FFmpeg coverage

Directory: ../../../ffmpeg/
File: src/libavcodec/wmavoice.c
Date: 2024-11-20 23:03:26
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];
253 ///< cache of the signal of previous
254 ///< superframes, used as a history for
255 ///< signal generation
256 float synth_history[MAX_LSPS]; ///< see #excitation_history
257 /**
258 * @}
259 *
260 * @name Postfilter values
261 *
262 * Variables used for postfilter implementation, mostly history for
263 * smoothing and so on, and context variables for FFT/iFFT.
264 * @{
265 */
266 AVTXContext *rdft, *irdft; ///< contexts for FFT-calculation in the
267 av_tx_fn rdft_fn, irdft_fn; ///< postfilter (for denoise filter)
268 AVTXContext *dct, *dst; ///< contexts for phase shift (in Hilbert
269 av_tx_fn dct_fn, dst_fn; ///< transform, part of postfilter)
270 float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi]
271 ///< range
272 float postfilter_agc; ///< gain control memory, used in
273 ///< #adaptive_gain_control()
274 float dcf_mem[2]; ///< DC filter history
275 float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE];
276 ///< zero filter output (i.e. excitation)
277 ///< by postfilter
278 float denoise_filter_cache[MAX_FRAMESIZE];
279 int denoise_filter_cache_size; ///< samples in #denoise_filter_cache
280 DECLARE_ALIGNED(32, float, tilted_lpcs_pf)[0x82];
281 ///< aligned buffer for LPC tilting
282 DECLARE_ALIGNED(32, float, denoise_coeffs_pf)[0x82];
283 ///< aligned buffer for denoise coefficients
284 DECLARE_ALIGNED(32, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16];
285 ///< aligned buffer for postfilter speech
286 ///< synthesis
287 /**
288 * @}
289 */
290 } WMAVoiceContext;
291
292 /**
293 * Set up the variable bit mode (VBM) tree from container extradata.
294 * @param gb bit I/O context.
295 * The bit context (s->gb) should be loaded with byte 23-46 of the
296 * container extradata (i.e. the ones containing the VBM tree).
297 * @param vbm_tree pointer to array to which the decoded VBM tree will be
298 * written.
299 * @return 0 on success, <0 on error.
300 */
301 8 static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25])
302 {
303 8 int cntr[8] = { 0 }, n, res;
304
305 8 memset(vbm_tree, 0xff, sizeof(vbm_tree[0]) * 25);
306
2/2
✓ Branch 0 taken 136 times.
✓ Branch 1 taken 8 times.
 144 for (n = 0; n < 17; n++) {
307 136 res = get_bits(gb, 3);
308
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 136 times.
 136 if (cntr[res] > 3) // should be >= 3 + (res == 7))
309 ✗ return -1;
310 136 vbm_tree[res * 3 + cntr[res]++] = n;
311 }
312 8 return 0;
313 }
314
315 5 static av_cold void wmavoice_init_static_data(void)
316 {
317 static const uint8_t bits[] = {
318 2, 2, 2, 4, 4, 4,
319 6, 6, 6, 8, 8, 8,
320 10, 10, 10, 12, 12, 12,
321 14, 14, 14, 14
322 };
323
324 5 VLC_INIT_STATIC_TABLE_FROM_LENGTHS(frame_type_vlc, VLC_NBITS,
325 FF_ARRAY_ELEMS(bits), bits,
326 1, NULL, 0, 0, 0, 0);
327 5 }
328
329 ✗ static av_cold void wmavoice_flush(AVCodecContext *ctx)
330 {
331 ✗ WMAVoiceContext *s = ctx->priv_data;
332 int n;
333
334 ✗ s->postfilter_agc = 0;
335 ✗ s->sframe_cache_size = 0;
336 ✗ s->skip_bits_next = 0;
337 ✗ for (n = 0; n < s->lsps; n++)
338 ✗ s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
339 ✗ memset(s->excitation_history, 0,
340 sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY);
341 ✗ memset(s->synth_history, 0,
342 sizeof(*s->synth_history) * MAX_LSPS);
343 ✗ memset(s->gain_pred_err, 0,
344 sizeof(s->gain_pred_err));
345
346 ✗ if (s->do_apf) {
347 ✗ memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0,
348 ✗ sizeof(*s->synth_filter_out_buf) * s->lsps);
349 ✗ memset(s->dcf_mem, 0,
350 sizeof(*s->dcf_mem) * 2);
351 ✗ memset(s->zero_exc_pf, 0,
352 ✗ sizeof(*s->zero_exc_pf) * s->history_nsamples);
353 ✗ memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache));
354 }
355 ✗ }
356
357 /**
358 * Set up decoder with parameters from demuxer (extradata etc.).
359 */
360 8 static av_cold int wmavoice_decode_init(AVCodecContext *ctx)
361 {
362 static AVOnce init_static_once = AV_ONCE_INIT;
363 int n, flags, pitch_range, lsp16_flag, ret;
364 8 WMAVoiceContext *s = ctx->priv_data;
365
366 8 ff_thread_once(&init_static_once, wmavoice_init_static_data);
367
368 /**
369 * Extradata layout:
370 * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c),
371 * - byte 19-22: flags field (annoyingly in LE; see below for known
372 * values),
373 * - byte 23-46: variable bitmode tree (really just 17 * 3 bits,
374 * rest is 0).
375 */
376
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
 8 if (ctx->extradata_size != 46) {
377 ✗ av_log(ctx, AV_LOG_ERROR,
378 "Invalid extradata size %d (should be 46)\n",
379 ctx->extradata_size);
380 ✗ return AVERROR_INVALIDDATA;
381 }
382
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)) {
383 ✗ av_log(ctx, AV_LOG_ERROR, "Invalid block alignment %d.\n", ctx->block_align);
384 ✗ return AVERROR_INVALIDDATA;
385 }
386
387 8 flags = AV_RL32(ctx->extradata + 18);
388 8 s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align);
389 8 s->do_apf = flags & 0x1;
390
1/2
✓ Branch 0 taken 8 times.
✗ Branch 1 not taken.
 8 if (s->do_apf) {
391 8 float scale = 1.0f;
392
393 8 ret = av_tx_init(&s->rdft, &s->rdft_fn, AV_TX_FLOAT_RDFT, 0, 1 << 7, &scale, 0);
394
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
 8 if (ret < 0)
395 ✗ return ret;
396
397 8 ret = av_tx_init(&s->irdft, &s->irdft_fn, AV_TX_FLOAT_RDFT, 1, 1 << 7, &scale, 0);
398
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
 8 if (ret < 0)
399 ✗ return ret;
400
401 8 scale = 1.0 / (1 << 6);
402 8 ret = av_tx_init(&s->dct, &s->dct_fn, AV_TX_FLOAT_DCT_I, 0, 1 << 6, &scale, 0);
403
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
 8 if (ret < 0)
404 ✗ return ret;
405
406 8 scale = 1.0 / (1 << 6);
407 8 ret = av_tx_init(&s->dst, &s->dst_fn, AV_TX_FLOAT_DST_I, 0, 1 << 6, &scale, 0);
408
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
 8 if (ret < 0)
409 ✗ return ret;
410
411 8 ff_sine_window_init(s->cos, 256);
412 8 memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0]));
413
2/2
✓ Branch 0 taken 2040 times.
✓ Branch 1 taken 8 times.
 2048 for (n = 0; n < 255; n++) {
414 2040 s->sin[n] = -s->sin[510 - n];
415 2040 s->cos[510 - n] = s->cos[n];
416 }
417 }
418 8 s->denoise_strength = (flags >> 2) & 0xF;
419
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
 8 if (s->denoise_strength >= 12) {
420 ✗ av_log(ctx, AV_LOG_ERROR,
421 "Invalid denoise filter strength %d (max=11)\n",
422 s->denoise_strength);
423 ✗ return AVERROR_INVALIDDATA;
424 }
425 8 s->denoise_tilt_corr = !!(flags & 0x40);
426 8 s->dc_level = (flags >> 7) & 0xF;
427 8 s->lsp_q_mode = !!(flags & 0x2000);
428 8 s->lsp_def_mode = !!(flags & 0x4000);
429 8 lsp16_flag = flags & 0x1000;
430
2/2
✓ Branch 0 taken 4 times.
✓ Branch 1 taken 4 times.
 8 if (lsp16_flag) {
431 4 s->lsps = 16;
432 } else {
433 4 s->lsps = 10;
434 }
435
2/2
✓ Branch 0 taken 104 times.
✓ Branch 1 taken 8 times.
 112 for (n = 0; n < s->lsps; n++)
436 104 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
437
438 8 init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3);
439
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 8 times.
 8 if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) {
440 ✗ av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n");
441 ✗ return AVERROR_INVALIDDATA;
442 }
443
444
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
 8 if (ctx->sample_rate >= INT_MAX / (256 * 37))
445 ✗ return AVERROR_INVALIDDATA;
446
447 8 s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8;
448 8 s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8;
449 8 pitch_range = s->max_pitch_val - s->min_pitch_val;
450
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
 8 if (pitch_range <= 0) {
451 ✗ av_log(ctx, AV_LOG_ERROR, "Invalid pitch range; broken extradata?\n");
452 ✗ return AVERROR_INVALIDDATA;
453 }
454 8 s->pitch_nbits = av_ceil_log2(pitch_range);
455 8 s->last_pitch_val = 40;
456 8 s->last_acb_type = ACB_TYPE_NONE;
457 8 s->history_nsamples = s->max_pitch_val + 8;
458
459
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) {
460 ✗ int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8,
461 ✗ max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8;
462
463 ✗ av_log(ctx, AV_LOG_ERROR,
464 "Unsupported samplerate %d (min=%d, max=%d)\n",
465 ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz
466
467 ✗ return AVERROR(ENOSYS);
468 }
469
470 8 s->block_conv_table[0] = s->min_pitch_val;
471 8 s->block_conv_table[1] = (pitch_range * 25) >> 6;
472 8 s->block_conv_table[2] = (pitch_range * 44) >> 6;
473 8 s->block_conv_table[3] = s->max_pitch_val - 1;
474 8 s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF;
475
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
 8 if (s->block_delta_pitch_hrange <= 0) {
476 ✗ av_log(ctx, AV_LOG_ERROR, "Invalid delta pitch hrange; broken extradata?\n");
477 ✗ return AVERROR_INVALIDDATA;
478 }
479 8 s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange);
480 8 s->block_pitch_range = s->block_conv_table[2] +
481 8 s->block_conv_table[3] + 1 +
482 8 2 * (s->block_conv_table[1] - 2 * s->min_pitch_val);
483 8 s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range);
484
485 8 av_channel_layout_uninit(&ctx->ch_layout);
486 8 ctx->ch_layout = (AVChannelLayout)AV_CHANNEL_LAYOUT_MONO;
487 8 ctx->sample_fmt = AV_SAMPLE_FMT_FLT;
488
489 8 return 0;
490 }
491
492 /**
493 * @name Postfilter functions
494 * Postfilter functions (gain control, wiener denoise filter, DC filter,
495 * kalman smoothening, plus surrounding code to wrap it)
496 * @{
497 */
498 /**
499 * Adaptive gain control (as used in postfilter).
