blob: 2ec4499981f52a318ec3ad714a4a8367e5b2beba
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 "avcodec.h" |
34 | #include "internal.h" |
35 | #include "get_bits.h" |
36 | #include "put_bits.h" |
37 | #include "wmavoice_data.h" |
38 | #include "celp_filters.h" |
39 | #include "acelp_vectors.h" |
40 | #include "acelp_filters.h" |
41 | #include "lsp.h" |
42 | #include "dct.h" |
43 | #include "rdft.h" |
44 | #include "sinewin.h" |
45 | |
46 | #define MAX_BLOCKS 8 ///< maximum number of blocks per frame |
47 | #define MAX_LSPS 16 ///< maximum filter order |
48 | #define MAX_LSPS_ALIGN16 16 ///< same as #MAX_LSPS; needs to be multiple |
49 | ///< of 16 for ASM input buffer alignment |
50 | #define MAX_FRAMES 3 ///< maximum number of frames per superframe |
51 | #define MAX_FRAMESIZE 160 ///< maximum number of samples per frame |
52 | #define MAX_SIGNAL_HISTORY 416 ///< maximum excitation signal history |
53 | #define MAX_SFRAMESIZE (MAX_FRAMESIZE * MAX_FRAMES) |
54 | ///< maximum number of samples per superframe |
55 | #define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that |
56 | ///< was split over two packets |
57 | #define VLC_NBITS 6 ///< number of bits to read per VLC iteration |
58 | |
59 | /** |
60 | * Frame type VLC coding. |
61 | */ |
62 | static VLC frame_type_vlc; |
63 | |
64 | /** |
65 | * Adaptive codebook types. |
66 | */ |
67 | enum { |
68 | ACB_TYPE_NONE = 0, ///< no adaptive codebook (only hardcoded fixed) |
69 | ACB_TYPE_ASYMMETRIC = 1, ///< adaptive codebook with per-frame pitch, which |
70 | ///< we interpolate to get a per-sample pitch. |
71 | ///< Signal is generated using an asymmetric sinc |
72 | ///< window function |
73 | ///< @note see #wmavoice_ipol1_coeffs |
74 | ACB_TYPE_HAMMING = 2 ///< Per-block pitch with signal generation using |
75 | ///< a Hamming sinc window function |
76 | ///< @note see #wmavoice_ipol2_coeffs |
77 | }; |
78 | |
79 | /** |
80 | * Fixed codebook types. |
81 | */ |
82 | enum { |
83 | FCB_TYPE_SILENCE = 0, ///< comfort noise during silence |
84 | ///< generated from a hardcoded (fixed) codebook |
85 | ///< with per-frame (low) gain values |
86 | FCB_TYPE_HARDCODED = 1, ///< hardcoded (fixed) codebook with per-block |
87 | ///< gain values |
88 | FCB_TYPE_AW_PULSES = 2, ///< Pitch-adaptive window (AW) pulse signals, |
89 | ///< used in particular for low-bitrate streams |
90 | FCB_TYPE_EXC_PULSES = 3, ///< Innovation (fixed) codebook pulse sets in |
91 | ///< combinations of either single pulses or |
92 | ///< pulse pairs |
93 | }; |
94 | |
95 | /** |
96 | * Description of frame types. |
97 | */ |
98 | static const struct frame_type_desc { |
99 | uint8_t n_blocks; ///< amount of blocks per frame (each block |
100 | ///< (contains 160/#n_blocks samples) |
101 | uint8_t log_n_blocks; ///< log2(#n_blocks) |
102 | uint8_t acb_type; ///< Adaptive codebook type (ACB_TYPE_*) |
103 | uint8_t fcb_type; ///< Fixed codebook type (FCB_TYPE_*) |
104 | uint8_t dbl_pulses; ///< how many pulse vectors have pulse pairs |
105 | ///< (rather than just one single pulse) |
106 | ///< only if #fcb_type == #FCB_TYPE_EXC_PULSES |
107 | } frame_descs[17] = { |
108 | { 1, 0, ACB_TYPE_NONE, FCB_TYPE_SILENCE, 0 }, |
109 | { 2, 1, ACB_TYPE_NONE, FCB_TYPE_HARDCODED, 0 }, |
110 | { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_AW_PULSES, 0 }, |
111 | { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2 }, |
112 | { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5 }, |
113 | { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 0 }, |
114 | { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2 }, |
115 | { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5 }, |
116 | { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0 }, |
117 | { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2 }, |
118 | { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5 }, |
119 | { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0 }, |
120 | { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2 }, |
121 | { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5 }, |
122 | { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0 }, |
123 | { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2 }, |
124 | { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5 } |
125 | }; |
126 | |
127 | /** |
128 | * WMA Voice decoding context. |
129 | */ |
130 | typedef struct WMAVoiceContext { |
131 | /** |
132 | * @name Global values specified in the stream header / extradata or used all over. |
133 | * @{ |
134 | */ |
135 | GetBitContext gb; ///< packet bitreader. During decoder init, |
136 | ///< it contains the extradata from the |
137 | ///< demuxer. During decoding, it contains |
138 | ///< packet data. |
139 | int8_t vbm_tree[25]; ///< converts VLC codes to frame type |
140 | |
141 | int spillover_bitsize; ///< number of bits used to specify |
142 | ///< #spillover_nbits in the packet header |
143 | ///< = ceil(log2(ctx->block_align << 3)) |
144 | int history_nsamples; ///< number of samples in history for signal |
145 | ///< prediction (through ACB) |
146 | |
147 | /* postfilter specific values */ |
148 | int do_apf; ///< whether to apply the averaged |
149 | ///< projection filter (APF) |
150 | int denoise_strength; ///< strength of denoising in Wiener filter |
151 | ///< [0-11] |
152 | int denoise_tilt_corr; ///< Whether to apply tilt correction to the |
153 | ///< Wiener filter coefficients (postfilter) |
154 | int dc_level; ///< Predicted amount of DC noise, based |
155 | ///< on which a DC removal filter is used |
156 | |
157 | int lsps; ///< number of LSPs per frame [10 or 16] |
158 | int lsp_q_mode; ///< defines quantizer defaults [0, 1] |
159 | int lsp_def_mode; ///< defines different sets of LSP defaults |
160 | ///< [0, 1] |
161 | |
162 | int min_pitch_val; ///< base value for pitch parsing code |
163 | int max_pitch_val; ///< max value + 1 for pitch parsing |
164 | int pitch_nbits; ///< number of bits used to specify the |
165 | ///< pitch value in the frame header |
166 | int block_pitch_nbits; ///< number of bits used to specify the |
167 | ///< first block's pitch value |
168 | int block_pitch_range; ///< range of the block pitch |
169 | int block_delta_pitch_nbits; ///< number of bits used to specify the |
170 | ///< delta pitch between this and the last |
171 | ///< block's pitch value, used in all but |
172 | ///< first block |
173 | int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is |
174 | ///< from -this to +this-1) |
175 | uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale |
176 | ///< conversion |
177 | |
178 | /** |
179 | * @} |
180 | * |
181 | * @name Packet values specified in the packet header or related to a packet. |
182 | * |
183 | * A packet is considered to be a single unit of data provided to this |
184 | * decoder by the demuxer. |
185 | * @{ |
186 | */ |
187 | int spillover_nbits; ///< number of bits of the previous packet's |
188 | ///< last superframe preceding this |
189 | ///< packet's first full superframe (useful |
190 | ///< for re-synchronization also) |
191 | int has_residual_lsps; ///< if set, superframes contain one set of |
192 | ///< LSPs that cover all frames, encoded as |
193 | ///< independent and residual LSPs; if not |
194 | ///< set, each frame contains its own, fully |
195 | ///< independent, LSPs |
196 | int skip_bits_next; ///< number of bits to skip at the next call |
197 | ///< to #wmavoice_decode_packet() (since |
198 | ///< they're part of the previous superframe) |
199 | |
200 | uint8_t sframe_cache[SFRAME_CACHE_MAXSIZE + AV_INPUT_BUFFER_PADDING_SIZE]; |
201 | ///< cache for superframe data split over |
202 | ///< multiple packets |
203 | int sframe_cache_size; ///< set to >0 if we have data from an |
204 | ///< (incomplete) superframe from a previous |
205 | ///< packet that spilled over in the current |
206 | ///< packet; specifies the amount of bits in |
207 | ///< #sframe_cache |
208 | PutBitContext pb; ///< bitstream writer for #sframe_cache |
209 | |
210 | /** |
211 | * @} |
212 | * |
213 | * @name Frame and superframe values |
214 | * Superframe and frame data - these can change from frame to frame, |
215 | * although some of them do in that case serve as a cache / history for |
216 | * the next frame or superframe. |
217 | * @{ |
218 | */ |
219 | double prev_lsps[MAX_LSPS]; ///< LSPs of the last frame of the previous |
220 | ///< superframe |
221 | int last_pitch_val; ///< pitch value of the previous frame |
222 | int last_acb_type; ///< frame type [0-2] of the previous frame |
223 | int pitch_diff_sh16; ///< ((cur_pitch_val - #last_pitch_val) |
224 | ///< << 16) / #MAX_FRAMESIZE |
225 | float silence_gain; ///< set for use in blocks if #ACB_TYPE_NONE |
226 | |
227 | int aw_idx_is_ext; ///< whether the AW index was encoded in |
228 | ///< 8 bits (instead of 6) |
229 | int aw_pulse_range; ///< the range over which #aw_pulse_set1() |
230 | ///< can apply the pulse, relative to the |
231 | ///< value in aw_first_pulse_off. The exact |
232 | ///< position of the first AW-pulse is within |
233 | ///< [pulse_off, pulse_off + this], and |
234 | ///< depends on bitstream values; [16 or 24] |
235 | int aw_n_pulses[2]; ///< number of AW-pulses in each block; note |
236 | ///< that this number can be negative (in |
237 | ///< which case it basically means "zero") |
238 | int aw_first_pulse_off[2]; ///< index of first sample to which to |
239 | ///< apply AW-pulses, or -0xff if unset |
240 | int aw_next_pulse_off_cache; ///< the position (relative to start of the |
241 | ///< second block) at which pulses should |
242 | ///< start to be positioned, serves as a |
243 | ///< cache for pitch-adaptive window pulses |
244 | ///< between blocks |
245 | |
246 | int frame_cntr; ///< current frame index [0 - 0xFFFE]; is |
247 | ///< only used for comfort noise in #pRNG() |
248 | int nb_superframes; ///< number of superframes in current packet |
249 | float gain_pred_err[6]; ///< cache for gain prediction |
