306 int pitch[MAX_BLOCKS], av_uninit(last_block_pitch);
1502
1503 /* Parse frame type ("frame header"), see frame_descs */
1504 3306 int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc, 6, 3)], block_nsamples;
1505
1506 3306 pitch[0] = INT_MAX;
1507
1508
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3306 times.
 3306 if (bd_idx < 0) {
1509 ✗ av_log(ctx, AV_LOG_ERROR,
1510 "Invalid frame type VLC code, skipping\n");
1511 ✗ return AVERROR_INVALIDDATA;
1512 }
1513
1514 3306 block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks;
1515
1516 /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */
1517
2/2
✓ Branch 0 taken 560 times.
✓ Branch 1 taken 2746 times.
 3306 if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) {
1518 /* Pitch is provided per frame, which is interpreted as the pitch of
1519 * the last sample of the last block of this frame. We can interpolate
1520 * the pitch of other blocks (and even pitch-per-sample) by gradually
1521 * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */
1522 560 n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1;
1523 560 log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1;
1524 560 cur_pitch_val = s->min_pitch_val + get_bits(gb, s->pitch_nbits);
1525
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 560 times.
 560 cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1);
1526
2/2
✓ Branch 0 taken 524 times.
✓ Branch 1 taken 36 times.
 560 if (s->last_acb_type == ACB_TYPE_NONE ||
1527 524 20 * abs(cur_pitch_val - s->last_pitch_val) >
1528
2/2
✓ Branch 0 taken 102 times.
✓ Branch 1 taken 422 times.
 524 (cur_pitch_val + s->last_pitch_val))
1529 138 s->last_pitch_val = cur_pitch_val;
1530
1531 /* pitch per block */
1532
2/2
✓ Branch 0 taken 1276 times.
✓ Branch 1 taken 560 times.
 1836 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1533 1276 int fac = n * 2 + 1;
1534
1535 1276 pitch[n] = (MUL16(fac, cur_pitch_val) +
1536 1276 MUL16((n_blocks_x2 - fac), s->last_pitch_val) +
1537 1276 frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2;
1538 }
1539
1540 /* "pitch-diff-per-sample" for calculation of pitch per sample */
1541 560 s->pitch_diff_sh16 =
1542 560 (cur_pitch_val - s->last_pitch_val) * (1 << 16) / MAX_FRAMESIZE;
1543 }
1544
1545 /* Global gain (if silence) and pitch-adaptive window coordinates */
1546
3/3
✓ Branch 0 taken 499 times.
✓ Branch 1 taken 341 times.
✓ Branch 2 taken 2466 times.
 3306 switch (frame_descs[bd_idx].fcb_type) {
1547 499 case FCB_TYPE_SILENCE:
1548 499 s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)];
1549 499 break;
1550 341 case FCB_TYPE_AW_PULSES:
1551 341 aw_parse_coords(s, gb, pitch);
1552 341 break;
1553 }
1554
1555
2/2
✓ Branch 0 taken 10783 times.
✓ Branch 1 taken 3306 times.
 14089 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1556 int bl_pitch_sh2;
1557
1558 /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */
1559
3/3
✓ Branch 0 taken 8464 times.
✓ Branch 1 taken 1276 times.
✓ Branch 2 taken 1043 times.
 10783 switch (frame_descs[bd_idx].acb_type) {
1560 8464 case ACB_TYPE_HAMMING: {
1561 /* Pitch is given per block. Per-block pitches are encoded as an
1562 * absolute value for the first block, and then delta values
1563 * relative to this value) for all subsequent blocks. The scale of
1564 * this pitch value is semi-logarithmic compared to its use in the
1565 * decoder, so we convert it to normal scale also. */
1566 int block_pitch,
1567 8464 t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2,
1568 8464 t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1,
1569 8464 t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1;
1570
1571
2/2
✓ Branch 0 taken 1975 times.
✓ Branch 1 taken 6489 times.