500 *
501 * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except
502 * that the energy here is calculated using sum(abs(...)), whereas the
503 * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)).
504 *
505 * @param out output buffer for filtered samples
506 * @param in input buffer containing the samples as they are after the
507 * postfilter steps so far
508 * @param speech_synth input buffer containing speech synth before postfilter
509 * @param size input buffer size
510 * @param alpha exponential filter factor
511 * @param gain_mem pointer to filter memory (single float)
512 */
513 6612 static void adaptive_gain_control(float *out, const float *in,
514 const float *speech_synth,
515 int size, float alpha, float *gain_mem)
516 {
517 int i;
518 6612 float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor;
519 6612 float mem = *gain_mem;
520
521
2/2
✓ Branch 0 taken 528960 times.
✓ Branch 1 taken 6612 times.
 535572 for (i = 0; i < size; i++) {
522 528960 speech_energy += fabsf(speech_synth[i]);
523 528960 postfilter_energy += fabsf(in[i]);
524 }
525
1/2
✓ Branch 0 taken 6612 times.
✗ Branch 1 not taken.
 6612 gain_scale_factor = postfilter_energy == 0.0 ? 0.0 :
526 6612 (1.0 - alpha) * speech_energy / postfilter_energy;
527
528
2/2
✓ Branch 0 taken 528960 times.
✓ Branch 1 taken 6612 times.
 535572 for (i = 0; i < size; i++) {
529 528960 mem = alpha * mem + gain_scale_factor;
530 528960 out[i] = in[i] * mem;
531 }
532
533 6612 *gain_mem = mem;
534 6612 }
535
536 /**
537 * Kalman smoothing function.
538 *
539 * This function looks back pitch +/- 3 samples back into history to find
540 * the best fitting curve (that one giving the optimal gain of the two
541 * signals, i.e. the highest dot product between the two), and then
542 * uses that signal history to smoothen the output of the speech synthesis
543 * filter.
544 *
545 * @param s WMA Voice decoding context
546 * @param pitch pitch of the speech signal
547 * @param in input speech signal
548 * @param out output pointer for smoothened signal
549 * @param size input/output buffer size
550 *
551 * @returns -1 if no smoothening took place, e.g. because no optimal
552 * fit could be found, or 0 on success.
553 */
554 5070 static int kalman_smoothen(WMAVoiceContext *s, int pitch,
555 const float *in, float *out, int size)
556 {
557 int n;
558 5070 float optimal_gain = 0, dot;
559
2/2
✓ Branch 0 taken 44 times.
✓ Branch 1 taken 5026 times.
 5070 const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)],
560 5070 *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)],
561 5070 *best_hist_ptr = NULL;
562
563 /* find best fitting point in history */
564 do {
565 35388 dot = avpriv_scalarproduct_float_c(in, ptr, size);
566
2/2
✓ Branch 0 taken 12328 times.
✓ Branch 1 taken 23060 times.
 35388 if (dot > optimal_gain) {
567 12328 optimal_gain = dot;
568 12328 best_hist_ptr = ptr;
569 }
570
2/2
✓ Branch 0 taken 30318 times.
✓ Branch 1 taken 5070 times.
 35388 } while (--ptr >= end);
571
572
2/2
✓ Branch 0 taken 26 times.
✓ Branch 1 taken 5044 times.
 5070 if (optimal_gain <= 0)
573 26 return -1;
574 5044 dot = avpriv_scalarproduct_float_c(best_hist_ptr, best_hist_ptr, size);
575
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 5044 times.
 5044 if (dot <= 0) // would be 1.0
576 ✗ return -1;
577
578
2/2
✓ Branch 0 taken 4872 times.
✓ Branch 1 taken 172 times.
 5044 if (optimal_gain <= dot) {
579 4872 dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000
580 } else
581 172 dot = 0.625;
582
583 /* actual smoothing */
584
2/2
✓ Branch 0 taken 403520 times.
✓ Branch 1 taken 5044 times.
 408564 for (n = 0; n < size; n++)
585 403520 out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]);
586
587 5044 return 0;
588 }
589
590 /**
591 * Get the tilt factor of a formant filter from its transfer function
592 * @see #tilt_factor() in amrnbdec.c, which does essentially the same,
593 * but somehow (??) it does a speech synthesis filter in the
594 * middle, which is missing here
595 *
596 * @param lpcs LPC coefficients
597 * @param n_lpcs Size of LPC buffer
598 * @returns the tilt factor
599 */
600 7098 static float tilt_factor(const float *lpcs, int n_lpcs)
601 {
602 float rh0, rh1;
603
604 7098 rh0 = 1.0 + avpriv_scalarproduct_float_c(lpcs, lpcs, n_lpcs);
605 7098 rh1 = lpcs[0] + avpriv_scalarproduct_float_c(lpcs, &lpcs[1], n_lpcs - 1);
606
607 7098 return rh1 / rh0;
608 }
609
610 /**
611 * Derive denoise filter coefficients (in real domain) from the LPCs.
612 */
613 5614 static void calc_input_response(WMAVoiceContext *s, float *lpcs_src,
614 int fcb_type, float *coeffs_dst, int remainder)
615 {
616 5614 float last_coeff, min = 15.0, max = -15.0;
617 float irange, angle_mul, gain_mul, range, sq;
618 5614 LOCAL_ALIGNED_32(float, coeffs, [0x82]);
619 5614 LOCAL_ALIGNED_32(float, lpcs, [0x82]);
620 5614 LOCAL_ALIGNED_32(float, lpcs_dct, [0x82]);
621 int n, idx;
622
623 5614 memcpy(coeffs, coeffs_dst, 0x82*sizeof(float));
624
625 /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */
626 5614 s->rdft_fn(s->rdft, lpcs, lpcs_src, sizeof(float));
627 #define log_range(var, assign) do { \
628 float tmp = log10f(assign); var = tmp; \
629 max = FFMAX(max, tmp); min = FFMIN(min, tmp); \
630 } while (0)
631
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]);
632
2/2
✓ Branch 0 taken 353682 times.
✓ Branch 1 taken 5614 times.
 359296 for (n = 1; n < 64; n++)
633
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] +
634 lpcs[n * 2 + 1] * lpcs[n * 2 + 1]);
635
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]);
636 #undef log_range
637 5614 range = max - min;
638 5614 lpcs[64] = last_coeff;
639
640 /* Now, use this spectrum to pick out these frequencies with higher
641 * (relative) power/energy (which we then take to be "not noise"),
642 * and set up a table (still in lpc[]) of (relative) gains per frequency.
643 * These frequencies will be maintained, while others ("noise") will be
644 * decreased in the filter output. */
645 5614 irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63]
646
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) :
647 (5.0 / 14.7));
648 5614 angle_mul = gain_mul * (8.0 * M_LN10 / M_PI);
649
2/2
✓ Branch 0 taken 364910 times.
✓ Branch 1 taken 5614 times.
 370524 for (n = 0; n <= 64; n++) {
650 float pwr;
651
652 364910 idx = lrint((max - lpcs[n]) * irange - 1);
653 364910 idx = FFMAX(0, idx);
654 364910 pwr = wmavoice_denoise_power_table[s->denoise_strength][idx];
655 364910 lpcs[n] = angle_mul * pwr;
656
657 /* 70.57 =~ 1/log10(1.0331663) */
658 364910 idx = av_clipd((pwr * gain_mul - 0.0295) * 70.570526123, 0, INT_MAX / 2);
659
660
2/2
✓ Branch 0 taken 9151 times.
✓ Branch 1 taken 355759 times.
 364910 if (idx > 127) { // fall back if index falls outside table range
661 9151 coeffs[n] = wmavoice_energy_table[127] *
662 9151 powf(1.0331663, idx - 127);
663 } else
664 355759 coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)];
665 }
666
667 /* calculate the Hilbert transform of the gains, which we do (since this
668 * is a sine input) by doing a phase shift (in theory, H(sin())=cos()).
669 * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the
670 * "moment" of the LPCs in this filter. */
671 5614 s->dct_fn(s->dct, lpcs_dct, lpcs, sizeof(float));
672 5614 s->dst_fn(s->dst, lpcs, lpcs_dct, sizeof(float));
673
674 /* Split out the coefficient indexes into phase/magnitude pairs */
675 5614 idx = 255 + av_clip(lpcs[64], -255, 255);
676 5614 coeffs[0] = coeffs[0] * s->cos[idx];
677 5614 idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255);
678 5614 last_coeff = coeffs[64] * s->cos[idx];
679 5614 for (n = 63;; n--) {
680 353682 idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255);
681 179648 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
682 179648 coeffs[n * 2] = coeffs[n] * s->cos[idx];
683
684
2/2
✓ Branch 0 taken 5614 times.
✓ Branch 1 taken 174034 times.
 179648 if (!--n) break;
685
686 174034 idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255);
687 174034 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
688 174034 coeffs[n * 2] = coeffs[n] * s->cos[idx];
689 }
690 5614 coeffs[64] = last_coeff;
691
692 /* move into real domain */
693 5614 s->irdft_fn(s->irdft, coeffs_dst, coeffs, sizeof(AVComplexFloat));
694
695 /* tilt correction and normalize scale */
696 5614 memset(&coeffs_dst[remainder], 0, sizeof(coeffs_dst[0]) * (128 - remainder));
697
2/2
✓ Branch 0 taken 1484 times.
✓ Branch 1 taken 4130 times.
 5614 if (s->denoise_tilt_corr) {
698 1484 float tilt_mem = 0;
699
700 1484 coeffs_dst[remainder - 1] = 0;
701 1484 ff_tilt_compensation(&tilt_mem,
702 1484 -1.8 * tilt_factor(coeffs_dst, remainder - 1),
703 coeffs_dst, remainder);
704 }
705 5614 sq = (1.0 / 64.0) * sqrtf(1 / avpriv_scalarproduct_float_c(coeffs_dst, coeffs_dst,
706 remainder));
707
2/2
✓ Branch 0 taken 263858 times.
✓ Branch 1 taken 5614 times.
 269472 for (n = 0; n < remainder; n++)
708 263858 coeffs_dst[n] *= sq;
709 5614 }
710
711 /**
712 * This function applies a Wiener filter on the (noisy) speech signal as
713 * a means to denoise it.