250 | float excitation_history[MAX_SIGNAL_HISTORY]; |
251 | ///< cache of the signal of previous |
252 | ///< superframes, used as a history for |
253 | ///< signal generation |
254 | float synth_history[MAX_LSPS]; ///< see #excitation_history |
255 | /** |
256 | * @} |
257 | * |
258 | * @name Postfilter values |
259 | * |
260 | * Variables used for postfilter implementation, mostly history for |
261 | * smoothing and so on, and context variables for FFT/iFFT. |
262 | * @{ |
263 | */ |
264 | RDFTContext rdft, irdft; ///< contexts for FFT-calculation in the |
265 | ///< postfilter (for denoise filter) |
266 | DCTContext dct, dst; ///< contexts for phase shift (in Hilbert |
267 | ///< transform, part of postfilter) |
268 | float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi] |
269 | ///< range |
270 | float postfilter_agc; ///< gain control memory, used in |
271 | ///< #adaptive_gain_control() |
272 | float dcf_mem[2]; ///< DC filter history |
273 | float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE]; |
274 | ///< zero filter output (i.e. excitation) |
275 | ///< by postfilter |
276 | float denoise_filter_cache[MAX_FRAMESIZE]; |
277 | int denoise_filter_cache_size; ///< samples in #denoise_filter_cache |
278 | DECLARE_ALIGNED(32, float, tilted_lpcs_pf)[0x80]; |
279 | ///< aligned buffer for LPC tilting |
280 | DECLARE_ALIGNED(32, float, denoise_coeffs_pf)[0x80]; |
281 | ///< aligned buffer for denoise coefficients |
282 | DECLARE_ALIGNED(32, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16]; |
283 | ///< aligned buffer for postfilter speech |
284 | ///< synthesis |
285 | /** |
286 | * @} |
287 | */ |
288 | } WMAVoiceContext; |
289 | |
290 | /** |
291 | * Set up the variable bit mode (VBM) tree from container extradata. |
292 | * @param gb bit I/O context. |
293 | * The bit context (s->gb) should be loaded with byte 23-46 of the |
294 | * container extradata (i.e. the ones containing the VBM tree). |
295 | * @param vbm_tree pointer to array to which the decoded VBM tree will be |
296 | * written. |
297 | * @return 0 on success, <0 on error. |
298 | */ |
299 | static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25]) |
300 | { |
301 | int cntr[8] = { 0 }, n, res; |
302 | |
303 | memset(vbm_tree, 0xff, sizeof(vbm_tree[0]) * 25); |
304 | for (n = 0; n < 17; n++) { |
305 | res = get_bits(gb, 3); |
306 | if (cntr[res] > 3) // should be >= 3 + (res == 7)) |
307 | return -1; |
308 | vbm_tree[res * 3 + cntr[res]++] = n; |
309 | } |
310 | return 0; |
311 | } |
312 | |
313 | static av_cold void wmavoice_init_static_data(AVCodec *codec) |
314 | { |
315 | static const uint8_t bits[] = { |
316 | 2, 2, 2, 4, 4, 4, |
317 | 6, 6, 6, 8, 8, 8, |
318 | 10, 10, 10, 12, 12, 12, |
319 | 14, 14, 14, 14 |
320 | }; |
321 | static const uint16_t codes[] = { |
322 | 0x0000, 0x0001, 0x0002, // 00/01/10 |
323 | 0x000c, 0x000d, 0x000e, // 11+00/01/10 |
324 | 0x003c, 0x003d, 0x003e, // 1111+00/01/10 |
325 | 0x00fc, 0x00fd, 0x00fe, // 111111+00/01/10 |
326 | 0x03fc, 0x03fd, 0x03fe, // 11111111+00/01/10 |
327 | 0x0ffc, 0x0ffd, 0x0ffe, // 1111111111+00/01/10 |
328 | 0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx |
329 | }; |
330 | |
331 | INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits), |
332 | bits, 1, 1, codes, 2, 2, 132); |
333 | } |
334 | |
335 | static av_cold void wmavoice_flush(AVCodecContext *ctx) |
336 | { |
337 | WMAVoiceContext *s = ctx->priv_data; |
338 | int n; |
339 | |
340 | s->postfilter_agc = 0; |
341 | s->sframe_cache_size = 0; |
342 | s->skip_bits_next = 0; |
343 | for (n = 0; n < s->lsps; n++) |
344 | s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0); |
345 | memset(s->excitation_history, 0, |
346 | sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY); |
347 | memset(s->synth_history, 0, |
348 | sizeof(*s->synth_history) * MAX_LSPS); |
349 | memset(s->gain_pred_err, 0, |
350 | sizeof(s->gain_pred_err)); |
351 | |
352 | if (s->do_apf) { |
353 | memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0, |
354 | sizeof(*s->synth_filter_out_buf) * s->lsps); |
355 | memset(s->dcf_mem, 0, |
356 | sizeof(*s->dcf_mem) * 2); |
357 | memset(s->zero_exc_pf, 0, |
358 | sizeof(*s->zero_exc_pf) * s->history_nsamples); |
359 | memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache)); |
360 | } |
361 | } |
362 | |
363 | /** |
364 | * Set up decoder with parameters from demuxer (extradata etc.). |
365 | */ |
366 | static av_cold int wmavoice_decode_init(AVCodecContext *ctx) |
367 | { |
368 | int n, flags, pitch_range, lsp16_flag; |
369 | WMAVoiceContext *s = ctx->priv_data; |
370 | |
371 | /** |
372 | * Extradata layout: |
373 | * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c), |
374 | * - byte 19-22: flags field (annoyingly in LE; see below for known |
375 | * values), |
376 | * - byte 23-46: variable bitmode tree (really just 17 * 3 bits, |
377 | * rest is 0). |
378 | */ |
379 | if (ctx->extradata_size != 46) { |
380 | av_log(ctx, AV_LOG_ERROR, |
381 | "Invalid extradata size %d (should be 46)\n", |
382 | ctx->extradata_size); |
383 | return AVERROR_INVALIDDATA; |
384 | } |
385 | if (ctx->block_align <= 0) { |
386 | av_log(ctx, AV_LOG_ERROR, "Invalid block alignment %d.\n", ctx->block_align); |
387 | return AVERROR_INVALIDDATA; |
388 | } |
389 | |
390 | flags = AV_RL32(ctx->extradata + 18); |
391 | s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align); |
392 | s->do_apf = flags & 0x1; |
393 | if (s->do_apf) { |
394 | ff_rdft_init(&s->rdft, 7, DFT_R2C); |
395 | ff_rdft_init(&s->irdft, 7, IDFT_C2R); |
396 | ff_dct_init(&s->dct, 6, DCT_I); |
397 | ff_dct_init(&s->dst, 6, DST_I); |
398 | |
399 | ff_sine_window_init(s->cos, 256); |
400 | memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0])); |
401 | for (n = 0; n < 255; n++) { |
402 | s->sin[n] = -s->sin[510 - n]; |
403 | s->cos[510 - n] = s->cos[n]; |
404 | } |
405 | } |
406 | s->denoise_strength = (flags >> 2) & 0xF; |
407 | if (s->denoise_strength >= 12) { |
408 | av_log(ctx, AV_LOG_ERROR, |
409 | "Invalid denoise filter strength %d (max=11)\n", |
410 | s->denoise_strength); |
411 | return AVERROR_INVALIDDATA; |
412 | } |
413 | s->denoise_tilt_corr = !!(flags & 0x40); |
414 | s->dc_level = (flags >> 7) & 0xF; |
415 | s->lsp_q_mode = !!(flags & 0x2000); |
416 | s->lsp_def_mode = !!(flags & 0x4000); |
417 | lsp16_flag = flags & 0x1000; |
418 | if (lsp16_flag) { |
419 | s->lsps = 16; |
420 | } else { |
421 | s->lsps = 10; |
422 | } |
423 | for (n = 0; n < s->lsps; n++) |
424 | s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0); |
425 | |
426 | init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3); |
427 | if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) { |
428 | av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n"); |
429 | return AVERROR_INVALIDDATA; |
430 | } |
431 | |
432 | s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8; |
433 | s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8; |
434 | pitch_range = s->max_pitch_val - s->min_pitch_val; |
435 | if (pitch_range <= 0) { |
436 | av_log(ctx, AV_LOG_ERROR, "Invalid pitch range; broken extradata?\n"); |
437 | return AVERROR_INVALIDDATA; |
438 | } |
439 | s->pitch_nbits = av_ceil_log2(pitch_range); |
440 | s->last_pitch_val = 40; |
441 | s->last_acb_type = ACB_TYPE_NONE; |
442 | s->history_nsamples = s->max_pitch_val + 8; |
443 | |
444 | if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) { |
445 | int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8, |
446 | max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8; |
447 | |
448 | av_log(ctx, AV_LOG_ERROR, |
449 | "Unsupported samplerate %d (min=%d, max=%d)\n", |
450 | ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz |
451 | |
452 | return AVERROR(ENOSYS); |
453 | } |
454 | |
455 | s->block_conv_table[0] = s->min_pitch_val; |
456 | s->block_conv_table[1] = (pitch_range * 25) >> 6; |
457 | s->block_conv_table[2] = (pitch_range * 44) >> 6; |
458 | s->block_conv_table[3] = s->max_pitch_val - 1; |
459 | s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF; |
460 | if (s->block_delta_pitch_hrange <= 0) { |
461 | av_log(ctx, AV_LOG_ERROR, "Invalid delta pitch hrange; broken extradata?\n"); |
462 | return AVERROR_INVALIDDATA; |
463 | } |
464 | s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange); |
465 | s->block_pitch_range = s->block_conv_table[2] + |
466 | s->block_conv_table[3] + 1 + |
467 | 2 * (s->block_conv_table[1] - 2 * s->min_pitch_val); |
468 | s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range); |
469 | |
470 | ctx->channels = 1; |
471 | ctx->channel_layout = AV_CH_LAYOUT_MONO; |
472 | ctx->sample_fmt = AV_SAMPLE_FMT_FLT; |
473 | |
474 | return 0; |
475 | } |
476 | |
477 | /** |
478 | * @name Postfilter functions |
479 | * Postfilter functions (gain control, wiener denoise filter, DC filter, |
480 | * kalman smoothening, plus surrounding code to wrap it) |
481 | * @{ |
482 | */ |
483 | /** |
484 | * Adaptive gain control (as used in postfilter). |
485 | * |
486 | * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except |
487 | * that the energy here is calculated using sum(abs(...)), whereas the |
488 | * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)). |
489 | * |
490 | * @param out output buffer for filtered samples |
491 | * @param in input buffer containing the samples as they are after the |
492 | * postfilter steps so far |
493 | * @param speech_synth input buffer containing speech synth before postfilter |
494 | * @param size input buffer size |
495 | * @param alpha exponential filter factor |
496 | * @param gain_mem pointer to filter memory (single float) |
497 | */ |
498 | static void adaptive_gain_control(float *out, const float *in, |
499 | const float *speech_synth, |
500 | int size, float alpha, float *gain_mem) |
501 | { |
502 | int i; |
503 | float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor; |
504 | float mem = *gain_mem; |
505 | |
506 | for (i = 0; i < size; i++) { |