 8464 if (n == 0) {
1572 1975 block_pitch = get_bits(gb, s->block_pitch_nbits);
1573 } else
1574 6489 block_pitch = last_block_pitch - s->block_delta_pitch_hrange +
1575 6489 get_bits(gb, s->block_delta_pitch_nbits);
1576 /* Convert last_ so that any next delta is within _range */
1577 8464 last_block_pitch = av_clip(block_pitch,
1578 s->block_delta_pitch_hrange,
1579 8464 s->block_pitch_range -
1580 8464 s->block_delta_pitch_hrange);
1581
1582 /* Convert semi-log-style scale back to normal scale */
1583
2/2
✓ Branch 0 taken 1491 times.
✓ Branch 1 taken 6973 times.
 8464 if (block_pitch < t1) {
1584 1491 bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch;
1585 } else {
1586 6973 block_pitch -= t1;
1587
2/2
✓ Branch 0 taken 5712 times.
✓ Branch 1 taken 1261 times.
 6973 if (block_pitch < t2) {
1588 5712 bl_pitch_sh2 =
1589 5712 (s->block_conv_table[1] << 2) + (block_pitch << 1);
1590 } else {
1591 1261 block_pitch -= t2;
1592
1/2
✓ Branch 0 taken 1261 times.
✗ Branch 1 not taken.
 1261 if (block_pitch < t3) {
1593 1261 bl_pitch_sh2 =
1594 1261 (s->block_conv_table[2] + block_pitch) << 2;
1595 } else
1596 ✗ bl_pitch_sh2 = s->block_conv_table[3] << 2;
1597 }
1598 }
1599 8464 pitch[n] = bl_pitch_sh2 >> 2;
1600 8464 break;
1601 }
1602
1603 1276 case ACB_TYPE_ASYMMETRIC: {
1604 1276 bl_pitch_sh2 = pitch[n] << 2;
1605 1276 break;
1606 }
1607
1608 1043 default: // ACB_TYPE_NONE has no pitch
1609 1043 bl_pitch_sh2 = 0;
1610 1043 break;
1611 }
1612
1613 10783 synth_block(s, gb, n, block_nsamples, bl_pitch_sh2,
1614 lsps, prev_lsps, &frame_descs[bd_idx],
1615 10783 &excitation[n * block_nsamples],
1616 10783 &synth[n * block_nsamples]);
1617 }
1618
1619 /* Averaging projection filter, if applicable. Else, just copy samples
1620 * from synthesis buffer */
1621
1/2
✓ Branch 0 taken 3306 times.
✗ Branch 1 not taken.
 3306 if (s->do_apf) {
1622 double i_lsps[MAX_LSPS];
1623 float lpcs[MAX_LSPS];
1624
1625
3/4
✓ Branch 0 taken 2535 times.
✓ Branch 1 taken 771 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 2535 times.
 3306 if(frame_descs[bd_idx].fcb_type >= FCB_TYPE_AW_PULSES && pitch[0] == INT_MAX)
1626 ✗ return AVERROR_INVALIDDATA;
1627
1628
2/2
✓ Branch 0 taken 42960 times.
✓ Branch 1 taken 3306 times.
 46266 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1629 42960 i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n]));
1630 3306 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1631 3306 postfilter(s, synth, samples, 80, lpcs,
1632 3306 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx],
1633 3306 frame_descs[bd_idx].fcb_type, pitch[0]);
1634
1635
2/2
✓ Branch 0 taken 42960 times.
✓ Branch 1 taken 3306 times.
 46266 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1636 42960 i_lsps[n] = cos(lsps[n]);
1637 3306 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1638 3306 postfilter(s, &synth[80], &samples[80], 80, lpcs,
1639 3306 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80],
1640 3306 frame_descs[bd_idx].fcb_type, pitch[0]);
1641 } else
1642 ✗ memcpy(samples, synth, 160 * sizeof(synth[0]));
1643
1644 /* Cache values for next frame */
1645 3306 s->frame_cntr++;
1646
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3306 times.
 3306 if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%)
1647 3306 s->last_acb_type = frame_descs[bd_idx].acb_type;
1648
3/4
✓ Branch 0 taken 771 times.
✓ Branch 1 taken 560 times.
✓ Branch 2 taken 1975 times.
✗ Branch 3 not taken.