714 *
715 * - take RDFT of LPCs to get the power spectrum of the noise + speech;
716 * - using this power spectrum, calculate (for each frequency) the Wiener
717 * filter gain, which depends on the frequency power and desired level
718 * of noise subtraction (when set too high, this leads to artifacts)
719 * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse
720 * of 4-8kHz);
721 * - by doing a phase shift, calculate the Hilbert transform of this array
722 * of per-frequency filter-gains to get the filtering coefficients;
723 * - smoothen/normalize/de-tilt these filter coefficients as desired;
724 * - take RDFT of noisy sound, apply the coefficients and take its IRDFT
725 * to get the denoised speech signal;
726 * - the leftover (i.e. output of the IRDFT on denoised speech data beyond
727 * the frame boundary) are saved and applied to subsequent frames by an
728 * overlap-add method (otherwise you get clicking-artifacts).
729 *
730 * @param s WMA Voice decoding context
731 * @param fcb_type Frame (codebook) type
732 * @param synth_pf input: the noisy speech signal, output: denoised speech
733 * data; should be 16-byte aligned (for ASM purposes)
734 * @param size size of the speech data
735 * @param lpcs LPCs used to synthesize this frame's speech data
736 */
737 6612 static void wiener_denoise(WMAVoiceContext *s, int fcb_type,
738 float *synth_pf, int size,
739 const float *lpcs)
740 {
741 int remainder, lim, n;
742
743
2/2
✓ Branch 0 taken 5614 times.
✓ Branch 1 taken 998 times.
 6612 if (fcb_type != FCB_TYPE_SILENCE) {
744 5614 LOCAL_ALIGNED_32(float, coeffs_f, [0x82]);
745 5614 LOCAL_ALIGNED_32(float, synth_f, [0x82]);
746 5614 float *tilted_lpcs = s->tilted_lpcs_pf,
747 5614 *coeffs = s->denoise_coeffs_pf, tilt_mem = 0;
748
749 5614 tilted_lpcs[0] = 1.0;
750 5614 memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps);
751 5614 memset(&tilted_lpcs[s->lsps + 1], 0,
752 5614 sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1));
753 5614 ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps),
754 5614 tilted_lpcs, s->lsps + 2);
755
756 /* The IRDFT output (127 samples for 7-bit filter) beyond the frame
757 * size is applied to the next frame. All input beyond this is zero,
758 * and thus all output beyond this will go towards zero, hence we can
759 * limit to min(size-1, 127-size) as a performance consideration. */
760
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 5614 times.
 5614 remainder = FFMIN(127 - size, size - 1);
761 5614 calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder);
762
763 /* apply coefficients (in frequency spectrum domain), i.e. complex
764 * number multiplication */
765 5614 memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size));
766 5614 s->rdft_fn(s->rdft, synth_f, synth_pf, sizeof(float));
767 5614 s->rdft_fn(s->rdft, coeffs_f, coeffs, sizeof(float));
768 5614 synth_f[0] *= coeffs_f[0];
769 5614 synth_f[1] *= coeffs_f[1];
770
2/2
✓ Branch 0 taken 359296 times.
✓ Branch 1 taken 5614 times.
 364910 for (n = 1; n <= 64; n++) {
771 359296 float v1 = synth_f[n * 2], v2 = synth_f[n * 2 + 1];
772 359296 synth_f[n * 2] = v1 * coeffs_f[n * 2] - v2 * coeffs_f[n * 2 + 1];
773 359296 synth_f[n * 2 + 1] = v2 * coeffs_f[n * 2] + v1 * coeffs_f[n * 2 + 1];
774 }
775 5614 s->irdft_fn(s->irdft, synth_pf, synth_f, sizeof(AVComplexFloat));
776 }
777
778 /* merge filter output with the history of previous runs */
779
2/2
✓ Branch 0 taken 5612 times.
✓ Branch 1 taken 1000 times.
 6612 if (s->denoise_filter_cache_size) {
780 5612 lim = FFMIN(s->denoise_filter_cache_size, size);
781
2/2
✓ Branch 0 taken 263764 times.
✓ Branch 1 taken 5612 times.
 269376 for (n = 0; n < lim; n++)
782 263764 synth_pf[n] += s->denoise_filter_cache[n];
783 5612 s->denoise_filter_cache_size -= lim;
784 5612 memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size],
785 5612 sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size);
786 }
787
788 /* move remainder of filter output into a cache for future runs */
789
2/2
✓ Branch 0 taken 5614 times.
✓ Branch 1 taken 998 times.
 6612 if (fcb_type != FCB_TYPE_SILENCE) {
790 5614 lim = FFMIN(remainder, s->denoise_filter_cache_size);
791
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 5614 times.
 5614 for (n = 0; n < lim; n++)
792 ✗ s->denoise_filter_cache[n] += synth_pf[size + n];
793
1/2
✓ Branch 0 taken 5614 times.
✗ Branch 1 not taken.
 5614 if (lim < remainder) {
794 5614 memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim],
795 5614 sizeof(s->denoise_filter_cache[0]) * (remainder - lim));
796 5614 s->denoise_filter_cache_size = remainder;
797 }
798 }
799 6612 }
800
801 /**
802 * Averaging projection filter, the postfilter used in WMAVoice.
803 *
804 * This uses the following steps:
805 * - A zero-synthesis filter (generate excitation from synth signal)
806 * - Kalman smoothing on excitation, based on pitch
807 * - Re-synthesized smoothened output
808 * - Iterative Wiener denoise filter
809 * - Adaptive gain filter
810 * - DC filter
811 *
812 * @param s WMAVoice decoding context
813 * @param synth Speech synthesis output (before postfilter)
814 * @param samples Output buffer for filtered samples
815 * @param size Buffer size of synth & samples
816 * @param lpcs Generated LPCs used for speech synthesis
817 * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned)
818 * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses)
819 * @param pitch Pitch of the input signal
820 */
821 6612 static void postfilter(WMAVoiceContext *s, const float *synth,
822 float *samples, int size,
823 const float *lpcs, float *zero_exc_pf,
824 int fcb_type, int pitch)
825 {
826 float synth_filter_in_buf[MAX_FRAMESIZE / 2],
827 6612 *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16],
828 6612 *synth_filter_in = zero_exc_pf;
829
830
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 6612 times.
 6612 av_assert0(size <= MAX_FRAMESIZE / 2);
831
832 /* generate excitation from input signal */
833 6612 ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps);
834
835
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 &&
836 5070 !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size))
837 5044 synth_filter_in = synth_filter_in_buf;
838
839 /* re-synthesize speech after smoothening, and keep history */
840 6612 ff_celp_lp_synthesis_filterf(synth_pf, lpcs,
841 synth_filter_in, size, s->lsps);
842 6612 memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps],
843 6612 sizeof(synth_pf[0]) * s->lsps);
844
845 6612 wiener_denoise(s, fcb_type, synth_pf, size, lpcs);
846
847 6612 adaptive_gain_control(samples, synth_pf, synth, size, 0.99,
848 &s->postfilter_agc);
849
850
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 6612 times.
 6612 if (s->dc_level > 8) {
851 /* remove ultra-low frequency DC noise / highpass filter;
852 * coefficients are identical to those used in SIPR decoding,
853 * and very closely resemble those used in AMR-NB decoding. */
854 ✗ ff_acelp_apply_order_2_transfer_function(samples, samples,
855 ✗ (const float[2]) { -1.99997, 1.0 },
856 ✗ (const float[2]) { -1.9330735188, 0.93589198496 },
857 ✗ 0.93980580475, s->dcf_mem, size);
858 }
859 6612 }
860 /**
861 * @}
862 */
863
864 /**
865 * Dequantize LSPs
866 * @param lsps output pointer to the array that will hold the LSPs
867 * @param num number of LSPs to be dequantized
868 * @param values quantized values, contains n_stages values
869 * @param sizes range (i.e. max value) of each quantized value
870 * @param n_stages number of dequantization runs
871 * @param table dequantization table to be used
872 * @param mul_q LSF multiplier
873 * @param base_q base (lowest) LSF values
874 */
875 4404 static void dequant_lsps(double *lsps, int num,
876 const uint16_t *values,
877 const uint16_t *sizes,
878 int n_stages, const uint8_t *table,
879 const double *mul_q,
880 const double *base_q)
881 {
882 int n, m;
883
884 4404 memset(lsps, 0, num * sizeof(*lsps));
885
2/2
✓ Branch 0 taken 8264 times.
✓ Branch 1 taken 4404 times.
 12668 for (n = 0; n < n_stages; n++) {
886 8264 const uint8_t *t_off = &table[values[n] * num];
887 8264 double base = base_q[n], mul = mul_q[n];
888
889
2/2
✓ Branch 0 taken 87100 times.
✓ Branch 1 taken 8264 times.
 95364 for (m = 0; m < num; m++)
890 87100 lsps[m] += base + mul * t_off[m];
891
892 8264 table += sizes[n] * num;
893 }
894 4404 }
895
896 /**
897 * @name LSP dequantization routines
898 * LSP dequantization routines, for 10/16LSPs and independent/residual coding.
899 * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;
900 * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits.
901 * @{
902 */
903 /**
904 * Parse 10 independently-coded LSPs.
905 */
906 552 static void dequant_lsp10i(GetBitContext *gb, double *lsps)
907 {
908 static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 };
909 static const double mul_lsf[4] = {
910 5.2187144800e-3, 1.4626986422e-3,
911 9.6179549166e-4, 1.1325736225e-3
912 };
913 static const double base_lsf[4] = {
914 M_PI * -2.15522e-1, M_PI * -6.1646e-2,
915 M_PI * -3.3486e-2, M_PI * -5.7408e-2
916 };
917 uint16_t v[4];
918
919 552 v[0] = get_bits(gb, 8);
920 552 v[1] = get_bits(gb, 6);
921 552 v[2] = get_bits(gb, 5);
922 552 v[3] = get_bits(gb, 5);
923
924 552 dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i,
925 mul_lsf, base_lsf);
926 552 }
927
928 /**
929 * Parse 10 independently-coded LSPs, and then derive the tables to
930 * generate LSPs for the other frames from them (residual coding).
931 */
932 552 static void dequant_lsp10r(GetBitContext *gb,
933 double *i_lsps, const double *old,
934 double *a1, double *a2, int q_mode)
935 {
936 static const uint16_t vec_sizes[3] = { 128, 64, 64 };
937 static const double mul_lsf[3] = {
938 2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3
939 };
940 static const double base_lsf[3] = {
941 M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2
942 };
943 552 const float (*ipol_tab)[2][10] = q_mode ?
944
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 552 times.