507 | speech_energy += fabsf(speech_synth[i]); |
508 | postfilter_energy += fabsf(in[i]); |
509 | } |
510 | gain_scale_factor = postfilter_energy == 0.0 ? 0.0 : |
511 | (1.0 - alpha) * speech_energy / postfilter_energy; |
512 | |
513 | for (i = 0; i < size; i++) { |
514 | mem = alpha * mem + gain_scale_factor; |
515 | out[i] = in[i] * mem; |
516 | } |
517 | |
518 | *gain_mem = mem; |
519 | } |
520 | |
521 | /** |
522 | * Kalman smoothing function. |
523 | * |
524 | * This function looks back pitch +/- 3 samples back into history to find |
525 | * the best fitting curve (that one giving the optimal gain of the two |
526 | * signals, i.e. the highest dot product between the two), and then |
527 | * uses that signal history to smoothen the output of the speech synthesis |
528 | * filter. |
529 | * |
530 | * @param s WMA Voice decoding context |
531 | * @param pitch pitch of the speech signal |
532 | * @param in input speech signal |
533 | * @param out output pointer for smoothened signal |
534 | * @param size input/output buffer size |
535 | * |
536 | * @returns -1 if no smoothening took place, e.g. because no optimal |
537 | * fit could be found, or 0 on success. |
538 | */ |
539 | static int kalman_smoothen(WMAVoiceContext *s, int pitch, |
540 | const float *in, float *out, int size) |
541 | { |
542 | int n; |
543 | float optimal_gain = 0, dot; |
544 | const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)], |
545 | *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)], |
546 | *best_hist_ptr = NULL; |
547 | |
548 | /* find best fitting point in history */ |
549 | do { |
550 | dot = avpriv_scalarproduct_float_c(in, ptr, size); |
551 | if (dot > optimal_gain) { |
552 | optimal_gain = dot; |
553 | best_hist_ptr = ptr; |
554 | } |
555 | } while (--ptr >= end); |
556 | |
557 | if (optimal_gain <= 0) |
558 | return -1; |
559 | dot = avpriv_scalarproduct_float_c(best_hist_ptr, best_hist_ptr, size); |
560 | if (dot <= 0) // would be 1.0 |
561 | return -1; |
562 | |
563 | if (optimal_gain <= dot) { |
564 | dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000 |
565 | } else |
566 | dot = 0.625; |
567 | |
568 | /* actual smoothing */ |
569 | for (n = 0; n < size; n++) |
570 | out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]); |
571 | |
572 | return 0; |
573 | } |
574 | |
575 | /** |
576 | * Get the tilt factor of a formant filter from its transfer function |
577 | * @see #tilt_factor() in amrnbdec.c, which does essentially the same, |
578 | * but somehow (??) it does a speech synthesis filter in the |
579 | * middle, which is missing here |
580 | * |
581 | * @param lpcs LPC coefficients |
582 | * @param n_lpcs Size of LPC buffer |
583 | * @returns the tilt factor |
584 | */ |
585 | static float tilt_factor(const float *lpcs, int n_lpcs) |
586 | { |
587 | float rh0, rh1; |
588 | |
589 | rh0 = 1.0 + avpriv_scalarproduct_float_c(lpcs, lpcs, n_lpcs); |
590 | rh1 = lpcs[0] + avpriv_scalarproduct_float_c(lpcs, &lpcs[1], n_lpcs - 1); |
591 | |
592 | return rh1 / rh0; |
593 | } |
594 | |
595 | /** |
596 | * Derive denoise filter coefficients (in real domain) from the LPCs. |
597 | */ |
598 | static void calc_input_response(WMAVoiceContext *s, float *lpcs, |
599 | int fcb_type, float *coeffs, int remainder) |
600 | { |
601 | float last_coeff, min = 15.0, max = -15.0; |
602 | float irange, angle_mul, gain_mul, range, sq; |
603 | int n, idx; |
604 | |
605 | /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */ |
606 | s->rdft.rdft_calc(&s->rdft, lpcs); |
607 | #define log_range(var, assign) do { \ |
608 | float tmp = log10f(assign); var = tmp; \ |
609 | max = FFMAX(max, tmp); min = FFMIN(min, tmp); \ |
610 | } while (0) |
611 | log_range(last_coeff, lpcs[1] * lpcs[1]); |
612 | for (n = 1; n < 64; n++) |
613 | log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] + |
614 | lpcs[n * 2 + 1] * lpcs[n * 2 + 1]); |
615 | log_range(lpcs[0], lpcs[0] * lpcs[0]); |
616 | #undef log_range |
617 | range = max - min; |
618 | lpcs[64] = last_coeff; |
619 | |
620 | /* Now, use this spectrum to pick out these frequencies with higher |
621 | * (relative) power/energy (which we then take to be "not noise"), |
622 | * and set up a table (still in lpc[]) of (relative) gains per frequency. |
623 | * These frequencies will be maintained, while others ("noise") will be |
624 | * decreased in the filter output. */ |
625 | irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63] |
626 | gain_mul = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) : |
627 | (5.0 / 14.7)); |
628 | angle_mul = gain_mul * (8.0 * M_LN10 / M_PI); |
629 | for (n = 0; n <= 64; n++) { |
630 | float pwr; |
631 | |
632 | idx = FFMAX(0, lrint((max - lpcs[n]) * irange) - 1); |
633 | pwr = wmavoice_denoise_power_table[s->denoise_strength][idx]; |
634 | lpcs[n] = angle_mul * pwr; |
635 | |
636 | /* 70.57 =~ 1/log10(1.0331663) */ |
637 | idx = (pwr * gain_mul - 0.0295) * 70.570526123; |
638 | if (idx > 127) { // fall back if index falls outside table range |
639 | coeffs[n] = wmavoice_energy_table[127] * |
640 | powf(1.0331663, idx - 127); |
641 | } else |
642 | coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)]; |
643 | } |
644 | |
645 | /* calculate the Hilbert transform of the gains, which we do (since this |
646 | * is a sine input) by doing a phase shift (in theory, H(sin())=cos()). |
647 | * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the |
648 | * "moment" of the LPCs in this filter. */ |
649 | s->dct.dct_calc(&s->dct, lpcs); |
650 | s->dst.dct_calc(&s->dst, lpcs); |
651 | |
652 | /* Split out the coefficient indexes into phase/magnitude pairs */ |
653 | idx = 255 + av_clip(lpcs[64], -255, 255); |
654 | coeffs[0] = coeffs[0] * s->cos[idx]; |
655 | idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255); |
656 | last_coeff = coeffs[64] * s->cos[idx]; |
657 | for (n = 63;; n--) { |
658 | idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255); |
659 | coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx]; |
660 | coeffs[n * 2] = coeffs[n] * s->cos[idx]; |
661 | |
662 | if (!--n) break; |
663 | |
664 | idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255); |
665 | coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx]; |
666 | coeffs[n * 2] = coeffs[n] * s->cos[idx]; |
667 | } |
668 | coeffs[1] = last_coeff; |
669 | |
670 | /* move into real domain */ |
671 | s->irdft.rdft_calc(&s->irdft, coeffs); |
672 | |
673 | /* tilt correction and normalize scale */ |
674 | memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder)); |
675 | if (s->denoise_tilt_corr) { |
676 | float tilt_mem = 0; |
677 | |
678 | coeffs[remainder - 1] = 0; |
679 | ff_tilt_compensation(&tilt_mem, |
680 | -1.8 * tilt_factor(coeffs, remainder - 1), |
681 | coeffs, remainder); |
682 | } |
683 | sq = (1.0 / 64.0) * sqrtf(1 / avpriv_scalarproduct_float_c(coeffs, coeffs, |
684 | remainder)); |
685 | for (n = 0; n < remainder; n++) |
686 | coeffs[n] *= sq; |
687 | } |
688 | |
689 | /** |
690 | * This function applies a Wiener filter on the (noisy) speech signal as |
691 | * a means to denoise it. |
692 | * |
693 | * - take RDFT of LPCs to get the power spectrum of the noise + speech; |
694 | * - using this power spectrum, calculate (for each frequency) the Wiener |
695 | * filter gain, which depends on the frequency power and desired level |
696 | * of noise subtraction (when set too high, this leads to artifacts) |
697 | * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse |
698 | * of 4-8kHz); |
699 | * - by doing a phase shift, calculate the Hilbert transform of this array |
700 | * of per-frequency filter-gains to get the filtering coefficients; |
701 | * - smoothen/normalize/de-tilt these filter coefficients as desired; |
702 | * - take RDFT of noisy sound, apply the coefficients and take its IRDFT |
703 | * to get the denoised speech signal; |
704 | * - the leftover (i.e. output of the IRDFT on denoised speech data beyond |
705 | * the frame boundary) are saved and applied to subsequent frames by an |
706 | * overlap-add method (otherwise you get clicking-artifacts). |
707 | * |
708 | * @param s WMA Voice decoding context |
709 | * @param fcb_type Frame (codebook) type |
710 | * @param synth_pf input: the noisy speech signal, output: denoised speech |
711 | * data; should be 16-byte aligned (for ASM purposes) |
712 | * @param size size of the speech data |
713 | * @param lpcs LPCs used to synthesize this frame's speech data |
714 | */ |
715 | static void wiener_denoise(WMAVoiceContext *s, int fcb_type, |
716 | float *synth_pf, int size, |
717 | const float *lpcs) |
718 | { |
719 | int remainder, lim, n; |
720 | |
721 | if (fcb_type != FCB_TYPE_SILENCE) { |
722 | float *tilted_lpcs = s->tilted_lpcs_pf, |
723 | *coeffs = s->denoise_coeffs_pf, tilt_mem = 0; |
724 | |
725 | tilted_lpcs[0] = 1.0; |
726 | memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps); |
727 | memset(&tilted_lpcs[s->lsps + 1], 0, |
728 | sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1)); |
729 | ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps), |
730 | tilted_lpcs, s->lsps + 2); |
731 | |
732 | /* The IRDFT output (127 samples for 7-bit filter) beyond the frame |
733 | * size is applied to the next frame. All input beyond this is zero, |
734 | * and thus all output beyond this will go towards zero, hence we can |
735 | * limit to min(size-1, 127-size) as a performance consideration. */ |
736 | remainder = FFMIN(127 - size, size - 1); |
737 | calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder); |
738 | |
739 | /* apply coefficients (in frequency spectrum domain), i.e. complex |
740 | * number multiplication */ |
741 | memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size)); |
742 | s->rdft.rdft_calc(&s->rdft, synth_pf); |
743 | s->rdft.rdft_calc(&s->rdft, coeffs); |
744 | synth_pf[0] *= coeffs[0]; |
745 | synth_pf[1] *= coeffs[1]; |
746 | for (n = 1; n < 64; n++) { |
747 | float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1]; |