 3306 switch (frame_descs[bd_idx].acb_type) {
1649 771 case ACB_TYPE_NONE:
1650 771 s->last_pitch_val = 0;
1651 771 break;
1652 560 case ACB_TYPE_ASYMMETRIC:
1653 560 s->last_pitch_val = cur_pitch_val;
1654 560 break;
1655 1975 case ACB_TYPE_HAMMING:
1656 1975 s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1];
1657 1975 break;
1658 }
1659
1660 3306 return 0;
1661 }
1662
1663 /**
1664 * Ensure minimum value for first item, maximum value for last value,
1665 * proper spacing between each value and proper ordering.
1666 *
1667 * @param lsps array of LSPs
1668 * @param num size of LSP array
1669 *
1670 * @note basically a double version of #ff_acelp_reorder_lsf(), might be
1671 * useful to put in a generic location later on. Parts are also
1672 * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(),
1673 * which is in float.
1674 */
1675 3306 static void stabilize_lsps(double *lsps, int num)
1676 {
1677 int n, m, l;
1678
1679 /* set minimum value for first, maximum value for last and minimum
1680 * spacing between LSF values.
1681 * Very similar to ff_set_min_dist_lsf(), but in double. */
1682
1/2
✓ Branch 0 taken 3306 times.
✗ Branch 1 not taken.
 3306 lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI);
1683
2/2
✓ Branch 0 taken 39654 times.
✓ Branch 1 taken 3306 times.
 42960 for (n = 1; n < num; n++)
1684
2/2
✓ Branch 0 taken 39138 times.
✓ Branch 1 taken 516 times.
 39654 lsps[n] = FFMAX(lsps[n], lsps[n - 1] + 0.0125 * M_PI);
1685
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3306 times.
 3306 lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI);
1686
1687 /* reorder (looks like one-time / non-recursed bubblesort).
1688 * Very similar to ff_sort_nearly_sorted_floats(), but in double. */
1689
2/2
✓ Branch 0 taken 39654 times.
✓ Branch 1 taken 3306 times.
 42960 for (n = 1; n < num; n++) {
1690
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 39654 times.
 39654 if (lsps[n] < lsps[n - 1]) {
1691 ✗ for (m = 1; m < num; m++) {
1692 ✗ double tmp = lsps[m];
1693 ✗ for (l = m - 1; l >= 0; l--) {
1694 ✗ if (lsps[l] <= tmp) break;
1695 ✗ lsps[l + 1] = lsps[l];
1696 }
1697 ✗ lsps[l + 1] = tmp;
1698 }
1699 ✗ break;
1700 }
1701 }
1702 3306 }
1703
1704 /**
1705 * Synthesize output samples for a single superframe. If we have any data
1706 * cached in s->sframe_cache, that will be used instead of whatever is loaded
1707 * in s->gb.
1708 *
1709 * WMA Voice superframes contain 3 frames, each containing 160 audio samples,
1710 * to give a total of 480 samples per frame. See #synth_frame() for frame
1711 * parsing. In addition to 3 frames, superframes can also contain the LSPs
1712 * (if these are globally specified for all frames (residually); they can
1713 * also be specified individually per-frame. See the s->has_residual_lsps
1714 * option), and can specify the number of samples encoded in this superframe
1715 * (if less than 480), usually used to prevent blanks at track boundaries.
1716 *
1717 * @param ctx WMA Voice decoder context
1718 * @return 0 on success, <0 on error or 1 if there was not enough data to
1719 * fully parse the superframe
1720 */
1721 1102 static int synth_superframe(AVCodecContext *ctx, AVFrame *frame,
1722 int *got_frame_ptr)
1723 {
1724 1102 WMAVoiceContext *s = ctx->priv_data;
1725 1102 GetBitContext *gb = &s->gb, s_gb;
1726 1102 int n, res, n_samples = MAX_SFRAMESIZE;
1727 double lsps[MAX_FRAMES][MAX_LSPS];
1728 2204 const double *mean_lsf = s->lsps == 16 ?
1729
2/2
✓ Branch 0 taken 550 times.
✓ Branch 1 taken 552 times.