 552 wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a;
945 uint16_t interpol, v[3];
946 int n;
947
948 552 dequant_lsp10i(gb, i_lsps);
949
950 552 interpol = get_bits(gb, 5);
951 552 v[0] = get_bits(gb, 7);
952 552 v[1] = get_bits(gb, 6);
953 552 v[2] = get_bits(gb, 6);
954
955
2/2
✓ Branch 0 taken 5520 times.
✓ Branch 1 taken 552 times.
 6072 for (n = 0; n < 10; n++) {
956 5520 double delta = old[n] - i_lsps[n];
957 5520 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
958 5520 a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
959 }
960
961 552 dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r,
962 mul_lsf, base_lsf);
963 552 }
964
965 /**
966 * Parse 16 independently-coded LSPs.
967 */
968 550 static void dequant_lsp16i(GetBitContext *gb, double *lsps)
969 {
970 static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 };
971 static const double mul_lsf[5] = {
972 3.3439586280e-3, 6.9908173703e-4,
973 3.3216608306e-3, 1.0334960326e-3,
974 3.1899104283e-3
975 };
976 static const double base_lsf[5] = {
977 M_PI * -1.27576e-1, M_PI * -2.4292e-2,
978 M_PI * -1.28094e-1, M_PI * -3.2128e-2,
979 M_PI * -1.29816e-1
980 };
981 uint16_t v[5];
982
983 550 v[0] = get_bits(gb, 8);
984 550 v[1] = get_bits(gb, 6);
985 550 v[2] = get_bits(gb, 7);
986 550 v[3] = get_bits(gb, 6);
987 550 v[4] = get_bits(gb, 7);
988
989 550 dequant_lsps( lsps, 5, v, vec_sizes, 2,
990 wmavoice_dq_lsp16i1, mul_lsf, base_lsf);
991 550 dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2,
992 wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]);
993 550 dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1,
994 wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]);
995 550 }
996
997 /**
998 * Parse 16 independently-coded LSPs, and then derive the tables to
999 * generate LSPs for the other frames from them (residual coding).
1000 */
1001 550 static void dequant_lsp16r(GetBitContext *gb,
1002 double *i_lsps, const double *old,
1003 double *a1, double *a2, int q_mode)
1004 {
1005 static const uint16_t vec_sizes[3] = { 128, 128, 128 };
1006 static const double mul_lsf[3] = {
1007 1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3
1008 };
1009 static const double base_lsf[3] = {
1010 M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2
1011 };
1012 550 const float (*ipol_tab)[2][16] = q_mode ?
1013
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 550 times.
 550 wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a;
1014 uint16_t interpol, v[3];
1015 int n;
1016
1017 550 dequant_lsp16i(gb, i_lsps);
1018
1019 550 interpol = get_bits(gb, 5);
1020 550 v[0] = get_bits(gb, 7);
1021 550 v[1] = get_bits(gb, 7);
1022 550 v[2] = get_bits(gb, 7);
1023
1024
2/2
✓ Branch 0 taken 8800 times.
✓ Branch 1 taken 550 times.
 9350 for (n = 0; n < 16; n++) {
1025 8800 double delta = old[n] - i_lsps[n];
1026 8800 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
1027 8800 a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
1028 }
1029
1030 550 dequant_lsps( a2, 10, v, vec_sizes, 1,
1031 wmavoice_dq_lsp16r1, mul_lsf, base_lsf);
1032 550 dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1,
1033 wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]);
1034 550 dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1,
1035 wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]);
1036 550 }
1037
1038 /**
1039 * @}
1040 * @name Pitch-adaptive window coding functions
1041 * The next few functions are for pitch-adaptive window coding.
1042 * @{
1043 */
1044 /**
1045 * Parse the offset of the first pitch-adaptive window pulses, and
1046 * the distribution of pulses between the two blocks in this frame.
1047 * @param s WMA Voice decoding context private data
1048 * @param gb bit I/O context
1049 * @param pitch pitch for each block in this frame
1050 */
1051 341 static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb,
1052 const int *pitch)
1053 {
1054 static const int16_t start_offset[94] = {
1055 -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11,
1056 13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26,
1057 27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43,
1058 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67,
1059 69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91,
1060 93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115,
1061 117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139,
1062 141, 143, 145, 147, 149, 151, 153, 155, 157, 159
1063 };
1064 int bits, offset;
1065
1066 /* position of pulse */
1067 341 s->aw_idx_is_ext = 0;
1068
2/2
✓ Branch 1 taken 10 times.
✓ Branch 2 taken 331 times.
 341 if ((bits = get_bits(gb, 6)) >= 54) {
1069 10 s->aw_idx_is_ext = 1;
1070 10 bits += (bits - 54) * 3 + get_bits(gb, 2);
1071 }
1072
1073 /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count
1074 * the distribution of the pulses in each block contained in this frame. */
1075
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;
1076
2/2
✓ Branch 0 taken 50 times.
✓ Branch 1 taken 341 times.
 391 for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ;
1077 341 s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0];
1078 341 s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2;
1079 341 offset += s->aw_n_pulses[0] * pitch[0];
1080 341 s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1];
1081 341 s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2;
1082
1083 /* if continuing from a position before the block, reset position to
1084 * start of block (when corrected for the range over which it can be
1085 * spread in aw_pulse_set1()). */
1086
2/2
✓ Branch 0 taken 331 times.
✓ Branch 1 taken 10 times.
 341 if (start_offset[bits] < MAX_FRAMESIZE / 2) {
1087
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)
1088 56 s->aw_first_pulse_off[1] -= pitch[1];
1089
2/2
✓ Branch 0 taken 50 times.
✓ Branch 1 taken 281 times.
 331 if (start_offset[bits] < 0)
1090
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)
1091 50 s->aw_first_pulse_off[0] -= pitch[0];
1092 }
1093 341 }
1094
1095 /**
1096 * Apply second set of pitch-adaptive window pulses.
1097 * @param s WMA Voice decoding context private data
1098 * @param gb bit I/O context
1099 * @param block_idx block index in frame [0, 1]
1100 * @param fcb structure containing fixed codebook vector info
1101 * @return -1 on error, 0 otherwise
1102 */
1103 682 static int aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb,
1104 int block_idx, AMRFixed *fcb)
1105 {
1106 uint16_t use_mask_mem[9]; // only 5 are used, rest is padding
1107 682 uint16_t *use_mask = use_mask_mem + 2;
1108 /* in this function, idx is the index in the 80-bit (+ padding) use_mask
1109 * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits
1110 * of idx are the position of the bit within a particular item in the
1111 * array (0 being the most significant bit, and 15 being the least
1112 * significant bit), and the remainder (>> 4) is the index in the
1113 * use_mask[]-array. This is faster and uses less memory than using a
1114 * 80-byte/80-int array. */
1115 682 int pulse_off = s->aw_first_pulse_off[block_idx],
1116 682 pulse_start, n, idx, range, aidx, start_off = 0;
1117
1118 /* set offset of first pulse to within this block */
1119
2/2
✓ Branch 0 taken 657 times.
✓ Branch 1 taken 25 times.
 682 if (s->aw_n_pulses[block_idx] > 0)
1120
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 657 times.
 657 while (pulse_off + s->aw_pulse_range < 1)
1121 ✗ pulse_off += fcb->pitch_lag;
1122
1123 /* find range per pulse */
1124
2/2
✓ Branch 0 taken 646 times.
✓ Branch 1 taken 36 times.
 682 if (s->aw_n_pulses[0] > 0) {
1125
2/2
✓ Branch 0 taken 323 times.
✓ Branch 1 taken 323 times.
 646 if (block_idx == 0) {
1126 323 range = 32;
1127 } else /* block_idx = 1 */ {
1128 323 range = 8;
1129
2/2
✓ Branch 0 taken 316 times.
✓ Branch 1 taken 7 times.
 323 if (s->aw_n_pulses[block_idx] > 0)
1130 316 pulse_off = s->aw_next_pulse_off_cache;
1131 }
1132 } else
1133 36 range = 16;
1134
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;
1135
1136 /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,
1137 * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus
1138 * we exclude that range from being pulsed again in this function. */
1139 682 memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0]));
1140 682 memset( use_mask, -1, 5 * sizeof(use_mask[0]));
1141 682 memset(&use_mask[5], 0, 2 * sizeof(use_mask[0]));
1142
2/2
✓ Branch 0 taken 657 times.
✓ Branch 1 taken 25 times.
 682 if (s->aw_n_pulses[block_idx] > 0)
1143
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) {
1144 911 int excl_range = s->aw_pulse_range; // always 16 or 24
1145 911 uint16_t *use_mask_ptr = &use_mask[idx >> 4];
1146 911 int first_sh = 16 - (idx & 15);
1147 911 *use_mask_ptr++ &= 0xFFFFu << first_sh;
1148 911 excl_range -= first_sh;
1149
2/2
✓ Branch 0 taken 468 times.
✓ Branch 1 taken 443 times.
 911 if (excl_range >= 16) {
1150 468 *use_mask_ptr++ = 0;
1151 468 *use_mask_ptr &= 0xFFFF >> (excl_range - 16);
1152 } else
1153 443 *use_mask_ptr &= 0xFFFF >> excl_range;
1154 }
1155
1156 /* find the 'aidx'th offset that is not excluded */
1157
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);
1158
2/2
✓ Branch 0 taken 16143 times.
✓ Branch 1 taken 682 times.
 16825 for (n = 0; n <= aidx; pulse_start++) {
1159
2/2
✓ Branch 0 taken 2315 times.
✓ Branch 1 taken 16143 times.
 18458 for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ;
1160
2/2
✓ Branch 0 taken 538 times.
✓ Branch 1 taken 15605 times.
 16143 if (idx >= MAX_FRAMESIZE / 2) { // find from zero
1161
2/2
✓ Branch 0 taken 415 times.
✓ Branch 1 taken 123 times.
 538 if (use_mask[0]) idx = 0x0F;
1162
2/2
✓ Branch 0 taken 105 times.
✓ Branch 1 taken 18 times.
 123 else if (use_mask[1]) idx = 0x1F;
1163
1/2
✓ Branch 0 taken 18 times.
✗ Branch 1 not taken.
 18 else if (use_mask[2]) idx = 0x2F;
1164 ✗ else if (use_mask[3]) idx = 0x3F;
1165 ✗ else if (use_mask[4]) idx = 0x4F;
1166 ✗ else return -1;
1167 538 idx -= av_log2_16bit(use_mask[idx >> 4]);
1168 }
1169
2/2
✓ Branch 0 taken 7465 times.
✓ Branch 1 taken 8678 times.
 16143 if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) {
1170 7465 use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15));
1171 7465 n++;
1172 7465 start_off = idx;
1173 }
1174 }
1175
1176 682 fcb->x[fcb->n] = start_off;
1177
2/2
✓ Branch 1 taken 365 times.
✓ Branch 2 taken 317 times.