748 | synth_pf[n * 2] = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1]; |
749 | synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1]; |
750 | } |
751 | s->irdft.rdft_calc(&s->irdft, synth_pf); |
752 | } |
753 | |
754 | /* merge filter output with the history of previous runs */ |
755 | if (s->denoise_filter_cache_size) { |
756 | lim = FFMIN(s->denoise_filter_cache_size, size); |
757 | for (n = 0; n < lim; n++) |
758 | synth_pf[n] += s->denoise_filter_cache[n]; |
759 | s->denoise_filter_cache_size -= lim; |
760 | memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size], |
761 | sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size); |
762 | } |
763 | |
764 | /* move remainder of filter output into a cache for future runs */ |
765 | if (fcb_type != FCB_TYPE_SILENCE) { |
766 | lim = FFMIN(remainder, s->denoise_filter_cache_size); |
767 | for (n = 0; n < lim; n++) |
768 | s->denoise_filter_cache[n] += synth_pf[size + n]; |
769 | if (lim < remainder) { |
770 | memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim], |
771 | sizeof(s->denoise_filter_cache[0]) * (remainder - lim)); |
772 | s->denoise_filter_cache_size = remainder; |
773 | } |
774 | } |
775 | } |
776 | |
777 | /** |
778 | * Averaging projection filter, the postfilter used in WMAVoice. |
779 | * |
780 | * This uses the following steps: |
781 | * - A zero-synthesis filter (generate excitation from synth signal) |
782 | * - Kalman smoothing on excitation, based on pitch |
783 | * - Re-synthesized smoothened output |
784 | * - Iterative Wiener denoise filter |
785 | * - Adaptive gain filter |
786 | * - DC filter |
787 | * |
788 | * @param s WMAVoice decoding context |
789 | * @param synth Speech synthesis output (before postfilter) |
790 | * @param samples Output buffer for filtered samples |
791 | * @param size Buffer size of synth & samples |
792 | * @param lpcs Generated LPCs used for speech synthesis |
793 | * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned) |
794 | * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses) |
795 | * @param pitch Pitch of the input signal |
796 | */ |
797 | static void postfilter(WMAVoiceContext *s, const float *synth, |
798 | float *samples, int size, |
799 | const float *lpcs, float *zero_exc_pf, |
800 | int fcb_type, int pitch) |
801 | { |
802 | float synth_filter_in_buf[MAX_FRAMESIZE / 2], |
803 | *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16], |
804 | *synth_filter_in = zero_exc_pf; |
805 | |
806 | av_assert0(size <= MAX_FRAMESIZE / 2); |
807 | |
808 | /* generate excitation from input signal */ |
809 | ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps); |
810 | |
811 | if (fcb_type >= FCB_TYPE_AW_PULSES && |
812 | !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size)) |
813 | synth_filter_in = synth_filter_in_buf; |
814 | |
815 | /* re-synthesize speech after smoothening, and keep history */ |
816 | ff_celp_lp_synthesis_filterf(synth_pf, lpcs, |
817 | synth_filter_in, size, s->lsps); |
818 | memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps], |
819 | sizeof(synth_pf[0]) * s->lsps); |
820 | |
821 | wiener_denoise(s, fcb_type, synth_pf, size, lpcs); |
822 | |
823 | adaptive_gain_control(samples, synth_pf, synth, size, 0.99, |
824 | &s->postfilter_agc); |
825 | |
826 | if (s->dc_level > 8) { |
827 | /* remove ultra-low frequency DC noise / highpass filter; |
828 | * coefficients are identical to those used in SIPR decoding, |
829 | * and very closely resemble those used in AMR-NB decoding. */ |
830 | ff_acelp_apply_order_2_transfer_function(samples, samples, |
831 | (const float[2]) { -1.99997, 1.0 }, |
832 | (const float[2]) { -1.9330735188, 0.93589198496 }, |
833 | 0.93980580475, s->dcf_mem, size); |
834 | } |
835 | } |
836 | /** |
837 | * @} |
838 | */ |
839 | |
840 | /** |
841 | * Dequantize LSPs |
842 | * @param lsps output pointer to the array that will hold the LSPs |
843 | * @param num number of LSPs to be dequantized |
844 | * @param values quantized values, contains n_stages values |
845 | * @param sizes range (i.e. max value) of each quantized value |
846 | * @param n_stages number of dequantization runs |
847 | * @param table dequantization table to be used |
848 | * @param mul_q LSF multiplier |
849 | * @param base_q base (lowest) LSF values |
850 | */ |
851 | static void dequant_lsps(double *lsps, int num, |
852 | const uint16_t *values, |
853 | const uint16_t *sizes, |
854 | int n_stages, const uint8_t *table, |
855 | const double *mul_q, |
856 | const double *base_q) |
857 | { |
858 | int n, m; |
859 | |
860 | memset(lsps, 0, num * sizeof(*lsps)); |
861 | for (n = 0; n < n_stages; n++) { |
862 | const uint8_t *t_off = &table[values[n] * num]; |
863 | double base = base_q[n], mul = mul_q[n]; |
864 | |
865 | for (m = 0; m < num; m++) |
866 | lsps[m] += base + mul * t_off[m]; |
867 | |
868 | table += sizes[n] * num; |
869 | } |
870 | } |
871 | |
872 | /** |
873 | * @name LSP dequantization routines |
874 | * LSP dequantization routines, for 10/16LSPs and independent/residual coding. |
875 | * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits; |
876 | * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits. |
877 | * @{ |
878 | */ |
879 | /** |
880 | * Parse 10 independently-coded LSPs. |
881 | */ |
882 | static void dequant_lsp10i(GetBitContext *gb, double *lsps) |
883 | { |
884 | static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 }; |
885 | static const double mul_lsf[4] = { |
886 | 5.2187144800e-3, 1.4626986422e-3, |
887 | 9.6179549166e-4, 1.1325736225e-3 |
888 | }; |
889 | static const double base_lsf[4] = { |
890 | M_PI * -2.15522e-1, M_PI * -6.1646e-2, |
891 | M_PI * -3.3486e-2, M_PI * -5.7408e-2 |
892 | }; |
893 | uint16_t v[4]; |
894 | |
895 | v[0] = get_bits(gb, 8); |
896 | v[1] = get_bits(gb, 6); |
897 | v[2] = get_bits(gb, 5); |
898 | v[3] = get_bits(gb, 5); |
899 | |
900 | dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i, |
901 | mul_lsf, base_lsf); |
902 | } |
903 | |
904 | /** |
905 | * Parse 10 independently-coded LSPs, and then derive the tables to |
906 | * generate LSPs for the other frames from them (residual coding). |
907 | */ |
908 | static void dequant_lsp10r(GetBitContext *gb, |
909 | double *i_lsps, const double *old, |
910 | double *a1, double *a2, int q_mode) |
911 | { |
912 | static const uint16_t vec_sizes[3] = { 128, 64, 64 }; |
913 | static const double mul_lsf[3] = { |
914 | 2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3 |
915 | }; |
916 | static const double base_lsf[3] = { |
917 | M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2 |
918 | }; |
919 | const float (*ipol_tab)[2][10] = q_mode ? |
920 | wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a; |
921 | uint16_t interpol, v[3]; |
922 | int n; |
923 | |
924 | dequant_lsp10i(gb, i_lsps); |
925 | |
926 | interpol = get_bits(gb, 5); |
927 | v[0] = get_bits(gb, 7); |
928 | v[1] = get_bits(gb, 6); |
929 | v[2] = get_bits(gb, 6); |
930 | |
931 | for (n = 0; n < 10; n++) { |
932 | double delta = old[n] - i_lsps[n]; |
933 | a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n]; |
934 | a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n]; |
935 | } |
936 | |
937 | dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r, |
938 | mul_lsf, base_lsf); |
939 | } |
940 | |
941 | /** |
942 | * Parse 16 independently-coded LSPs. |
943 | */ |
944 | static void dequant_lsp16i(GetBitContext *gb, double *lsps) |
945 | { |
946 | static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 }; |
947 | static const double mul_lsf[5] = { |
948 | 3.3439586280e-3, 6.9908173703e-4, |
949 | 3.3216608306e-3, 1.0334960326e-3, |
950 | 3.1899104283e-3 |
951 | }; |
952 | static const double base_lsf[5] = { |
953 | M_PI * -1.27576e-1, M_PI * -2.4292e-2, |
954 | M_PI * -1.28094e-1, M_PI * -3.2128e-2, |
955 | M_PI * -1.29816e-1 |
956 | }; |
957 | uint16_t v[5]; |
958 | |
959 | v[0] = get_bits(gb, 8); |
960 | v[1] = get_bits(gb, 6); |
961 | v[2] = get_bits(gb, 7); |
962 | v[3] = get_bits(gb, 6); |
963 | v[4] = get_bits(gb, 7); |
964 | |
965 | dequant_lsps( lsps, 5, v, vec_sizes, 2, |
966 | wmavoice_dq_lsp16i1, mul_lsf, base_lsf); |
967 | dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2, |
968 | wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]); |
969 | dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1, |
970 | wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]); |
971 | } |
972 | |
973 | /** |
974 | * Parse 16 independently-coded LSPs, and then derive the tables to |
975 | * generate LSPs for the other frames from them (residual coding). |
976 | */ |
977 | static void dequant_lsp16r(GetBitContext *gb, |
978 | double *i_lsps, const double *old, |
979 | double *a1, double *a2, int q_mode) |
980 | { |
981 | static const uint16_t vec_sizes[3] = { 128, 128, 128 }; |
982 | static const double mul_lsf[3] = { |
983 | 1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3 |
984 | }; |
985 | static const double base_lsf[3] = { |
986 | M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2 |
987 | }; |
988 | const float (*ipol_tab)[2][16] = q_mode ? |
989 | wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a; |
990 | uint16_t interpol, v[3]; |
991 | int n; |
992 | |
993 | dequant_lsp16i(gb, i_lsps); |
994 | |
995 | interpol = get_bits(gb, 5); |
996 | v[0] = get_bits(gb, 7); |
997 | v[1] = get_bits(gb, 7); |
998 | v[2] = get_bits(gb, 7); |
999 | |
1000 | for (n = 0; n < 16; n++) { |
1001 | double delta = old[n] - i_lsps[n]; |
1002 | a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n]; |
1003 | a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n]; |
1004 | } |
1005 | |
1006 | dequant_lsps( a2, 10, v, vec_sizes, 1, |
1007 | wmavoice_dq_lsp16r1, mul_lsf, base_lsf); |
1008 | dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1, |
1009 | wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]); |
1010 | dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1, |