 1102 wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode];
1730 float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12];
1731 float synth[MAX_LSPS + MAX_SFRAMESIZE];
1732 float *samples;
1733
1734 1102 memcpy(synth, s->synth_history,
1735 1102 s->lsps * sizeof(*synth));
1736 1102 memcpy(excitation, s->excitation_history,
1737 1102 s->history_nsamples * sizeof(*excitation));
1738
1739
2/2
✓ Branch 0 taken 185 times.
✓ Branch 1 taken 917 times.
 1102 if (s->sframe_cache_size > 0) {
1740 185 gb = &s_gb;
1741 185 init_get_bits(gb, s->sframe_cache, s->sframe_cache_size);
1742 185 s->sframe_cache_size = 0;
1743 }
1744
1745 /* First bit is speech/music bit, it differentiates between WMAVoice
1746 * speech samples (the actual codec) and WMAVoice music samples, which
1747 * are really WMAPro-in-WMAVoice-superframes. I've never seen those in
1748 * the wild yet. */
1749
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 1102 times.
 1102 if (!get_bits1(gb)) {
1750 ✗ avpriv_request_sample(ctx, "WMAPro-in-WMAVoice");
1751 ✗ return AVERROR_PATCHWELCOME;
1752 }
1753
1754 /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */
1755
2/2
✓ Branch 1 taken 3 times.
✓ Branch 2 taken 1099 times.
 1102 if (get_bits1(gb)) {
1756
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 3 times.
 3 if ((n_samples = get_bits(gb, 12)) > MAX_SFRAMESIZE) {
1757 ✗ av_log(ctx, AV_LOG_ERROR,
1758 "Superframe encodes > %d samples (%d), not allowed\n",
1759 MAX_SFRAMESIZE, n_samples);
1760 ✗ return AVERROR_INVALIDDATA;
1761 }
1762 }
1763
1764 /* Parse LSPs, if global for the superframe (can also be per-frame). */
1765
1/2
✓ Branch 0 taken 1102 times.
✗ Branch 1 not taken.
 1102 if (s->has_residual_lsps) {
1766 double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2];
1767
1768
2/2
✓ Branch 0 taken 14320 times.
✓ Branch 1 taken 1102 times.
 15422 for (n = 0; n < s->lsps; n++)
1769 14320 prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n];
1770
1771
2/2
✓ Branch 0 taken 552 times.
✓ Branch 1 taken 550 times.
 1102 if (s->lsps == 10) {
1772 552 dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1773 } else /* s->lsps == 16 */
1774 550 dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1775
1776
2/2
✓ Branch 0 taken 14320 times.
✓ Branch 1 taken 1102 times.
 15422 for (n = 0; n < s->lsps; n++) {
1777 14320 lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]);
1778 14320 lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]);
1779 14320 lsps[2][n] += mean_lsf[n];
1780 }
1781
2/2
✓ Branch 0 taken 3306 times.
✓ Branch 1 taken 1102 times.
 4408 for (n = 0; n < 3; n++)
1782 3306 stabilize_lsps(lsps[n], s->lsps);
1783 }
1784
1785 /* synth_superframe can run multiple times per packet
1786 * free potential previous frame */
1787 1102 av_frame_unref(frame);
1788
1789 /* get output buffer */
1790 1102 frame->nb_samples = MAX_SFRAMESIZE;
1791
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 1102 times.
 1102 if ((res = ff_get_buffer(ctx, frame, 0)) < 0)
1792 ✗ return res;
1793 1102 frame->nb_samples = n_samples;
1794 1102 samples = (float *)frame->data[0];
1795
1796 /* Parse frames, optionally preceded by per-frame (independent) LSPs. */
1797
2/2
✓ Branch 0 taken 3306 times.
✓ Branch 1 taken 1102 times.
 4408 for (n = 0; n < 3; n++) {
1798
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3306 times.