 682 fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0;
1178 682 fcb->n++;
1179
1180 /* set offset for next block, relative to start of that block */
1181 682 n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag;
1182
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;
1183 682 return 0;
1184 }
1185
1186 /**
1187 * Apply first set of pitch-adaptive window pulses.
1188 * @param s WMA Voice decoding context private data
1189 * @param gb bit I/O context
1190 * @param block_idx block index in frame [0, 1]
1191 * @param fcb storage location for fixed codebook pulse info
1192 */
1193 682 static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb,
1194 int block_idx, AMRFixed *fcb)
1195 {
1196
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));
1197 float v;
1198
1199
2/2
✓ Branch 0 taken 657 times.
✓ Branch 1 taken 25 times.
 682 if (s->aw_n_pulses[block_idx] > 0) {
1200 int n, v_mask, i_mask, sh, n_pulses;
1201
1202
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
1203 652 n_pulses = 3;
1204 652 v_mask = 8;
1205 652 i_mask = 7;
1206 652 sh = 4;
1207 } else { // 4 pulses, 1:sign + 2:index each
1208 5 n_pulses = 4;
1209 5 v_mask = 4;
1210 5 i_mask = 3;
1211 5 sh = 3;
1212 }
1213
1214
2/2
✓ Branch 0 taken 1976 times.
✓ Branch 1 taken 657 times.
 2633 for (n = n_pulses - 1; n >= 0; n--, val >>= sh) {
1215
2/2
✓ Branch 0 taken 950 times.
✓ Branch 1 taken 1026 times.
 1976 fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0;
1216 1976 fcb->x[fcb->n] = (val & i_mask) * n_pulses + n +
1217 1976 s->aw_first_pulse_off[block_idx];
1218
2/2
✓ Branch 0 taken 241 times.
✓ Branch 1 taken 1976 times.
 2217 while (fcb->x[fcb->n] < 0)
1219 241 fcb->x[fcb->n] += fcb->pitch_lag;
1220
2/2
✓ Branch 0 taken 1959 times.
✓ Branch 1 taken 17 times.
 1976 if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2)
1221 1959 fcb->n++;
1222 }
1223 } else {
1224 25 int num2 = (val & 0x1FF) >> 1, delta, idx;
1225
1226
2/2
✓ Branch 0 taken 4 times.
✓ Branch 1 taken 21 times.
 25 if (num2 < 1 * 79) { delta = 1; idx = num2 + 1; }
1227
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; }
1228
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; }
1229 5 else { delta = 7; idx = num2 + 1 - 3 * 75; }
1230
2/2
✓ Branch 0 taken 11 times.
✓ Branch 1 taken 14 times.
 25 v = (val & 0x200) ? -1.0 : 1.0;
1231
1232 25 fcb->no_repeat_mask |= 3 << fcb->n;
1233 25 fcb->x[fcb->n] = idx - delta;
1234 25 fcb->y[fcb->n] = v;
1235 25 fcb->x[fcb->n + 1] = idx;
1236
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 18 times.
 25 fcb->y[fcb->n + 1] = (val & 1) ? -v : v;
1237 25 fcb->n += 2;
1238 }
1239 682 }
1240
1241 /**
1242 * @}
1243 *
1244 * Generate a random number from frame_cntr and block_idx, which will live
1245 * in the range [0, 1000 - block_size] (so it can be used as an index in a
1246 * table of size 1000 of which you want to read block_size entries).
1247 *
1248 * @param frame_cntr current frame number
1249 * @param block_num current block index
1250 * @param block_size amount of entries we want to read from a table
1251 * that has 1000 entries
1252 * @return a (non-)random number in the [0, 1000 - block_size] range.
1253 */
1254 499 static int pRNG(int frame_cntr, int block_num, int block_size)
1255 {
1256 /* array to simplify the calculation of z:
1257 * y = (x % 9) * 5 + 6;
1258 * z = (49995 * x) / y;
1259 * Since y only has 9 values, we can remove the division by using a
1260 * LUT and using FASTDIV-style divisions. For each of the 9 values
1261 * of y, we can rewrite z as:
1262 * z = x * (49995 / y) + x * ((49995 % y) / y)
1263 * In this table, each col represents one possible value of y, the
1264 * first number is 49995 / y, and the second is the FASTDIV variant
1265 * of 49995 % y / y. */
1266 static const unsigned int div_tbl[9][2] = {
1267 { 8332, 3 * 715827883U }, // y = 6
1268 { 4545, 0 * 390451573U }, // y = 11
1269 { 3124, 11 * 268435456U }, // y = 16
1270 { 2380, 15 * 204522253U }, // y = 21
1271 { 1922, 23 * 165191050U }, // y = 26
1272 { 1612, 23 * 138547333U }, // y = 31
1273 { 1388, 27 * 119304648U }, // y = 36
1274 { 1219, 16 * 104755300U }, // y = 41
1275 { 1086, 39 * 93368855U } // y = 46
1276 };
1277 499 unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr;
1278
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,
1279 // so this is effectively a modulo (%)
1280 499 y = x - 9 * MULH(477218589, x); // x % 9
1281 499 z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1]));
1282 // z = x * 49995 / (y * 5 + 6)
1283 499 return z % (1000 - block_size);
1284 }
1285
1286 /**
1287 * Parse hardcoded signal for a single block.
1288 * @note see #synth_block().
1289 */
1290 1043 static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb,
1291 int block_idx, int size,
1292 const struct frame_type_desc *frame_desc,
1293 float *excitation)
1294 {
1295 float gain;
1296 int n, r_idx;
1297
1298
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 1043 times.
 1043 av_assert0(size <= MAX_FRAMESIZE);
1299
1300 /* Set the offset from which we start reading wmavoice_std_codebook */
1301
2/2
✓ Branch 0 taken 499 times.
✓ Branch 1 taken 544 times.
 1043 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1302 499 r_idx = pRNG(s->frame_cntr, block_idx, size);
1303 499 gain = s->silence_gain;
1304 } else /* FCB_TYPE_HARDCODED */ {
1305 544 r_idx = get_bits(gb, 8);
1306 544 gain = wmavoice_gain_universal[get_bits(gb, 6)];
1307 }
1308
1309 /* Clear gain prediction parameters */
1310 1043 memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err));
1311
1312 /* Apply gain to hardcoded codebook and use that as excitation signal */
1313
2/2
✓ Branch 0 taken 123360 times.
✓ Branch 1 taken 1043 times.
 124403 for (n = 0; n < size; n++)
1314 123360 excitation[n] = wmavoice_std_codebook[r_idx + n] * gain;
1315 1043 }
1316
1317 /**
1318 * Parse FCB/ACB signal for a single block.
1319 * @note see #synth_block().
1320 */
1321 9740 static void synth_block_fcb_acb(WMAVoiceContext *s, GetBitContext *gb,
1322 int block_idx, int size,
1323 int block_pitch_sh2,
1324 const struct frame_type_desc *frame_desc,
1325 float *excitation)
1326 {
1327 static const float gain_coeff[6] = {
1328 0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458
1329 };
1330 float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain;
1331 int n, idx, gain_weight;
1332 AMRFixed fcb;
1333
1334
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 9740 times.
 9740 av_assert0(size <= MAX_FRAMESIZE / 2);
1335 9740 memset(pulses, 0, sizeof(*pulses) * size);
1336
1337 9740 fcb.pitch_lag = block_pitch_sh2 >> 2;
1338 9740 fcb.pitch_fac = 1.0;
1339 9740 fcb.no_repeat_mask = 0;
1340 9740 fcb.n = 0;
1341
1342 /* For the other frame types, this is where we apply the innovation
1343 * (fixed) codebook pulses of the speech signal. */
1344
2/2
✓ Branch 0 taken 682 times.
✓ Branch 1 taken 9058 times.
 9740 if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1345 682 aw_pulse_set1(s, gb, block_idx, &fcb);
1346
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 682 times.
 682 if (aw_pulse_set2(s, gb, block_idx, &fcb)) {
1347 /* Conceal the block with silence and return.
1348 * Skip the correct amount of bits to read the next
1349 * block from the correct offset. */
1350 ✗ int r_idx = pRNG(s->frame_cntr, block_idx, size);
1351
1352 ✗ for (n = 0; n < size; n++)
1353 ✗ excitation[n] =
1354 ✗ wmavoice_std_codebook[r_idx + n] * s->silence_gain;
1355 ✗ skip_bits(gb, 7 + 1);
1356 ✗ return;
1357 }
1358 } else /* FCB_TYPE_EXC_PULSES */ {
1359 9058 int offset_nbits = 5 - frame_desc->log_n_blocks;
1360
1361 9058 fcb.no_repeat_mask = -1;
1362 /* similar to ff_decode_10_pulses_35bits(), but with single pulses
1363 * (instead of double) for a subset of pulses */
1364
2/2
✓ Branch 0 taken 45290 times.
✓ Branch 1 taken 9058 times.
 54348 for (n = 0; n < 5; n++) {
1365 float sign;
1366 int pos1, pos2;
1367
1368
2/2
✓ Branch 1 taken 22708 times.
✓ Branch 2 taken 22582 times.
 45290 sign = get_bits1(gb) ? 1.0 : -1.0;
1369 45290 pos1 = get_bits(gb, offset_nbits);
1370 45290 fcb.x[fcb.n] = n + 5 * pos1;
1371 45290 fcb.y[fcb.n++] = sign;
1372
2/2
✓ Branch 0 taken 36270 times.
✓ Branch 1 taken 9020 times.
 45290 if (n < frame_desc->dbl_pulses) {
1373 36270 pos2 = get_bits(gb, offset_nbits);
1374 36270 fcb.x[fcb.n] = n + 5 * pos2;
1375
2/2
✓ Branch 0 taken 21490 times.
✓ Branch 1 taken 14780 times.
 36270 fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign;
1376 }
1377 }
1378 }
1379 9740 ff_set_fixed_vector(pulses, &fcb, 1.0, size);
1380
1381 /* Calculate gain for adaptive & fixed codebook signal.
1382 * see ff_amr_set_fixed_gain(). */
1383 9740 idx = get_bits(gb, 7);
1384 9740 fcb_gain = expf(avpriv_scalarproduct_float_c(s->gain_pred_err,
1385 9740 gain_coeff, 6) -
1386 9740 5.2409161640 + wmavoice_gain_codebook_fcb[idx]);
1387 9740 acb_gain = wmavoice_gain_codebook_acb[idx];
1388 9740 pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx],
1389 -2.9957322736 /* log(0.05) */,
1390 1.6094379124 /* log(5.0) */);
1391
1392 9740 gain_weight = 8 >> frame_desc->log_n_blocks;
1393 9740 memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err,
1394 9740 sizeof(*s->gain_pred_err) * (6 - gain_weight));
1395
2/2
✓ Branch 0 taken 20280 times.