1011 | wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]); |
1012 | } |
1013 | |
1014 | /** |
1015 | * @} |
1016 | * @name Pitch-adaptive window coding functions |
1017 | * The next few functions are for pitch-adaptive window coding. |
1018 | * @{ |
1019 | */ |
1020 | /** |
1021 | * Parse the offset of the first pitch-adaptive window pulses, and |
1022 | * the distribution of pulses between the two blocks in this frame. |
1023 | * @param s WMA Voice decoding context private data |
1024 | * @param gb bit I/O context |
1025 | * @param pitch pitch for each block in this frame |
1026 | */ |
1027 | static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb, |
1028 | const int *pitch) |
1029 | { |
1030 | static const int16_t start_offset[94] = { |
1031 | -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11, |
1032 | 13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26, |
1033 | 27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43, |
1034 | 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67, |
1035 | 69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91, |
1036 | 93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115, |
1037 | 117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139, |
1038 | 141, 143, 145, 147, 149, 151, 153, 155, 157, 159 |
1039 | }; |
1040 | int bits, offset; |
1041 | |
1042 | /* position of pulse */ |
1043 | s->aw_idx_is_ext = 0; |
1044 | if ((bits = get_bits(gb, 6)) >= 54) { |
1045 | s->aw_idx_is_ext = 1; |
1046 | bits += (bits - 54) * 3 + get_bits(gb, 2); |
1047 | } |
1048 | |
1049 | /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count |
1050 | * the distribution of the pulses in each block contained in this frame. */ |
1051 | s->aw_pulse_range = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16; |
1052 | for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ; |
1053 | s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0]; |
1054 | s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2; |
1055 | offset += s->aw_n_pulses[0] * pitch[0]; |
1056 | s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1]; |
1057 | s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2; |
1058 | |
1059 | /* if continuing from a position before the block, reset position to |
1060 | * start of block (when corrected for the range over which it can be |
1061 | * spread in aw_pulse_set1()). */ |
1062 | if (start_offset[bits] < MAX_FRAMESIZE / 2) { |
1063 | while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0) |
1064 | s->aw_first_pulse_off[1] -= pitch[1]; |
1065 | if (start_offset[bits] < 0) |
1066 | while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0) |
1067 | s->aw_first_pulse_off[0] -= pitch[0]; |
1068 | } |
1069 | } |
1070 | |
1071 | /** |
1072 | * Apply second set of pitch-adaptive window pulses. |
1073 | * @param s WMA Voice decoding context private data |
1074 | * @param gb bit I/O context |
1075 | * @param block_idx block index in frame [0, 1] |
1076 | * @param fcb structure containing fixed codebook vector info |
1077 | * @return -1 on error, 0 otherwise |
1078 | */ |
1079 | static int aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb, |
1080 | int block_idx, AMRFixed *fcb) |
1081 | { |
1082 | uint16_t use_mask_mem[9]; // only 5 are used, rest is padding |
1083 | uint16_t *use_mask = use_mask_mem + 2; |
1084 | /* in this function, idx is the index in the 80-bit (+ padding) use_mask |
1085 | * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits |
1086 | * of idx are the position of the bit within a particular item in the |
1087 | * array (0 being the most significant bit, and 15 being the least |
1088 | * significant bit), and the remainder (>> 4) is the index in the |
1089 | * use_mask[]-array. This is faster and uses less memory than using a |
1090 | * 80-byte/80-int array. */ |
1091 | int pulse_off = s->aw_first_pulse_off[block_idx], |
1092 | pulse_start, n, idx, range, aidx, start_off = 0; |
1093 | |
1094 | /* set offset of first pulse to within this block */ |
1095 | if (s->aw_n_pulses[block_idx] > 0) |
1096 | while (pulse_off + s->aw_pulse_range < 1) |
1097 | pulse_off += fcb->pitch_lag; |
1098 | |
1099 | /* find range per pulse */ |
1100 | if (s->aw_n_pulses[0] > 0) { |
1101 | if (block_idx == 0) { |
1102 | range = 32; |
1103 | } else /* block_idx = 1 */ { |
1104 | range = 8; |
1105 | if (s->aw_n_pulses[block_idx] > 0) |
1106 | pulse_off = s->aw_next_pulse_off_cache; |
1107 | } |
1108 | } else |
1109 | range = 16; |
1110 | pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0; |
1111 | |
1112 | /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly, |
1113 | * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus |
1114 | * we exclude that range from being pulsed again in this function. */ |
1115 | memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0])); |
1116 | memset( use_mask, -1, 5 * sizeof(use_mask[0])); |
1117 | memset(&use_mask[5], 0, 2 * sizeof(use_mask[0])); |
1118 | if (s->aw_n_pulses[block_idx] > 0) |
1119 | for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) { |
1120 | int excl_range = s->aw_pulse_range; // always 16 or 24 |
1121 | uint16_t *use_mask_ptr = &use_mask[idx >> 4]; |
1122 | int first_sh = 16 - (idx & 15); |
1123 | *use_mask_ptr++ &= 0xFFFFu << first_sh; |
1124 | excl_range -= first_sh; |
1125 | if (excl_range >= 16) { |
1126 | *use_mask_ptr++ = 0; |
1127 | *use_mask_ptr &= 0xFFFF >> (excl_range - 16); |
1128 | } else |
1129 | *use_mask_ptr &= 0xFFFF >> excl_range; |
1130 | } |
1131 | |
1132 | /* find the 'aidx'th offset that is not excluded */ |
1133 | aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4); |
1134 | for (n = 0; n <= aidx; pulse_start++) { |
1135 | for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ; |
1136 | if (idx >= MAX_FRAMESIZE / 2) { // find from zero |
1137 | if (use_mask[0]) idx = 0x0F; |
1138 | else if (use_mask[1]) idx = 0x1F; |
1139 | else if (use_mask[2]) idx = 0x2F; |
1140 | else if (use_mask[3]) idx = 0x3F; |
1141 | else if (use_mask[4]) idx = 0x4F; |
1142 | else return -1; |
1143 | idx -= av_log2_16bit(use_mask[idx >> 4]); |
1144 | } |
1145 | if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) { |
1146 | use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15)); |
1147 | n++; |
1148 | start_off = idx; |
1149 | } |
1150 | } |
1151 | |
1152 | fcb->x[fcb->n] = start_off; |
1153 | fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0; |
1154 | fcb->n++; |
1155 | |
1156 | /* set offset for next block, relative to start of that block */ |
1157 | n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag; |
1158 | s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0; |
1159 | return 0; |
1160 | } |
1161 | |
1162 | /** |
1163 | * Apply first set of pitch-adaptive window pulses. |
1164 | * @param s WMA Voice decoding context private data |
1165 | * @param gb bit I/O context |
1166 | * @param block_idx block index in frame [0, 1] |
1167 | * @param fcb storage location for fixed codebook pulse info |
1168 | */ |
1169 | static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb, |
1170 | int block_idx, AMRFixed *fcb) |
1171 | { |
1172 | int val = get_bits(gb, 12 - 2 * (s->aw_idx_is_ext && !block_idx)); |
1173 | float v; |
1174 | |
1175 | if (s->aw_n_pulses[block_idx] > 0) { |
1176 | int n, v_mask, i_mask, sh, n_pulses; |
1177 | |
1178 | if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each |
1179 | n_pulses = 3; |
1180 | v_mask = 8; |
1181 | i_mask = 7; |
1182 | sh = 4; |
1183 | } else { // 4 pulses, 1:sign + 2:index each |
1184 | n_pulses = 4; |
1185 | v_mask = 4; |
1186 | i_mask = 3; |
1187 | sh = 3; |
1188 | } |
1189 | |
1190 | for (n = n_pulses - 1; n >= 0; n--, val >>= sh) { |
1191 | fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0; |
1192 | fcb->x[fcb->n] = (val & i_mask) * n_pulses + n + |
1193 | s->aw_first_pulse_off[block_idx]; |
1194 | while (fcb->x[fcb->n] < 0) |
1195 | fcb->x[fcb->n] += fcb->pitch_lag; |
1196 | if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2) |
1197 | fcb->n++; |
1198 | } |
1199 | } else { |
1200 | int num2 = (val & 0x1FF) >> 1, delta, idx; |
1201 | |
1202 | if (num2 < 1 * 79) { delta = 1; idx = num2 + 1; } |
1203 | else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; } |
1204 | else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; } |
1205 | else { delta = 7; idx = num2 + 1 - 3 * 75; } |
1206 | v = (val & 0x200) ? -1.0 : 1.0; |
1207 | |
1208 | fcb->no_repeat_mask |= 3 << fcb->n; |
1209 | fcb->x[fcb->n] = idx - delta; |
1210 | fcb->y[fcb->n] = v; |
1211 | fcb->x[fcb->n + 1] = idx; |
1212 | fcb->y[fcb->n + 1] = (val & 1) ? -v : v; |
1213 | fcb->n += 2; |
1214 | } |
1215 | } |
1216 | |
1217 | /** |
1218 | * @} |
1219 | * |
1220 | * Generate a random number from frame_cntr and block_idx, which will live |
1221 | * in the range [0, 1000 - block_size] (so it can be used as an index in a |
1222 | * table of size 1000 of which you want to read block_size entries). |
1223 | * |
1224 | * @param frame_cntr current frame number |
1225 | * @param block_num current block index |
1226 | * @param block_size amount of entries we want to read from a table |
1227 | * that has 1000 entries |
1228 | * @return a (non-)random number in the [0, 1000 - block_size] range. |
1229 | */ |
1230 | static int pRNG(int frame_cntr, int block_num, int block_size) |
1231 | { |
1232 | /* array to simplify the calculation of z: |
1233 | * y = (x % 9) * 5 + 6; |
1234 | * z = (49995 * x) / y; |
1235 | * Since y only has 9 values, we can remove the division by using a |
1236 | * LUT and using FASTDIV-style divisions. For each of the 9 values |
1237 | * of y, we can rewrite z as: |
1238 | * z = x * (49995 / y) + x * ((49995 % y) / y) |
1239 | * In this table, each col represents one possible value of y, the |
1240 | * first number is 49995 / y, and the second is the FASTDIV variant |
1241 | * of 49995 % y / y. */ |
1242 | static const unsigned int div_tbl[9][2] = { |
1243 | { 8332, 3 * 715827883U }, // y = 6 |
1244 | { 4545, 0 * 390451573U }, // y = 11 |