 3306 if (!s->has_residual_lsps) {
1799 int m;
1800
1801 ✗ if (s->lsps == 10) {
1802 ✗ dequant_lsp10i(gb, lsps[n]);
1803 } else /* s->lsps == 16 */
1804 ✗ dequant_lsp16i(gb, lsps[n]);
1805
1806 ✗ for (m = 0; m < s->lsps; m++)
1807 ✗ lsps[n][m] += mean_lsf[m];
1808 ✗ stabilize_lsps(lsps[n], s->lsps);
1809 }
1810
1811
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3306 times.
 4408 if ((res = synth_frame(ctx, gb, n,
1812 3306 &samples[n * MAX_FRAMESIZE],
1813 3306 lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1],
1814 3306 &excitation[s->history_nsamples + n * MAX_FRAMESIZE],
1815
2/2
✓ Branch 0 taken 1102 times.
✓ Branch 1 taken 2204 times.
 3306 &synth[s->lsps + n * MAX_FRAMESIZE]))) {
1816 ✗ *got_frame_ptr = 0;
1817 ✗ return res;
1818 }
1819 }
1820
1821 /* Statistics? FIXME - we don't check for length, a slight overrun
1822 * will be caught by internal buffer padding, and anything else
1823 * will be skipped, not read. */
1824
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 1102 times.
 1102 if (get_bits1(gb)) {
1825 ✗ res = get_bits(gb, 4);
1826 ✗ skip_bits(gb, 10 * (res + 1));
1827 }
1828
1829
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 1102 times.
 1102 if (get_bits_left(gb) < 0) {
1830 ✗ wmavoice_flush(ctx);
1831 ✗ return AVERROR_INVALIDDATA;
1832 }
1833
1834 1102 *got_frame_ptr = 1;
1835
1836 /* Update history */
1837 1102 memcpy(s->prev_lsps, lsps[2],
1838 1102 s->lsps * sizeof(*s->prev_lsps));
1839 1102 memcpy(s->synth_history, &synth[MAX_SFRAMESIZE],
1840 1102 s->lsps * sizeof(*synth));
1841 1102 memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE],
1842 1102 s->history_nsamples * sizeof(*excitation));
1843
1/2
✓ Branch 0 taken 1102 times.
✗ Branch 1 not taken.
 1102 if (s->do_apf)
1844 1102 memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE],
1845 1102 s->history_nsamples * sizeof(*s->zero_exc_pf));
1846
1847 1102 return 0;
1848 }
1849
1850 /**
1851 * Parse the packet header at the start of each packet (input data to this
1852 * decoder).
1853 *
1854 * @param s WMA Voice decoding context private data
1855 * @return <0 on error, nb_superframes on success.
1856 */
1857 186 static int parse_packet_header(WMAVoiceContext *s)
1858 {
1859 186 GetBitContext *gb = &s->gb;
1860 186 unsigned int res, n_superframes = 0;
1861
1862 186 skip_bits(gb, 4); // packet sequence number
1863 186 s->has_residual_lsps = get_bits1(gb);
1864 do {
1865
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 186 times.
 186 if (get_bits_left(gb) < 6 + s->spillover_bitsize)
1866 ✗ return AVERROR_INVALIDDATA;
1867
1868 186 res = get_bits(gb, 6); // number of superframes per packet
1869 // (minus first one if there is spillover)
1870 186 n_superframes += res;
1871
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 186 times.
 186 } while (res == 0x3F);
1872 186 s->spillover_nbits = get_bits(gb, s->spillover_bitsize);
1873
1874
1/2
✓ Branch 1 taken 186 times.
✗ Branch 2 not taken.
 186 return get_bits_left(gb) >= 0 ? n_superframes : AVERROR_INVALIDDATA;
1875 }
1876
1877 /**
1878 * Copy (unaligned) bits from gb/data/size to pb.
1879 *
1880 * @param pb target buffer to copy bits into
1881 * @param data source buffer to copy bits from
1882 * @param size size of the source data, in bytes
1883 * @param gb bit I/O context specifying the current position in the source.
1884 * data. This function might use this to align the bit position to
1885 * a whole-byte boundary before calling #ff_copy_bits() on aligned
1886 * source data
1887 * @param nbits the amount of bits to copy from source to target
1888 *
1889 * @note after calling this function, the current position in the input bit
1890 * I/O context is undefined.