✓ Branch 1 taken 9740 times.
 30020 for (n = 0; n < gain_weight; n++)
1396 20280 s->gain_pred_err[n] = pred_err;
1397
1398 /* Calculation of adaptive codebook */
1399
2/2
✓ Branch 0 taken 1276 times.
✓ Branch 1 taken 8464 times.
 9740 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1400 int len;
1401
2/2
✓ Branch 0 taken 17876 times.
✓ Branch 1 taken 1276 times.
 19152 for (n = 0; n < size; n += len) {
1402 int next_idx_sh16;
1403 17876 int abs_idx = block_idx * size + n;
1404 17876 int pitch_sh16 = (s->last_pitch_val << 16) +
1405 17876 s->pitch_diff_sh16 * abs_idx;
1406 17876 int pitch = (pitch_sh16 + 0x6FFF) >> 16;
1407 17876 int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000;
1408 17876 idx = idx_sh16 >> 16;
1409
2/2
✓ Branch 0 taken 17442 times.
✓ Branch 1 taken 434 times.
 17876 if (s->pitch_diff_sh16) {
1410
2/2
✓ Branch 0 taken 10526 times.
✓ Branch 1 taken 6916 times.
 17442 if (s->pitch_diff_sh16 > 0) {
1411 10526 next_idx_sh16 = (idx_sh16) &~ 0xFFFF;
1412 } else
1413 6916 next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF;
1414 17442 len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8,
1415 1, size - n);
1416 } else
1417 434 len = size;
1418
1419 17876 ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch],
1420 wmavoice_ipol1_coeffs, 17,
1421 idx, 9, len);
1422 }
1423 } else /* ACB_TYPE_HAMMING */ {
1424 8464 int block_pitch = block_pitch_sh2 >> 2;
1425 8464 idx = block_pitch_sh2 & 3;
1426
2/2
✓ Branch 0 taken 3652 times.
✓ Branch 1 taken 4812 times.
 8464 if (idx) {
1427 3652 ff_acelp_interpolatef(excitation, &excitation[-block_pitch],
1428 wmavoice_ipol2_coeffs, 4,
1429 idx, 8, size);
1430 } else
1431 4812 av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch,
1432 sizeof(float) * size);
1433 }
1434
1435 /* Interpolate ACB/FCB and use as excitation signal */
1436 9740 ff_weighted_vector_sumf(excitation, excitation, pulses,
1437 acb_gain, fcb_gain, size);
1438 }
1439
1440 /**
1441 * Parse data in a single block.
1442 *
1443 * @param s WMA Voice decoding context private data
1444 * @param gb bit I/O context
1445 * @param block_idx index of the to-be-read block
1446 * @param size amount of samples to be read in this block
1447 * @param block_pitch_sh2 pitch for this block << 2
1448 * @param lsps LSPs for (the end of) this frame
1449 * @param prev_lsps LSPs for the last frame
1450 * @param frame_desc frame type descriptor
1451 * @param excitation target memory for the ACB+FCB interpolated signal
1452 * @param synth target memory for the speech synthesis filter output
1453 * @return 0 on success, <0 on error.
1454 */
1455 10783 static void synth_block(WMAVoiceContext *s, GetBitContext *gb,
1456 int block_idx, int size,
1457 int block_pitch_sh2,
1458 const double *lsps, const double *prev_lsps,
1459 const struct frame_type_desc *frame_desc,
1460 float *excitation, float *synth)
1461 {
1462 double i_lsps[MAX_LSPS];
1463 float lpcs[MAX_LSPS];
1464 float fac;
1465 int n;
1466
1467
2/2
✓ Branch 0 taken 1043 times.
✓ Branch 1 taken 9740 times.
 10783 if (frame_desc->acb_type == ACB_TYPE_NONE)
1468 1043 synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation);
1469 else
1470 9740 synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2,
1471 frame_desc, excitation);
1472
1473 /* convert interpolated LSPs to LPCs */
1474 10783 fac = (block_idx + 0.5) / frame_desc->n_blocks;
1475
2/2
✓ Branch 0 taken 140776 times.
✓ Branch 1 taken 10783 times.
 151559 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1476 140776 i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n]));
1477 10783 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1478
1479 /* Speech synthesis */
1480 10783 ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps);
1481 10783 }
1482
1483 /**
1484 * Synthesize output samples for a single frame.
1485 *
1486 * @param ctx WMA Voice decoder context
1487 * @param gb bit I/O context (s->gb or one for cross-packet superframes)
1488 * @param frame_idx Frame number within superframe [0-2]
1489 * @param samples pointer to output sample buffer, has space for at least 160
1490 * samples
1491 * @param lsps LSP array
1492 * @param prev_lsps array of previous frame's LSPs
1493 * @param excitation target buffer for excitation signal
1494 * @param synth target buffer for synthesized speech data
1495 * @return 0 on success, <0 on error.
1496 */
1497 3306 static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx,
1498 float *samples,
1499 const double *lsps, const double *prev_lsps,
1500 float *excitation, float *synth)
1501 {
1502 3306 WMAVoiceContext *s = ctx->priv_data;
1503 3306 int n, n_blocks_x2, log_n_blocks_x2, av_uninit(cur_pitch_val);
1504 3306 int pitch[MAX_BLOCKS], av_uninit(last_block_pitch);
1505
1506 /* Parse frame type ("frame header"), see frame_descs */
1507 3306 int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc, 6, 3)], block_nsamples;
1508
1509 3306 pitch[0] = INT_MAX;
1510
1511
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3306 times.
 3306 if (bd_idx < 0) {
1512 ✗ av_log(ctx, AV_LOG_ERROR,
1513 "Invalid frame type VLC code, skipping\n");
1514 ✗ return AVERROR_INVALIDDATA;
1515 }
1516
1517 3306 block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks;
1518
1519 /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */
1520
2/2
✓ Branch 0 taken 560 times.
✓ Branch 1 taken 2746 times.
 3306 if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) {
1521 /* Pitch is provided per frame, which is interpreted as the pitch of
1522 * the last sample of the last block of this frame. We can interpolate
1523 * the pitch of other blocks (and even pitch-per-sample) by gradually
1524 * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */
1525 560 n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1;
1526 560 log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1;
1527 560 cur_pitch_val = s->min_pitch_val + get_bits(gb, s->pitch_nbits);
1528
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 560 times.
 560 cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1);
1529
2/2
✓ Branch 0 taken 524 times.
✓ Branch 1 taken 36 times.
 560 if (s->last_acb_type == ACB_TYPE_NONE ||
1530 524 20 * abs(cur_pitch_val - s->last_pitch_val) >
1531
2/2
✓ Branch 0 taken 102 times.
✓ Branch 1 taken 422 times.
 524 (cur_pitch_val + s->last_pitch_val))
1532 138 s->last_pitch_val = cur_pitch_val;
1533
1534 /* pitch per block */
1535
2/2
✓ Branch 0 taken 1276 times.
✓ Branch 1 taken 560 times.
 1836 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1536 1276 int fac = n * 2 + 1;
1537
1538 1276 pitch[n] = (MUL16(fac, cur_pitch_val) +
1539 1276 MUL16((n_blocks_x2 - fac), s->last_pitch_val) +
1540 1276 frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2;
1541 }
1542
1543 /* "pitch-diff-per-sample" for calculation of pitch per sample */
1544 560 s->pitch_diff_sh16 =
1545 560 (cur_pitch_val - s->last_pitch_val) * (1 << 16) / MAX_FRAMESIZE;
1546 }
1547
1548 /* Global gain (if silence) and pitch-adaptive window coordinates */
1549
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) {
1550 499 case FCB_TYPE_SILENCE:
1551 499 s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)];
1552 499 break;
1553 341 case FCB_TYPE_AW_PULSES:
1554 341 aw_parse_coords(s, gb, pitch);
1555 341 break;
1556 }
1557
1558
2/2
✓ Branch 0 taken 10783 times.
✓ Branch 1 taken 3306 times.
 14089 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1559 int bl_pitch_sh2;
1560
1561 /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */
1562
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) {
1563 8464 case ACB_TYPE_HAMMING: {
1564 /* Pitch is given per block. Per-block pitches are encoded as an
1565 * absolute value for the first block, and then delta values
1566 * relative to this value) for all subsequent blocks. The scale of
1567 * this pitch value is semi-logarithmic compared to its use in the
1568 * decoder, so we convert it to normal scale also. */
1569 int block_pitch,
1570 8464 t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2,
1571 8464 t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1,
1572 8464 t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1;
1573
1574
2/2
✓ Branch 0 taken 1975 times.
✓ Branch 1 taken 6489 times.
 8464 if (n == 0) {
1575 1975 block_pitch = get_bits(gb, s->block_pitch_nbits);
1576 } else
1577 6489 block_pitch = last_block_pitch - s->block_delta_pitch_hrange +
1578 6489 get_bits(gb, s->block_delta_pitch_nbits);
1579 /* Convert last_ so that any next delta is within _range */
1580 8464 last_block_pitch = av_clip(block_pitch,
1581 s->block_delta_pitch_hrange,
1582 8464 s->block_pitch_range -
1583 8464 s->block_delta_pitch_hrange);
1584
1585 /* Convert semi-log-style scale back to normal scale */
1586
2/2
✓ Branch 0 taken 1491 times.
✓ Branch 1 taken 6973 times.
 8464 if (block_pitch < t1) {
1587 1491 bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch;
1588 } else {
1589 6973 block_pitch -= t1;
1590
2/2
✓ Branch 0 taken 5712 times.
✓ Branch 1 taken 1261 times.
 6973 if (block_pitch < t2) {
1591 5712 bl_pitch_sh2 =
1592 5712 (s->block_conv_table[1] << 2) + (block_pitch << 1);
1593 } else {
1594 1261 block_pitch -= t2;
1595
1/2
✓ Branch 0 taken 1261 times.
✗ Branch 1 not taken.
 1261 if (block_pitch < t3) {
1596 1261 bl_pitch_sh2 =
1597 1261 (s->block_conv_table[2] + block_pitch) << 2;
1598 } else
1599 ✗ bl_pitch_sh2 = s->block_conv_table[3] << 2;
1600 }
1601 }
1602 8464 pitch[n] = bl_pitch_sh2 >> 2;
1603 8464 break;
1604 }
1605
1606 1276 case ACB_TYPE_ASYMMETRIC: {
1607 1276 bl_pitch_sh2 = pitch[n] << 2;
1608 1276 break;
1609 }
1610
1611 1043 default: // ACB_TYPE_NONE has no pitch
1612 1043 bl_pitch_sh2 = 0;
1613 1043 break;
1614 }
1615
1616 10783 synth_block(s, gb, n, block_nsamples, bl_pitch_sh2,
1617 lsps, prev_lsps, &frame_descs[bd_idx],
1618 10783 &excitation[n * block_nsamples],
1619 10783 &synth[n * block_nsamples]);
1620 }
1621
1622 /* Averaging projection filter, if applicable. Else, just copy samples
1623 * from synthesis buffer */
1624
1/2
✓ Branch 0 taken 3306 times.