1245 | { 3124, 11 * 268435456U }, // y = 16 |
1246 | { 2380, 15 * 204522253U }, // y = 21 |
1247 | { 1922, 23 * 165191050U }, // y = 26 |
1248 | { 1612, 23 * 138547333U }, // y = 31 |
1249 | { 1388, 27 * 119304648U }, // y = 36 |
1250 | { 1219, 16 * 104755300U }, // y = 41 |
1251 | { 1086, 39 * 93368855U } // y = 46 |
1252 | }; |
1253 | unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr; |
1254 | if (x >= 0xFFFF) x -= 0xFFFF; // max value of x is 8*1877+0xFFFE=0x13AA6, |
1255 | // so this is effectively a modulo (%) |
1256 | y = x - 9 * MULH(477218589, x); // x % 9 |
1257 | z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1])); |
1258 | // z = x * 49995 / (y * 5 + 6) |
1259 | return z % (1000 - block_size); |
1260 | } |
1261 | |
1262 | /** |
1263 | * Parse hardcoded signal for a single block. |
1264 | * @note see #synth_block(). |
1265 | */ |
1266 | static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb, |
1267 | int block_idx, int size, |
1268 | const struct frame_type_desc *frame_desc, |
1269 | float *excitation) |
1270 | { |
1271 | float gain; |
1272 | int n, r_idx; |
1273 | |
1274 | av_assert0(size <= MAX_FRAMESIZE); |
1275 | |
1276 | /* Set the offset from which we start reading wmavoice_std_codebook */ |
1277 | if (frame_desc->fcb_type == FCB_TYPE_SILENCE) { |
1278 | r_idx = pRNG(s->frame_cntr, block_idx, size); |
1279 | gain = s->silence_gain; |
1280 | } else /* FCB_TYPE_HARDCODED */ { |
1281 | r_idx = get_bits(gb, 8); |
1282 | gain = wmavoice_gain_universal[get_bits(gb, 6)]; |
1283 | } |
1284 | |
1285 | /* Clear gain prediction parameters */ |
1286 | memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err)); |
1287 | |
1288 | /* Apply gain to hardcoded codebook and use that as excitation signal */ |
1289 | for (n = 0; n < size; n++) |
1290 | excitation[n] = wmavoice_std_codebook[r_idx + n] * gain; |
1291 | } |
1292 | |
1293 | /** |
1294 | * Parse FCB/ACB signal for a single block. |
1295 | * @note see #synth_block(). |
1296 | */ |
1297 | static void synth_block_fcb_acb(WMAVoiceContext *s, GetBitContext *gb, |
1298 | int block_idx, int size, |
1299 | int block_pitch_sh2, |
1300 | const struct frame_type_desc *frame_desc, |
1301 | float *excitation) |
1302 | { |
1303 | static const float gain_coeff[6] = { |
1304 | 0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458 |
1305 | }; |
1306 | float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain; |
1307 | int n, idx, gain_weight; |
1308 | AMRFixed fcb; |
1309 | |
1310 | av_assert0(size <= MAX_FRAMESIZE / 2); |
1311 | memset(pulses, 0, sizeof(*pulses) * size); |
1312 | |
1313 | fcb.pitch_lag = block_pitch_sh2 >> 2; |
1314 | fcb.pitch_fac = 1.0; |
1315 | fcb.no_repeat_mask = 0; |
1316 | fcb.n = 0; |
1317 | |
1318 | /* For the other frame types, this is where we apply the innovation |
1319 | * (fixed) codebook pulses of the speech signal. */ |
1320 | if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) { |
1321 | aw_pulse_set1(s, gb, block_idx, &fcb); |
1322 | if (aw_pulse_set2(s, gb, block_idx, &fcb)) { |
1323 | /* Conceal the block with silence and return. |
1324 | * Skip the correct amount of bits to read the next |
1325 | * block from the correct offset. */ |
1326 | int r_idx = pRNG(s->frame_cntr, block_idx, size); |
1327 | |
1328 | for (n = 0; n < size; n++) |
1329 | excitation[n] = |
1330 | wmavoice_std_codebook[r_idx + n] * s->silence_gain; |
1331 | skip_bits(gb, 7 + 1); |
1332 | return; |
1333 | } |
1334 | } else /* FCB_TYPE_EXC_PULSES */ { |
1335 | int offset_nbits = 5 - frame_desc->log_n_blocks; |
1336 | |
1337 | fcb.no_repeat_mask = -1; |
1338 | /* similar to ff_decode_10_pulses_35bits(), but with single pulses |
1339 | * (instead of double) for a subset of pulses */ |
1340 | for (n = 0; n < 5; n++) { |
1341 | float sign; |
1342 | int pos1, pos2; |
1343 | |
1344 | sign = get_bits1(gb) ? 1.0 : -1.0; |
1345 | pos1 = get_bits(gb, offset_nbits); |
1346 | fcb.x[fcb.n] = n + 5 * pos1; |
1347 | fcb.y[fcb.n++] = sign; |
1348 | if (n < frame_desc->dbl_pulses) { |
1349 | pos2 = get_bits(gb, offset_nbits); |
1350 | fcb.x[fcb.n] = n + 5 * pos2; |
1351 | fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign; |
1352 | } |
1353 | } |
1354 | } |
1355 | ff_set_fixed_vector(pulses, &fcb, 1.0, size); |
1356 | |
1357 | /* Calculate gain for adaptive & fixed codebook signal. |
1358 | * see ff_amr_set_fixed_gain(). */ |
1359 | idx = get_bits(gb, 7); |
1360 | fcb_gain = expf(avpriv_scalarproduct_float_c(s->gain_pred_err, |
1361 | gain_coeff, 6) - |
1362 | 5.2409161640 + wmavoice_gain_codebook_fcb[idx]); |
1363 | acb_gain = wmavoice_gain_codebook_acb[idx]; |
1364 | pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx], |
1365 | -2.9957322736 /* log(0.05) */, |
1366 | 1.6094379124 /* log(5.0) */); |
1367 | |
1368 | gain_weight = 8 >> frame_desc->log_n_blocks; |
1369 | memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err, |
1370 | sizeof(*s->gain_pred_err) * (6 - gain_weight)); |
1371 | for (n = 0; n < gain_weight; n++) |
1372 | s->gain_pred_err[n] = pred_err; |
1373 | |
1374 | /* Calculation of adaptive codebook */ |
1375 | if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) { |
1376 | int len; |
1377 | for (n = 0; n < size; n += len) { |
1378 | int next_idx_sh16; |
1379 | int abs_idx = block_idx * size + n; |
1380 | int pitch_sh16 = (s->last_pitch_val << 16) + |
1381 | s->pitch_diff_sh16 * abs_idx; |
1382 | int pitch = (pitch_sh16 + 0x6FFF) >> 16; |
1383 | int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000; |
1384 | idx = idx_sh16 >> 16; |
1385 | if (s->pitch_diff_sh16) { |
1386 | if (s->pitch_diff_sh16 > 0) { |
1387 | next_idx_sh16 = (idx_sh16) &~ 0xFFFF; |
1388 | } else |
1389 | next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF; |
1390 | len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8, |
1391 | 1, size - n); |
1392 | } else |
1393 | len = size; |
1394 | |
1395 | ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch], |
1396 | wmavoice_ipol1_coeffs, 17, |
1397 | idx, 9, len); |
1398 | } |
1399 | } else /* ACB_TYPE_HAMMING */ { |
1400 | int block_pitch = block_pitch_sh2 >> 2; |
1401 | idx = block_pitch_sh2 & 3; |
1402 | if (idx) { |
1403 | ff_acelp_interpolatef(excitation, &excitation[-block_pitch], |
1404 | wmavoice_ipol2_coeffs, 4, |
1405 | idx, 8, size); |
1406 | } else |
1407 | av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch, |
1408 | sizeof(float) * size); |
1409 | } |
1410 | |
1411 | /* Interpolate ACB/FCB and use as excitation signal */ |
1412 | ff_weighted_vector_sumf(excitation, excitation, pulses, |
1413 | acb_gain, fcb_gain, size); |
1414 | } |
1415 | |
1416 | /** |
1417 | * Parse data in a single block. |
1418 | * |
1419 | * @param s WMA Voice decoding context private data |
1420 | * @param gb bit I/O context |
1421 | * @param block_idx index of the to-be-read block |
1422 | * @param size amount of samples to be read in this block |
1423 | * @param block_pitch_sh2 pitch for this block << 2 |
1424 | * @param lsps LSPs for (the end of) this frame |
1425 | * @param prev_lsps LSPs for the last frame |
1426 | * @param frame_desc frame type descriptor |
1427 | * @param excitation target memory for the ACB+FCB interpolated signal |
1428 | * @param synth target memory for the speech synthesis filter output |
1429 | * @return 0 on success, <0 on error. |
1430 | */ |
1431 | static void synth_block(WMAVoiceContext *s, GetBitContext *gb, |
1432 | int block_idx, int size, |
1433 | int block_pitch_sh2, |
1434 | const double *lsps, const double *prev_lsps, |
1435 | const struct frame_type_desc *frame_desc, |
1436 | float *excitation, float *synth) |
1437 | { |
1438 | double i_lsps[MAX_LSPS]; |
1439 | float lpcs[MAX_LSPS]; |
1440 | float fac; |
1441 | int n; |
1442 | |
1443 | if (frame_desc->acb_type == ACB_TYPE_NONE) |
1444 | synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation); |
1445 | else |
1446 | synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2, |
1447 | frame_desc, excitation); |
1448 | |
1449 | /* convert interpolated LSPs to LPCs */ |
1450 | fac = (block_idx + 0.5) / frame_desc->n_blocks; |
1451 | for (n = 0; n < s->lsps; n++) // LSF -> LSP |
1452 | i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n])); |
1453 | ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1); |
1454 | |
1455 | /* Speech synthesis */ |
1456 | ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps); |
1457 | } |
1458 | |
1459 | /** |
1460 | * Synthesize output samples for a single frame. |
1461 | * |
1462 | * @param ctx WMA Voice decoder context |
1463 | * @param gb bit I/O context (s->gb or one for cross-packet superframes) |
1464 | * @param frame_idx Frame number within superframe [0-2] |
1465 | * @param samples pointer to output sample buffer, has space for at least 160 |
1466 | * samples |
1467 | * @param lsps LSP array |
1468 | * @param prev_lsps array of previous frame's LSPs |
1469 | * @param excitation target buffer for excitation signal |
1470 | * @param synth target buffer for synthesized speech data |
1471 | * @return 0 on success, <0 on error. |
1472 | */ |
1473 | static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx, |
1474 | float *samples, |
1475 | const double *lsps, const double *prev_lsps, |
1476 | float *excitation, float *synth) |
1477 | { |
1478 | WMAVoiceContext *s = ctx->priv_data; |
1479 | int n, n_blocks_x2, log_n_blocks_x2, av_uninit(cur_pitch_val); |
1480 | int pitch[MAX_BLOCKS], av_uninit(last_block_pitch); |
1481 | |
1482 | /* Parse frame type ("frame header"), see frame_descs */ |
1483 | int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)], block_nsamples; |
1484 | |
1485 | if (bd_idx < 0) { |
1486 | av_log(ctx, AV_LOG_ERROR, |
1487 | "Invalid frame type VLC code, skipping\n"); |
1488 | return AVERROR_INVALIDDATA; |
1489 | } |
1490 | |
1491 | block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks; |
1492 | |
1493 | /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */ |
1494 | if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) { |
1495 | /* Pitch is provided per frame, which is interpreted as the pitch of |
1496 | * the last sample of the last block of this frame. We can interpolate |
1497 | * the pitch of other blocks (and even pitch-per-sample) by gradually |
1498 | * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */ |
1499 | n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1; |
1500 | log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1; |
1501 | cur_pitch_val = s->min_pitch_val + get_bits(gb, s->pitch_nbits); |
1502 | cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1); |
1503 | if (s->last_acb_type == ACB_TYPE_NONE || |
1504 | 20 * abs(cur_pitch_val - s->last_pitch_val) > |
1505 | (cur_pitch_val + s->last_pitch_val)) |
1506 | s->last_pitch_val = cur_pitch_val; |
1507 | |
1508 | /* pitch per block */ |
1509 | for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) { |
1510 | int fac = n * 2 + 1; |
1511 | |
1512 | pitch[n] = (MUL16(fac, cur_pitch_val) + |
1513 | MUL16((n_blocks_x2 - fac), s->last_pitch_val) + |
1514 | frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2; |
1515 | } |
1516 | |
1517 | /* "pitch-diff-per-sample" for calculation of pitch per sample */ |
1518 | s->pitch_diff_sh16 = |
1519 | ((cur_pitch_val - s->last_pitch_val) << 16) / MAX_FRAMESIZE; |
1520 | } |
1521 | |
1522 | /* Global gain (if silence) and pitch-adaptive window coordinates */ |
1523 | switch (frame_descs[bd_idx].fcb_type) { |
1524 | case FCB_TYPE_SILENCE: |
1525 | s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)]; |
1526 | break; |
1527 | case FCB_TYPE_AW_PULSES: |
1528 | aw_parse_coords(s, gb, pitch); |
1529 | break; |
1530 | } |
1531 | |
1532 | for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) { |
1533 | int bl_pitch_sh2; |
1534 | |
1535 | /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */ |
1536 | switch (frame_descs[bd_idx].acb_type) { |
1537 | case ACB_TYPE_HAMMING: { |
1538 | /* Pitch is given per block. Per-block pitches are encoded as an |
1539 | * absolute value for the first block, and then delta values |
1540 | * relative to this value) for all subsequent blocks. The scale of |
1541 | * this pitch value is semi-logarithmic compared to its use in the |
1542 | * decoder, so we convert it to normal scale also. */ |
1543 | int block_pitch, |
1544 | t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2, |
1545 | t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1, |
1546 | t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1; |
1547 | |
1548 | if (n == 0) { |
1549 | block_pitch = get_bits(gb, s->block_pitch_nbits); |
1550 | } else |
1551 | block_pitch = last_block_pitch - s->block_delta_pitch_hrange + |
1552 | get_bits(gb, s->block_delta_pitch_nbits); |
1553 | /* Convert last_ so that any next delta is within _range */ |
1554 | last_block_pitch = av_clip(block_pitch, |
1555 | s->block_delta_pitch_hrange, |
1556 | s->block_pitch_range - |
1557 | s->block_delta_pitch_hrange); |
1558 | |
1559 | /* Convert semi-log-style scale back to normal scale */ |
1560 | if (block_pitch < t1) { |
1561 | bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch; |
1562 | } else { |
1563 | block_pitch -= t1; |
1564 | if (block_pitch < t2) { |
1565 | bl_pitch_sh2 = |
1566 | (s->block_conv_table[1] << 2) + (block_pitch << 1); |
1567 | } else { |
1568 | block_pitch -= t2; |
1569 | if (block_pitch < t3) { |
1570 | bl_pitch_sh2 = |
1571 | (s->block_conv_table[2] + block_pitch) << 2; |
1572 | } else |
1573 | bl_pitch_sh2 = s->block_conv_table[3] << 2; |
1574 | } |
1575 | } |
1576 | pitch[n] = bl_pitch_sh2 >> 2; |
1577 | break; |
1578 | } |
1579 | |
1580 | case ACB_TYPE_ASYMMETRIC: { |
1581 | bl_pitch_sh2 = pitch[n] << 2; |
1582 | break; |
1583 | } |
1584 | |
1585 | default: // ACB_TYPE_NONE has no pitch |
1586 | bl_pitch_sh2 = 0; |
1587 | break; |
1588 | } |
1589 | |
1590 | synth_block(s, gb, n, block_nsamples, bl_pitch_sh2, |
1591 | lsps, prev_lsps, &frame_descs[bd_idx], |
1592 | &excitation[n * block_nsamples], |
1593 | &synth[n * block_nsamples]); |
1594 | } |
1595 | |
1596 | /* Averaging projection filter, if applicable. Else, just copy samples |
1597 | * from synthesis buffer */ |
1598 | if (s->do_apf) { |
1599 | double i_lsps[MAX_LSPS]; |
1600 | float lpcs[MAX_LSPS]; |
1601 | |
1602 | for (n = 0; n < s->lsps; n++) // LSF -> LSP |
1603 | i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n])); |
1604 | ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1); |
1605 | postfilter(s, synth, samples, 80, lpcs, |
1606 | &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx], |
1607 | frame_descs[bd_idx].fcb_type, pitch[0]); |
1608 | |
1609 | for (n = 0; n < s->lsps; n++) // LSF -> LSP |
1610 | i_lsps[n] = cos(lsps[n]); |
1611 | ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1); |
1612 | postfilter(s, &synth[80], &samples[80], 80, lpcs, |
1613 | &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80], |
1614 | frame_descs[bd_idx].fcb_type, pitch[0]); |
1615 | } else |
1616 | memcpy(samples, synth, 160 * sizeof(synth[0])); |
1617 | |
1618 | /* Cache values for next frame */ |
1619 | s->frame_cntr++; |
1620 | if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%) |
1621 | s->last_acb_type = frame_descs[bd_idx].acb_type; |
1622 | switch (frame_descs[bd_idx].acb_type) { |
1623 | case ACB_TYPE_NONE: |
1624 | s->last_pitch_val = 0; |
1625 | break; |
1626 | case ACB_TYPE_ASYMMETRIC: |
1627 | s->last_pitch_val = cur_pitch_val; |
1628 | break; |
1629 | case ACB_TYPE_HAMMING: |
1630 | s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1]; |
1631 | break; |
1632 | } |
1633 | |
1634 | return 0; |
1635 | } |
1636 | |
1637 | /** |
1638 | * Ensure minimum value for first item, maximum value for last value, |
1639 | * proper spacing between each value and proper ordering. |
1640 | * |
1641 | * @param lsps array of LSPs |
1642 | * @param num size of LSP array |
1643 | * |
1644 | * @note basically a double version of #ff_acelp_reorder_lsf(), might be |
1645 | * useful to put in a generic location later on. Parts are also |
1646 | * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(), |
1647 | * which is in float. |
1648 | */ |
1649 | static void stabilize_lsps(double *lsps, int num) |
1650 | { |
1651 | int n, m, l; |
1652 | |
1653 | /* set minimum value for first, maximum value for last and minimum |
1654 | * spacing between LSF values. |
1655 | * Very similar to ff_set_min_dist_lsf(), but in double. */ |
1656 | lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI); |
1657 | for (n = 1; n < num; n++) |
1658 | lsps[n] = FFMAX(lsps[n], lsps[n - 1] + 0.0125 * M_PI); |
1659 | lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI); |
1660 | |
1661 | /* reorder (looks like one-time / non-recursed bubblesort). |
1662 | * Very similar to ff_sort_nearly_sorted_floats(), but in double. */ |
1663 | for (n = 1; n < num; n++) { |
1664 | if (lsps[n] < lsps[n - 1]) { |
1665 | for (m = 1; m < num; m++) { |
1666 | double tmp = lsps[m]; |
1667 | for (l = m - 1; l >= 0; l--) { |
1668 | if (lsps[l] <= tmp) break; |
1669 | lsps[l + 1] = lsps[l]; |
1670 | } |
1671 | lsps[l + 1] = tmp; |
1672 | } |
1673 | break; |
1674 | } |
1675 | } |
1676 | } |
1677 | |
1678 | /** |
1679 | * Synthesize output samples for a single superframe. If we have any data |
1680 | * cached in s->sframe_cache, that will be used instead of whatever is loaded |
1681 | * in s->gb. |
1682 | * |
1683 | * WMA Voice superframes contain 3 frames, each containing 160 audio samples, |
1684 | * to give a total of 480 samples per frame. See #synth_frame() for frame |
1685 | * parsing. In addition to 3 frames, superframes can also contain the LSPs |
1686 | * (if these are globally specified for all frames (residually); they can |
1687 | * also be specified individually per-frame. See the s->has_residual_lsps |
1688 | * option), and can specify the number of samples encoded in this superframe |
1689 | * (if less than 480), usually used to prevent blanks at track boundaries. |
1690 | * |
1691 | * @param ctx WMA Voice decoder context |
1692 | * @return 0 on success, <0 on error or 1 if there was not enough data to |
1693 | * fully parse the superframe |
1694 | */ |
1695 | static int synth_superframe(AVCodecContext *ctx, AVFrame *frame, |
1696 | int *got_frame_ptr) |
1697 | { |
1698 | WMAVoiceContext *s = ctx->priv_data; |
1699 | GetBitContext *gb = &s->gb, s_gb; |
1700 | int n, res, n_samples = MAX_SFRAMESIZE; |
1701 | double lsps[MAX_FRAMES][MAX_LSPS]; |
1702 | const double *mean_lsf = s->lsps == 16 ? |
1703 | wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode]; |
1704 | float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12]; |
1705 | float synth[MAX_LSPS + MAX_SFRAMESIZE]; |
1706 | float *samples; |
1707 | |
1708 | memcpy(synth, s->synth_history, |
1709 | s->lsps * sizeof(*synth)); |
1710 | memcpy(excitation, s->excitation_history, |
1711 | s->history_nsamples * sizeof(*excitation)); |
1712 | |
1713 | if (s->sframe_cache_size > 0) { |
1714 | gb = &s_gb; |
1715 | init_get_bits(gb, s->sframe_cache, s->sframe_cache_size); |
1716 | s->sframe_cache_size = 0; |
1717 | } |
1718 | |
1719 | /* First bit is speech/music bit, it differentiates between WMAVoice |
1720 | * speech samples (the actual codec) and WMAVoice music samples, which |
1721 | * are really WMAPro-in-WMAVoice-superframes. I've never seen those in |
1722 | * the wild yet. */ |
1723 | if (!get_bits1(gb)) { |
1724 | avpriv_request_sample(ctx, "WMAPro-in-WMAVoice"); |
1725 | return AVERROR_PATCHWELCOME; |
1726 | } |
1727 | |
1728 | /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */ |
1729 | if (get_bits1(gb)) { |
1730 | if ((n_samples = get_bits(gb, 12)) > MAX_SFRAMESIZE) { |
1731 | av_log(ctx, AV_LOG_ERROR, |
1732 | "Superframe encodes > %d samples (%d), not allowed\n", |
1733 | MAX_SFRAMESIZE, n_samples); |
1734 | return AVERROR_INVALIDDATA; |
1735 | } |
1736 | } |
1737 | |
1738 | /* Parse LSPs, if global for the superframe (can also be per-frame). */ |