1891 */
1892 370 static void copy_bits(PutBitContext *pb,
1893 const uint8_t *data, int size,
1894 GetBitContext *gb, int nbits)
1895 {
1896 int rmn_bytes, rmn_bits;
1897
1898 370 rmn_bits = rmn_bytes = get_bits_left(gb);
1899
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 370 times.
 370 if (rmn_bits < nbits)
1900 ✗ return;
1901
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 370 times.
 370 if (nbits > put_bits_left(pb))
1902 ✗ return;
1903 370 rmn_bits &= 7; rmn_bytes >>= 3;
1904
2/2
✓ Branch 0 taken 290 times.
✓ Branch 1 taken 80 times.
 370 if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0)
1905 290 put_bits(pb, rmn_bits, get_bits(gb, rmn_bits));
1906 370 ff_copy_bits(pb, data + size - rmn_bytes,
1907 370 FFMIN(nbits - rmn_bits, rmn_bytes << 3));
1908 }
1909
1910 /**
1911 * Packet decoding: a packet is anything that the (ASF) demuxer contains,
1912 * and we expect that the demuxer / application provides it to us as such
1913 * (else you'll probably get garbage as output). Every packet has a size of
1914 * ctx->block_align bytes, starts with a packet header (see
1915 * #parse_packet_header()), and then a series of superframes. Superframe
1916 * boundaries may exceed packets, i.e. superframes can split data over
1917 * multiple (two) packets.
1918 *
1919 * For more information about frames, see #synth_superframe().
1920 */
1921 1291 static int wmavoice_decode_packet(AVCodecContext *ctx, AVFrame *frame,
1922 int *got_frame_ptr, AVPacket *avpkt)
1923 {
1924 1291 WMAVoiceContext *s = ctx->priv_data;
1925 1291 GetBitContext *gb = &s->gb;
1926 1291 const uint8_t *buf = avpkt->data;
1927 uint8_t dummy[1];
1928 int size, res, pos;
1929
1930 /* Packets are sometimes a multiple of ctx->block_align, with a packet
1931 * header at each ctx->block_align bytes. However, FFmpeg's ASF demuxer
1932 * feeds us ASF packets, which may concatenate multiple "codec" packets
1933 * in a single "muxer" packet, so we artificially emulate that by
1934 * capping the packet size at ctx->block_align. */
1935
2/2
✓ Branch 0 taken 180 times.
✓ Branch 1 taken 1291 times.
 1471 for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align);
1936
2/2
✓ Branch 0 taken 5 times.
✓ Branch 1 taken 1286 times.
 1291 buf = size ? buf : dummy;
1937 1291 res = init_get_bits8(&s->gb, buf, size);
1938
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 1291 times.
 1291 if (res < 0)
1939 ✗ return res;
1940
1941 /* size == ctx->block_align is used to indicate whether we are dealing with
1942 * a new packet or a packet of which we already read the packet header
1943 * previously. */
1944
2/2
✓ Branch 0 taken 191 times.
✓ Branch 1 taken 1100 times.
 1291 if (!(size % ctx->block_align)) { // new packet header
1945
2/2
✓ Branch 0 taken 5 times.
✓ Branch 1 taken 186 times.
 191 if (!size) {
1946 5 s->spillover_nbits = 0;
1947 5 s->nb_superframes = 0;
1948 } else {
1949
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 186 times.
 186 if ((res = parse_packet_header(s)) < 0)
1950 ✗ return res;
1951 186 s->nb_superframes = res;
1952 }
1953
1954 /* If the packet header specifies a s->spillover_nbits, then we want
1955 * to push out all data of the previous packet (+ spillover) before
1956 * continuing to parse new superframes in the current packet. */
1957
2/2
✓ Branch 0 taken 185 times.
✓ Branch 1 taken 6 times.
 191 if (s->sframe_cache_size > 0) {
1958 185 int cnt = get_bits_count(gb);
1959
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 185 times.