✗ Branch 1 not taken.
 3306 if (s->do_apf) {
1625 double i_lsps[MAX_LSPS];
1626 float lpcs[MAX_LSPS];
1627
1628
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)
1629 ✗ return AVERROR_INVALIDDATA;
1630
1631
2/2
✓ Branch 0 taken 42960 times.
✓ Branch 1 taken 3306 times.
 46266 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1632 42960 i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n]));
1633 3306 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1634 3306 postfilter(s, synth, samples, 80, lpcs,
1635 3306 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx],
1636 3306 frame_descs[bd_idx].fcb_type, pitch[0]);
1637
1638
2/2
✓ Branch 0 taken 42960 times.
✓ Branch 1 taken 3306 times.
 46266 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1639 42960 i_lsps[n] = cos(lsps[n]);
1640 3306 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1641 3306 postfilter(s, &synth[80], &samples[80], 80, lpcs,
1642 3306 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80],
1643 3306 frame_descs[bd_idx].fcb_type, pitch[0]);
1644 } else
1645 ✗ memcpy(samples, synth, 160 * sizeof(synth[0]));
1646
1647 /* Cache values for next frame */
1648 3306 s->frame_cntr++;
1649
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3306 times.
 3306 if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%)
1650 3306 s->last_acb_type = frame_descs[bd_idx].acb_type;
1651
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) {
1652 771 case ACB_TYPE_NONE:
1653 771 s->last_pitch_val = 0;
1654 771 break;
1655 560 case ACB_TYPE_ASYMMETRIC:
1656 560 s->last_pitch_val = cur_pitch_val;
1657 560 break;
1658 1975 case ACB_TYPE_HAMMING:
1659 1975 s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1];
1660 1975 break;
1661 }
1662
1663 3306 return 0;
1664 }
1665
1666 /**
1667 * Ensure minimum value for first item, maximum value for last value,
1668 * proper spacing between each value and proper ordering.
1669 *
1670 * @param lsps array of LSPs
1671 * @param num size of LSP array
1672 *
1673 * @note basically a double version of #ff_acelp_reorder_lsf(), might be
1674 * useful to put in a generic location later on. Parts are also
1675 * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(),
1676 * which is in float.
1677 */
1678 3306 static void stabilize_lsps(double *lsps, int num)
1679 {
1680 int n, m, l;
1681
1682 /* set minimum value for first, maximum value for last and minimum
1683 * spacing between LSF values.
1684 * Very similar to ff_set_min_dist_lsf(), but in double. */
1685
1/2
✓ Branch 0 taken 3306 times.
✗ Branch 1 not taken.
 3306 lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI);
1686
2/2
✓ Branch 0 taken 39654 times.
✓ Branch 1 taken 3306 times.
 42960 for (n = 1; n < num; n++)
1687
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);
1688
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3306 times.
 3306 lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI);
1689
1690 /* reorder (looks like one-time / non-recursed bubblesort).
1691 * Very similar to ff_sort_nearly_sorted_floats(), but in double. */
1692
2/2
✓ Branch 0 taken 39654 times.
✓ Branch 1 taken 3306 times.
 42960 for (n = 1; n < num; n++) {
1693
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 39654 times.
 39654 if (lsps[n] < lsps[n - 1]) {
1694 ✗ for (m = 1; m < num; m++) {
1695 ✗ double tmp = lsps[m];
1696 ✗ for (l = m - 1; l >= 0; l--) {
1697 ✗ if (lsps[l] <= tmp) break;
1698 ✗ lsps[l + 1] = lsps[l];
1699 }
1700 ✗ lsps[l + 1] = tmp;
1701 }
1702 ✗ break;
1703 }
1704 }
1705 3306 }
1706
1707 /**
1708 * Synthesize output samples for a single superframe. If we have any data
1709 * cached in s->sframe_cache, that will be used instead of whatever is loaded
1710 * in s->gb.
1711 *
1712 * WMA Voice superframes contain 3 frames, each containing 160 audio samples,
1713 * to give a total of 480 samples per frame. See #synth_frame() for frame
1714 * parsing. In addition to 3 frames, superframes can also contain the LSPs
1715 * (if these are globally specified for all frames (residually); they can
1716 * also be specified individually per-frame. See the s->has_residual_lsps
1717 * option), and can specify the number of samples encoded in this superframe
1718 * (if less than 480), usually used to prevent blanks at track boundaries.
1719 *
1720 * @param ctx WMA Voice decoder context
1721 * @return 0 on success, <0 on error or 1 if there was not enough data to
1722 * fully parse the superframe
1723 */
1724 1102 static int synth_superframe(AVCodecContext *ctx, AVFrame *frame,
1725 int *got_frame_ptr)
1726 {
1727 1102 WMAVoiceContext *s = ctx->priv_data;
1728 1102 GetBitContext *gb = &s->gb, s_gb;
1729 1102 int n, res, n_samples = MAX_SFRAMESIZE;
1730 double lsps[MAX_FRAMES][MAX_LSPS];
1731 2204 const double *mean_lsf = s->lsps == 16 ?
1732
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];
1733 float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12];
1734 float synth[MAX_LSPS + MAX_SFRAMESIZE];
1735 float *samples;
1736
1737 1102 memcpy(synth, s->synth_history,
1738 1102 s->lsps * sizeof(*synth));
1739 1102 memcpy(excitation, s->excitation_history,
1740 1102 s->history_nsamples * sizeof(*excitation));
1741
1742
2/2
✓ Branch 0 taken 185 times.
✓ Branch 1 taken 917 times.
 1102 if (s->sframe_cache_size > 0) {
1743 185 gb = &s_gb;
1744 185 init_get_bits(gb, s->sframe_cache, s->sframe_cache_size);
1745 185 s->sframe_cache_size = 0;
1746 }
1747
1748 /* First bit is speech/music bit, it differentiates between WMAVoice
1749 * speech samples (the actual codec) and WMAVoice music samples, which
1750 * are really WMAPro-in-WMAVoice-superframes. I've never seen those in
1751 * the wild yet. */
1752
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 1102 times.
 1102 if (!get_bits1(gb)) {
1753 ✗ avpriv_request_sample(ctx, "WMAPro-in-WMAVoice");
1754 ✗ return AVERROR_PATCHWELCOME;
1755 }
1756
1757 /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */
1758
2/2
✓ Branch 1 taken 3 times.
✓ Branch 2 taken 1099 times.
 1102 if (get_bits1(gb)) {
1759
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 3 times.
 3 if ((n_samples = get_bits(gb, 12)) > MAX_SFRAMESIZE) {
1760 ✗ av_log(ctx, AV_LOG_ERROR,
1761 "Superframe encodes > %d samples (%d), not allowed\n",
1762 MAX_SFRAMESIZE, n_samples);
1763 ✗ return AVERROR_INVALIDDATA;
1764 }
1765 }
1766
1767 /* Parse LSPs, if global for the superframe (can also be per-frame). */
1768
1/2
✓ Branch 0 taken 1102 times.
✗ Branch 1 not taken.
 1102 if (s->has_residual_lsps) {
1769 double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2];
1770
1771
2/2
✓ Branch 0 taken 14320 times.
✓ Branch 1 taken 1102 times.
 15422 for (n = 0; n < s->lsps; n++)
1772 14320 prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n];
1773
1774
2/2
✓ Branch 0 taken 552 times.
✓ Branch 1 taken 550 times.
 1102 if (s->lsps == 10) {
1775 552 dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1776 } else /* s->lsps == 16 */
1777 550 dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1778
1779
2/2
✓ Branch 0 taken 14320 times.
✓ Branch 1 taken 1102 times.
 15422 for (n = 0; n < s->lsps; n++) {
1780 14320 lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]);
1781 14320 lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]);
1782 14320 lsps[2][n] += mean_lsf[n];
1783 }
1784
2/2
✓ Branch 0 taken 3306 times.
✓ Branch 1 taken 1102 times.
 4408 for (n = 0; n < 3; n++)
1785 3306 stabilize_lsps(lsps[n], s->lsps);
1786 }
1787
1788 /* synth_superframe can run multiple times per packet
1789 * free potential previous frame */
1790 1102 av_frame_unref(frame);
1791
1792 /* get output buffer */
1793 1102 frame->nb_samples = MAX_SFRAMESIZE;
1794
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 1102 times.
 1102 if ((res = ff_get_buffer(ctx, frame, 0)) < 0)
1795 ✗ return res;
1796 1102 frame->nb_samples = n_samples;
1797 1102 samples = (float *)frame->data[0];
1798
1799 /* Parse frames, optionally preceded by per-frame (independent) LSPs. */
1800
2/2
✓ Branch 0 taken 3306 times.
✓ Branch 1 taken 1102 times.
 4408 for (n = 0; n < 3; n++) {
1801
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3306 times.
 3306 if (!s->has_residual_lsps) {
1802 int m;
1803
1804 ✗ if (s->lsps == 10) {
1805 ✗ dequant_lsp10i(gb, lsps[n]);
1806 } else /* s->lsps == 16 */
1807 ✗ dequant_lsp16i(gb, lsps[n]);
1808
1809 ✗ for (m = 0; m < s->lsps; m++)
1810 ✗ lsps[n][m] += mean_lsf[m];
1811 ✗ stabilize_lsps(lsps[n], s->lsps);
1812 }
1813
1814
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3306 times.
 4408 if ((res = synth_frame(ctx, gb, n,
1815 3306 &samples[n * MAX_FRAMESIZE],
1816 3306 lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1],
1817 3306 &excitation[s->history_nsamples + n * MAX_FRAMESIZE],
1818
2/2
✓ Branch 0 taken 1102 times.
✓ Branch 1 taken 2204 times.
 3306 &synth[s->lsps + n * MAX_FRAMESIZE]))) {
1819 ✗ *got_frame_ptr = 0;
1820 ✗ return res;
1821 }
1822 }
1823
1824 /* Statistics? FIXME - we don't check for length, a slight overrun
1825 * will be caught by internal buffer padding, and anything else
1826 * will be skipped, not read. */
1827
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 1102 times.
 1102 if (get_bits1(gb)) {
1828 ✗ res = get_bits(gb, 4);
1829 ✗ skip_bits(gb, 10 * (res + 1));
1830 }
1831
1832
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 1102 times.