1739 | if (s->has_residual_lsps) { |
1740 | double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2]; |
1741 | |
1742 | for (n = 0; n < s->lsps; n++) |
1743 | prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n]; |
1744 | |
1745 | if (s->lsps == 10) { |
1746 | dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode); |
1747 | } else /* s->lsps == 16 */ |
1748 | dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode); |
1749 | |
1750 | for (n = 0; n < s->lsps; n++) { |
1751 | lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]); |
1752 | lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]); |
1753 | lsps[2][n] += mean_lsf[n]; |
1754 | } |
1755 | for (n = 0; n < 3; n++) |
1756 | stabilize_lsps(lsps[n], s->lsps); |
1757 | } |
1758 | |
1759 | /* get output buffer */ |
1760 | frame->nb_samples = MAX_SFRAMESIZE; |
1761 | if ((res = ff_get_buffer(ctx, frame, 0)) < 0) |
1762 | return res; |
1763 | frame->nb_samples = n_samples; |
1764 | samples = (float *)frame->data[0]; |
1765 | |
1766 | /* Parse frames, optionally preceded by per-frame (independent) LSPs. */ |
1767 | for (n = 0; n < 3; n++) { |
1768 | if (!s->has_residual_lsps) { |
1769 | int m; |
1770 | |
1771 | if (s->lsps == 10) { |
1772 | dequant_lsp10i(gb, lsps[n]); |
1773 | } else /* s->lsps == 16 */ |
1774 | dequant_lsp16i(gb, lsps[n]); |
1775 | |
1776 | for (m = 0; m < s->lsps; m++) |
1777 | lsps[n][m] += mean_lsf[m]; |
1778 | stabilize_lsps(lsps[n], s->lsps); |
1779 | } |
1780 | |
1781 | if ((res = synth_frame(ctx, gb, n, |
1782 | &samples[n * MAX_FRAMESIZE], |
1783 | lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1], |
1784 | &excitation[s->history_nsamples + n * MAX_FRAMESIZE], |
1785 | &synth[s->lsps + n * MAX_FRAMESIZE]))) { |
1786 | *got_frame_ptr = 0; |
1787 | return res; |
1788 | } |
1789 | } |
1790 | |
1791 | /* Statistics? FIXME - we don't check for length, a slight overrun |
1792 | * will be caught by internal buffer padding, and anything else |
1793 | * will be skipped, not read. */ |
1794 | if (get_bits1(gb)) { |
1795 | res = get_bits(gb, 4); |
1796 | skip_bits(gb, 10 * (res + 1)); |
1797 | } |
1798 | |
1799 | if (get_bits_left(gb) < 0) { |
1800 | wmavoice_flush(ctx); |
1801 | return AVERROR_INVALIDDATA; |
1802 | } |
1803 | |
1804 | *got_frame_ptr = 1; |
1805 | |
1806 | /* Update history */ |
1807 | memcpy(s->prev_lsps, lsps[2], |
1808 | s->lsps * sizeof(*s->prev_lsps)); |
1809 | memcpy(s->synth_history, &synth[MAX_SFRAMESIZE], |
1810 | s->lsps * sizeof(*synth)); |
1811 | memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE], |
1812 | s->history_nsamples * sizeof(*excitation)); |
1813 | if (s->do_apf) |
1814 | memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE], |
1815 | s->history_nsamples * sizeof(*s->zero_exc_pf)); |
1816 | |
1817 | return 0; |
1818 | } |
1819 | |
1820 | /** |
1821 | * Parse the packet header at the start of each packet (input data to this |
1822 | * decoder). |
1823 | * |
1824 | * @param s WMA Voice decoding context private data |
1825 | * @return <0 on error, nb_superframes on success. |
1826 | */ |
1827 | static int parse_packet_header(WMAVoiceContext *s) |
1828 | { |
1829 | GetBitContext *gb = &s->gb; |
1830 | unsigned int res, n_superframes = 0; |
1831 | |
1832 | skip_bits(gb, 4); // packet sequence number |
1833 | s->has_residual_lsps = get_bits1(gb); |
1834 | do { |
1835 | res = get_bits(gb, 6); // number of superframes per packet |
1836 | // (minus first one if there is spillover) |
1837 | n_superframes += res; |
1838 | } while (res == 0x3F); |
1839 | s->spillover_nbits = get_bits(gb, s->spillover_bitsize); |
1840 | |
1841 | return get_bits_left(gb) >= 0 ? n_superframes : AVERROR_INVALIDDATA; |
1842 | } |
1843 | |
1844 | /** |
1845 | * Copy (unaligned) bits from gb/data/size to pb. |
1846 | * |
1847 | * @param pb target buffer to copy bits into |
1848 | * @param data source buffer to copy bits from |
1849 | * @param size size of the source data, in bytes |
1850 | * @param gb bit I/O context specifying the current position in the source. |
1851 | * data. This function might use this to align the bit position to |
1852 | * a whole-byte boundary before calling #avpriv_copy_bits() on aligned |
1853 | * source data |
1854 | * @param nbits the amount of bits to copy from source to target |
1855 | * |
1856 | * @note after calling this function, the current position in the input bit |
1857 | * I/O context is undefined. |
1858 | */ |
1859 | static void copy_bits(PutBitContext *pb, |
1860 | const uint8_t *data, int size, |
1861 | GetBitContext *gb, int nbits) |
1862 | { |
1863 | int rmn_bytes, rmn_bits; |
1864 | |
1865 | rmn_bits = rmn_bytes = get_bits_left(gb); |
1866 | if (rmn_bits < nbits) |
1867 | return; |
1868 | if (nbits > pb->size_in_bits - put_bits_count(pb)) |
1869 | return; |
1870 | rmn_bits &= 7; rmn_bytes >>= 3; |
1871 | if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0) |
1872 | put_bits(pb, rmn_bits, get_bits(gb, rmn_bits)); |
1873 | avpriv_copy_bits(pb, data + size - rmn_bytes, |
1874 | FFMIN(nbits - rmn_bits, rmn_bytes << 3)); |
1875 | } |
1876 | |
1877 | /** |
1878 | * Packet decoding: a packet is anything that the (ASF) demuxer contains, |
1879 | * and we expect that the demuxer / application provides it to us as such |
1880 | * (else you'll probably get garbage as output). Every packet has a size of |
1881 | * ctx->block_align bytes, starts with a packet header (see |
1882 | * #parse_packet_header()), and then a series of superframes. Superframe |
1883 | * boundaries may exceed packets, i.e. superframes can split data over |
1884 | * multiple (two) packets. |
1885 | * |
1886 | * For more information about frames, see #synth_superframe(). |
1887 | */ |
1888 | static int wmavoice_decode_packet(AVCodecContext *ctx, void *data, |
1889 | int *got_frame_ptr, AVPacket *avpkt) |
1890 | { |
1891 | WMAVoiceContext *s = ctx->priv_data; |
1892 | GetBitContext *gb = &s->gb; |
1893 | int size, res, pos; |
1894 | |
1895 | /* Packets are sometimes a multiple of ctx->block_align, with a packet |
1896 | * header at each ctx->block_align bytes. However, FFmpeg's ASF demuxer |
1897 | * feeds us ASF packets, which may concatenate multiple "codec" packets |
1898 | * in a single "muxer" packet, so we artificially emulate that by |
1899 | * capping the packet size at ctx->block_align. */ |
1900 | for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align); |
1901 | init_get_bits(&s->gb, avpkt->data, size << 3); |
1902 | |
1903 | /* size == ctx->block_align is used to indicate whether we are dealing with |
1904 | * a new packet or a packet of which we already read the packet header |
1905 | * previously. */ |
1906 | if (!(size % ctx->block_align)) { // new packet header |
1907 | if (!size) { |
1908 | s->spillover_nbits = 0; |
1909 | s->nb_superframes = 0; |
1910 | } else { |
1911 | if ((res = parse_packet_header(s)) < 0) |
1912 | return res; |
1913 | s->nb_superframes = res; |
1914 | } |
1915 | |
1916 | /* If the packet header specifies a s->spillover_nbits, then we want |
1917 | * to push out all data of the previous packet (+ spillover) before |
1918 | * continuing to parse new superframes in the current packet. */ |
1919 | if (s->sframe_cache_size > 0) { |
1920 | int cnt = get_bits_count(gb); |
1921 | if (cnt + s->spillover_nbits > avpkt->size * 8) { |
1922 | s->spillover_nbits = avpkt->size * 8 - cnt; |
1923 | } |
1924 | copy_bits(&s->pb, avpkt->data, size, gb, s->spillover_nbits); |
1925 | flush_put_bits(&s->pb); |
1926 | s->sframe_cache_size += s->spillover_nbits; |
1927 | if ((res = synth_superframe(ctx, data, got_frame_ptr)) == 0 && |
1928 | *got_frame_ptr) { |
1929 | cnt += s->spillover_nbits; |
1930 | s->skip_bits_next = cnt & 7; |
1931 | res = cnt >> 3; |
1932 | return res; |
1933 | } else |
1934 | skip_bits_long (gb, s->spillover_nbits - cnt + |
1935 | get_bits_count(gb)); // resync |
1936 | } else if (s->spillover_nbits) { |
1937 | skip_bits_long(gb, s->spillover_nbits); // resync |
1938 | } |
1939 | } else if (s->skip_bits_next) |
1940 | skip_bits(gb, s->skip_bits_next); |
1941 | |
1942 | /* Try parsing superframes in current packet */ |
1943 | s->sframe_cache_size = 0; |
1944 | s->skip_bits_next = 0; |
1945 | pos = get_bits_left(gb); |
1946 | if (s->nb_superframes-- == 0) { |
1947 | *got_frame_ptr = 0; |
1948 | return size; |
1949 | } else if (s->nb_superframes > 0) { |
1950 | if ((res = synth_superframe(ctx, data, got_frame_ptr)) < 0) { |
1951 | return res; |
1952 | } else if (*got_frame_ptr) { |
1953 | int cnt = get_bits_count(gb); |
1954 | s->skip_bits_next = cnt & 7; |
1955 | res = cnt >> 3; |
1956 | return res; |
1957 | } |
1958 | } else if ((s->sframe_cache_size = pos) > 0) { |
1959 | /* ... cache it for spillover in next packet */ |
1960 | init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE); |
1961 | copy_bits(&s->pb, avpkt->data, size, gb, s->sframe_cache_size); |
1962 | // FIXME bad - just copy bytes as whole and add use the |
1963 | // skip_bits_next field |
1964 | } |
1965 | |
1966 | return size; |
1967 | } |
1968 | |
1969 | static av_cold int wmavoice_decode_end(AVCodecContext *ctx) |
1970 | { |
1971 | WMAVoiceContext *s = ctx->priv_data; |
1972 | |
1973 | if (s->do_apf) { |
1974 | ff_rdft_end(&s->rdft); |
1975 | ff_rdft_end(&s->irdft); |
1976 | ff_dct_end(&s->dct); |
1977 | ff_dct_end(&s->dst); |
1978 | } |
1979 | |
1980 | return 0; |
1981 | } |
1982 | |
1983 | AVCodec ff_wmavoice_decoder = { |
1984 | .name = "wmavoice", |
1985 | .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"), |
1986 | .type = AVMEDIA_TYPE_AUDIO, |
1987 | .id = AV_CODEC_ID_WMAVOICE, |
1988 | .priv_data_size = sizeof(WMAVoiceContext), |
1989 | .init = wmavoice_decode_init, |
1990 | .init_static_data = wmavoice_init_static_data, |
1991 | .close = wmavoice_decode_end, |
1992 | .decode = wmavoice_decode_packet, |
1993 | .capabilities = AV_CODEC_CAP_SUBFRAMES | AV_CODEC_CAP_DR1 | AV_CODEC_CAP_DELAY, |
1994 | .flush = wmavoice_flush, |
1995 | }; |
1996 |