 185 if (cnt + s->spillover_nbits > avpkt->size * 8) {
1960 ✗ s->spillover_nbits = avpkt->size * 8 - cnt;
1961 }
1962 185 copy_bits(&s->pb, buf, size, gb, s->spillover_nbits);
1963 185 flush_put_bits(&s->pb);
1964 185 s->sframe_cache_size += s->spillover_nbits;
1965
1/2
✓ Branch 1 taken 185 times.
✗ Branch 2 not taken.
 185 if ((res = synth_superframe(ctx, frame, got_frame_ptr)) == 0 &&
1966
1/2
✓ Branch 0 taken 185 times.
✗ Branch 1 not taken.
 185 *got_frame_ptr) {
1967 185 cnt += s->spillover_nbits;
1968 185 s->skip_bits_next = cnt & 7;
1969 185 res = cnt >> 3;
1970 185 return res;
1971 } else
1972 ✗ skip_bits_long (gb, s->spillover_nbits - cnt +
1973 ✗ get_bits_count(gb)); // resync
1974
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 6 times.
 6 } else if (s->spillover_nbits) {
1975 ✗ skip_bits_long(gb, s->spillover_nbits); // resync
1976 }
1977
2/2
✓ Branch 0 taken 971 times.
✓ Branch 1 taken 129 times.
 1100 } else if (s->skip_bits_next)
1978 971 skip_bits(gb, s->skip_bits_next);
1979
1980 /* Try parsing superframes in current packet */
1981 1106 s->sframe_cache_size = 0;
1982 1106 s->skip_bits_next = 0;
1983 1106 pos = get_bits_left(gb);
1984
2/2
✓ Branch 0 taken 4 times.
✓ Branch 1 taken 1102 times.
 1106 if (s->nb_superframes-- == 0) {
1985 4 *got_frame_ptr = 0;
1986 4 return size;
1987
2/2
✓ Branch 0 taken 917 times.
✓ Branch 1 taken 185 times.
 1102 } else if (s->nb_superframes > 0) {
1988
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 917 times.
 917 if ((res = synth_superframe(ctx, frame, got_frame_ptr)) < 0) {
1989 ✗ return res;
1990
1/2
✓ Branch 0 taken 917 times.
✗ Branch 1 not taken.
 917 } else if (*got_frame_ptr) {
1991 917 int cnt = get_bits_count(gb);
1992 917 s->skip_bits_next = cnt & 7;
1993 917 res = cnt >> 3;
1994 917 return res;
1995 }
1996
1/2
✓ Branch 0 taken 185 times.
✗ Branch 1 not taken.
 185 } else if ((s->sframe_cache_size = pos) > 0) {
1997 /* ... cache it for spillover in next packet */
1998 185 init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE);
1999 185 copy_bits(&s->pb, buf, size, gb, s->sframe_cache_size);
2000 // FIXME bad - just copy bytes as whole and add use the
2001 // skip_bits_next field
2002 }
2003
2004 185 return size;
2005 }
2006
2007 8 static av_cold int wmavoice_decode_end(AVCodecContext *ctx)
2008 {
2009 8 WMAVoiceContext *s = ctx->priv_data;
2010
2011
1/2
✓ Branch 0 taken 8 times.
✗ Branch 1 not taken.
 8 if (s->do_apf) {
2012 8 av_tx_uninit(&s->rdft);
2013 8 av_tx_uninit(&s->irdft);
2014 8 av_tx_uninit(&s->dct);
2015 8 av_tx_uninit(&s->dst);
2016 }
2017
2018 8 return 0;
2019 }
2020
2021 const FFCodec ff_wmavoice_decoder = {
2022 .p.name = "wmavoice",
2023 CODEC_LONG_NAME("Windows Media Audio Voice"),
2024 .p.type = AVMEDIA_TYPE_AUDIO,
2025 .p.id = AV_CODEC_ID_WMAVOICE,
2026 .priv_data_size = sizeof(WMAVoiceContext),
2027 .init = wmavoice_decode_init,
2028 .close = wmavoice_decode_end,
2029 FF_CODEC_DECODE_CB(wmavoice_decode_packet),
2030 .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_DELAY,
2031 .caps_internal = FF_CODEC_CAP_INIT_CLEANUP,
2032 .flush = wmavoice_flush,
2033 };
2034