 1102 if (get_bits_left(gb) < 0) {
1833 ✗ wmavoice_flush(ctx);
1834 ✗ return AVERROR_INVALIDDATA;
1835 }
1836
1837 1102 *got_frame_ptr = 1;
1838
1839 /* Update history */
1840 1102 memcpy(s->prev_lsps, lsps[2],
1841 1102 s->lsps * sizeof(*s->prev_lsps));
1842 1102 memcpy(s->synth_history, &synth[MAX_SFRAMESIZE],
1843 1102 s->lsps * sizeof(*synth));
1844 1102 memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE],
1845 1102 s->history_nsamples * sizeof(*excitation));
1846
1/2
✓ Branch 0 taken 1102 times.
✗ Branch 1 not taken.
 1102 if (s->do_apf)
1847 1102 memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE],
1848 1102 s->history_nsamples * sizeof(*s->zero_exc_pf));
1849
1850 1102 return 0;
1851 }
1852
1853 /**
1854 * Parse the packet header at the start of each packet (input data to this
1855 * decoder).
1856 *
1857 * @param s WMA Voice decoding context private data
1858 * @return <0 on error, nb_superframes on success.
1859 */
1860 186 static int parse_packet_header(WMAVoiceContext *s)
1861 {
1862 186 GetBitContext *gb = &s->gb;
1863 186 unsigned int res, n_superframes = 0;
1864
1865 186 skip_bits(gb, 4); // packet sequence number
1866 186 s->has_residual_lsps = get_bits1(gb);
1867 do {
1868
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 186 times.
 186 if (get_bits_left(gb) < 6 + s->spillover_bitsize)
1869 ✗ return AVERROR_INVALIDDATA;
1870
1871 186 res = get_bits(gb, 6); // number of superframes per packet
1872 // (minus first one if there is spillover)
1873 186 n_superframes += res;
1874
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 186 times.
 186 } while (res == 0x3F);
1875 186 s->spillover_nbits = get_bits(gb, s->spillover_bitsize);
1876
1877
1/2
✓ Branch 1 taken 186 times.
✗ Branch 2 not taken.
 186 return get_bits_left(gb) >= 0 ? n_superframes : AVERROR_INVALIDDATA;
1878 }
1879
1880 /**
1881 * Copy (unaligned) bits from gb/data/size to pb.
1882 *
1883 * @param pb target buffer to copy bits into
1884 * @param data source buffer to copy bits from
1885 * @param size size of the source data, in bytes
1886 * @param gb bit I/O context specifying the current position in the source.
1887 * data. This function might use this to align the bit position to
1888 * a whole-byte boundary before calling #ff_copy_bits() on aligned
1889 * source data
1890 * @param nbits the amount of bits to copy from source to target
1891 *
1892 * @note after calling this function, the current position in the input bit
1893 * I/O context is undefined.
1894 */
1895 370 static void copy_bits(PutBitContext *pb,
1896 const uint8_t *data, int size,
1897 GetBitContext *gb, int nbits)
1898 {
1899 int rmn_bytes, rmn_bits;
1900
1901 370 rmn_bits = rmn_bytes = get_bits_left(gb);
1902
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 370 times.
 370 if (rmn_bits < nbits)
1903 ✗ return;
1904
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 370 times.
 370 if (nbits > put_bits_left(pb))
1905 ✗ return;
1906 370 rmn_bits &= 7; rmn_bytes >>= 3;
1907
2/2
✓ Branch 0 taken 290 times.
✓ Branch 1 taken 80 times.
 370 if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0)
1908 290 put_bits(pb, rmn_bits, get_bits(gb, rmn_bits));
1909 370 ff_copy_bits(pb, data + size - rmn_bytes,
1910 370 FFMIN(nbits - rmn_bits, rmn_bytes << 3));
1911 }
1912
1913 /**
1914 * Packet decoding: a packet is anything that the (ASF) demuxer contains,
1915 * and we expect that the demuxer / application provides it to us as such
1916 * (else you'll probably get garbage as output). Every packet has a size of
1917 * ctx->block_align bytes, starts with a packet header (see
1918 * #parse_packet_header()), and then a series of superframes. Superframe
1919 * boundaries may exceed packets, i.e. superframes can split data over
1920 * multiple (two) packets.
1921 *
1922 * For more information about frames, see #synth_superframe().
1923 */
1924 1291 static int wmavoice_decode_packet(AVCodecContext *ctx, AVFrame *frame,
1925 int *got_frame_ptr, AVPacket *avpkt)
1926 {
1927 1291 WMAVoiceContext *s = ctx->priv_data;
1928 1291 GetBitContext *gb = &s->gb;
1929 1291 const uint8_t *buf = avpkt->data;
1930 uint8_t dummy[1];
1931 int size, res, pos;
1932
1933 /* Packets are sometimes a multiple of ctx->block_align, with a packet
1934 * header at each ctx->block_align bytes. However, FFmpeg's ASF demuxer
1935 * feeds us ASF packets, which may concatenate multiple "codec" packets
1936 * in a single "muxer" packet, so we artificially emulate that by
1937 * capping the packet size at ctx->block_align. */
1938
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);
1939
2/2
✓ Branch 0 taken 5 times.
✓ Branch 1 taken 1286 times.
 1291 buf = size ? buf : dummy;
1940 1291 res = init_get_bits8(&s->gb, buf, size);
1941
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 1291 times.
 1291 if (res < 0)
1942 ✗ return res;
1943
1944 /* size == ctx->block_align is used to indicate whether we are dealing with
1945 * a new packet or a packet of which we already read the packet header
1946 * previously. */
1947
2/2
✓ Branch 0 taken 191 times.
✓ Branch 1 taken 1100 times.
 1291 if (!(size % ctx->block_align)) { // new packet header
1948
2/2
✓ Branch 0 taken 5 times.
✓ Branch 1 taken 186 times.
 191 if (!size) {
1949 5 s->spillover_nbits = 0;
1950 5 s->nb_superframes = 0;
1951 } else {
1952
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 186 times.
 186 if ((res = parse_packet_header(s)) < 0)
1953 ✗ return res;
1954 186 s->nb_superframes = res;
1955 }
1956
1957 /* If the packet header specifies a s->spillover_nbits, then we want
1958 * to push out all data of the previous packet (+ spillover) before
1959 * continuing to parse new superframes in the current packet. */
1960
2/2
✓ Branch 0 taken 185 times.
✓ Branch 1 taken 6 times.
 191 if (s->sframe_cache_size > 0) {
1961 185 int cnt = get_bits_count(gb);
1962
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 185 times.
 185 if (cnt + s->spillover_nbits > avpkt->size * 8) {
1963 ✗ s->spillover_nbits = avpkt->size * 8 - cnt;
1964 }
1965 185 copy_bits(&s->pb, buf, size, gb, s->spillover_nbits);
1966 185 flush_put_bits(&s->pb);
1967 185 s->sframe_cache_size += s->spillover_nbits;
1968
1/2
✓ Branch 1 taken 185 times.
✗ Branch 2 not taken.
 185 if ((res = synth_superframe(ctx, frame, got_frame_ptr)) == 0 &&
1969
1/2
✓ Branch 0 taken 185 times.
✗ Branch 1 not taken.
 185 *got_frame_ptr) {
1970 185 cnt += s->spillover_nbits;
1971 185 s->skip_bits_next = cnt & 7;
1972 185 res = cnt >> 3;
1973 185 return res;
1974 } else
1975 ✗ skip_bits_long (gb, s->spillover_nbits - cnt +
1976 ✗ get_bits_count(gb)); // resync
1977
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 6 times.
 6 } else if (s->spillover_nbits) {
1978 ✗ skip_bits_long(gb, s->spillover_nbits); // resync
1979 }
1980
2/2
✓ Branch 0 taken 971 times.
✓ Branch 1 taken 129 times.
 1100 } else if (s->skip_bits_next)
1981 971 skip_bits(gb, s->skip_bits_next);
1982
1983 /* Try parsing superframes in current packet */
1984 1106 s->sframe_cache_size = 0;
1985 1106 s->skip_bits_next = 0;
1986 1106 pos = get_bits_left(gb);
1987
2/2
✓ Branch 0 taken 4 times.
✓ Branch 1 taken 1102 times.
 1106 if (s->nb_superframes-- == 0) {
1988 4 *got_frame_ptr = 0;
1989 4 return size;
1990
2/2
✓ Branch 0 taken 917 times.
✓ Branch 1 taken 185 times.
 1102 } else if (s->nb_superframes > 0) {
1991
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 917 times.
 917 if ((res = synth_superframe(ctx, frame, got_frame_ptr)) < 0) {
1992 ✗ return res;
1993
1/2
✓ Branch 0 taken 917 times.
✗ Branch 1 not taken.
 917 } else if (*got_frame_ptr) {
1994 917 int cnt = get_bits_count(gb);
1995 917 s->skip_bits_next = cnt & 7;
1996 917 res = cnt >> 3;
1997 917 return res;
1998 }
1999
1/2
✓ Branch 0 taken 185 times.
✗ Branch 1 not taken.
 185 } else if ((s->sframe_cache_size = pos) > 0) {
2000 /* ... cache it for spillover in next packet */
2001 185 init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE);
2002 185 copy_bits(&s->pb, buf, size, gb, s->sframe_cache_size);
2003 // FIXME bad - just copy bytes as whole and add use the
2004 // skip_bits_next field
2005 }
2006
2007 185 return size;
2008 }
2009
2010 8 static av_cold int wmavoice_decode_end(AVCodecContext *ctx)
2011 {
2012 8 WMAVoiceContext *s = ctx->priv_data;
2013
2014
1/2
✓ Branch 0 taken 8 times.
✗ Branch 1 not taken.
 8 if (s->do_apf) {
2015 8 av_tx_uninit(&s->rdft);
2016 8 av_tx_uninit(&s->irdft);
2017 8 av_tx_uninit(&s->dct);
2018 8 av_tx_uninit(&s->dst);
2019 }
2020
2021 8 return 0;
2022 }
2023
2024 const FFCodec ff_wmavoice_decoder = {
2025 .p.name = "wmavoice",
2026 CODEC_LONG_NAME("Windows Media Audio Voice"),
2027 .p.type = AVMEDIA_TYPE_AUDIO,
2028 .p.id = AV_CODEC_ID_WMAVOICE,
2029 .priv_data_size = sizeof(WMAVoiceContext),
2030 .init = wmavoice_decode_init,
2031 .close = wmavoice_decode_end,
2032 FF_CODEC_DECODE_CB(wmavoice_decode_packet),
2033 .p.capabilities =
2034 #if FF_API_SUBFRAMES
2035 AV_CODEC_CAP_SUBFRAMES |
2036 #endif
2037 AV_CODEC_CAP_DR1 | AV_CODEC_CAP_DELAY,
2038 .caps_internal = FF_CODEC_CAP_INIT_CLEANUP,
2039 .flush = wmavoice_flush,
2040 };
2041