diff options
author | Sean Bartell <wingedtachikoma@gmail.com> | 2011-06-25 21:32:25 -0400 |
---|---|---|
committer | Nils Wallménius <nils@rockbox.org> | 2012-04-25 22:13:20 +0200 |
commit | f40bfc9267b13b54e6379dfe7539447662879d24 (patch) | |
tree | 9b20069d5e62809ff434061ad730096836f916f2 /lib/rbcodec/codecs/libwmavoice/wmavoice.c | |
parent | a0009907de7a0107d49040d8a180f140e2eff299 (diff) | |
download | rockbox-f40bfc9267b13b54e6379dfe7539447662879d24.tar.gz rockbox-f40bfc9267b13b54e6379dfe7539447662879d24.zip |
Add codecs to librbcodec.
Change-Id: Id7f4717d51ed02d67cb9f9cb3c0ada4a81843f97
Reviewed-on: http://gerrit.rockbox.org/137
Reviewed-by: Nils Wallménius <nils@rockbox.org>
Tested-by: Nils Wallménius <nils@rockbox.org>
Diffstat (limited to 'lib/rbcodec/codecs/libwmavoice/wmavoice.c')
-rw-r--r-- | lib/rbcodec/codecs/libwmavoice/wmavoice.c | 2060 |
1 files changed, 2060 insertions, 0 deletions
diff --git a/lib/rbcodec/codecs/libwmavoice/wmavoice.c b/lib/rbcodec/codecs/libwmavoice/wmavoice.c new file mode 100644 index 0000000000..4d31334c36 --- /dev/null +++ b/lib/rbcodec/codecs/libwmavoice/wmavoice.c | |||
@@ -0,0 +1,2060 @@ | |||
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 | #include "wmavoice.h" | ||
30 | #include "get_bits.h" | ||
31 | #include "put_bits.h" | ||
32 | #include "wmavoice_data.h" | ||
33 | #include "celp_math.h" | ||
34 | #include "celp_filters.h" | ||
35 | #include "acelp_vectors.h" | ||
36 | #include "acelp_filters.h" | ||
37 | #include "lsp.h" | ||
38 | #include "libavutil/lzo.h" | ||
39 | #include "avfft.h" | ||
40 | #include "fft.h" | ||
41 | |||
42 | #define MAX_BLOCKS 8 ///< maximum number of blocks per frame | ||
43 | #define MAX_LSPS 16 ///< maximum filter order | ||
44 | #define MAX_LSPS_ALIGN16 16 ///< same as #MAX_LSPS; needs to be multiple | ||
45 | ///< of 16 for ASM input buffer alignment | ||
46 | #define MAX_FRAMES 3 ///< maximum number of frames per superframe | ||
47 | #define MAX_FRAMESIZE 160 ///< maximum number of samples per frame | ||
48 | #define MAX_SIGNAL_HISTORY 416 ///< maximum excitation signal history | ||
49 | #define MAX_SFRAMESIZE (MAX_FRAMESIZE * MAX_FRAMES) | ||
50 | ///< maximum number of samples per superframe | ||
51 | #define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that | ||
52 | ///< was split over two packets | ||
53 | #define VLC_NBITS 6 ///< number of bits to read per VLC iteration | ||
54 | |||
55 | /** | ||
56 | * Frame type VLC coding. | ||
57 | */ | ||
58 | static VLC frame_type_vlc; | ||
59 | |||
60 | /** | ||
61 | * Adaptive codebook types. | ||
62 | */ | ||
63 | enum { | ||
64 | ACB_TYPE_NONE = 0, ///< no adaptive codebook (only hardcoded fixed) | ||
65 | ACB_TYPE_ASYMMETRIC = 1, ///< adaptive codebook with per-frame pitch, which | ||
66 | ///< we interpolate to get a per-sample pitch. | ||
67 | ///< Signal is generated using an asymmetric sinc | ||
68 | ///< window function | ||
69 | ///< @note see #wmavoice_ipol1_coeffs | ||
70 | ACB_TYPE_HAMMING = 2 ///< Per-block pitch with signal generation using | ||
71 | ///< a Hamming sinc window function | ||
72 | ///< @note see #wmavoice_ipol2_coeffs | ||
73 | }; | ||
74 | |||
75 | /** | ||
76 | * Fixed codebook types. | ||
77 | */ | ||
78 | enum { | ||
79 | FCB_TYPE_SILENCE = 0, ///< comfort noise during silence | ||
80 | ///< generated from a hardcoded (fixed) codebook | ||
81 | ///< with per-frame (low) gain values | ||
82 | FCB_TYPE_HARDCODED = 1, ///< hardcoded (fixed) codebook with per-block | ||
83 | ///< gain values | ||
84 | FCB_TYPE_AW_PULSES = 2, ///< Pitch-adaptive window (AW) pulse signals, | ||
85 | ///< used in particular for low-bitrate streams | ||
86 | FCB_TYPE_EXC_PULSES = 3, ///< Innovation (fixed) codebook pulse sets in | ||
87 | ///< combinations of either single pulses or | ||
88 | ///< pulse pairs | ||
89 | }; | ||
90 | |||
91 | /** | ||
92 | * Description of frame types. | ||
93 | */ | ||
94 | static const struct frame_type_desc { | ||
95 | uint8_t n_blocks; ///< amount of blocks per frame (each block | ||
96 | ///< (contains 160/#n_blocks samples) | ||
97 | uint8_t log_n_blocks; ///< log2(#n_blocks) | ||
98 | uint8_t acb_type; ///< Adaptive codebook type (ACB_TYPE_*) | ||
99 | uint8_t fcb_type; ///< Fixed codebook type (FCB_TYPE_*) | ||
100 | uint8_t dbl_pulses; ///< how many pulse vectors have pulse pairs | ||
101 | ///< (rather than just one single pulse) | ||
102 | ///< only if #fcb_type == #FCB_TYPE_EXC_PULSES | ||
103 | uint16_t frame_size; ///< the amount of bits that make up the block | ||
104 | ///< data (per frame) | ||
105 | } frame_descs[17] = { | ||
106 | { 1, 0, ACB_TYPE_NONE, FCB_TYPE_SILENCE, 0, 0 }, | ||
107 | { 2, 1, ACB_TYPE_NONE, FCB_TYPE_HARDCODED, 0, 28 }, | ||
108 | { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_AW_PULSES, 0, 46 }, | ||
109 | { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 80 }, | ||
110 | { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 104 }, | ||
111 | { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 0, 108 }, | ||
112 | { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 132 }, | ||
113 | { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 168 }, | ||
114 | { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 64 }, | ||
115 | { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 80 }, | ||
116 | { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 104 }, | ||
117 | { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 108 }, | ||
118 | { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 132 }, | ||
119 | { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 168 }, | ||
120 | { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 176 }, | ||
121 | { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 208 }, | ||
122 | { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 256 } | ||
123 | }; | ||
124 | |||
125 | /** | ||
126 | * WMA Voice decoding context. | ||
127 | */ | ||
128 | typedef struct { | ||
129 | /** | ||
130 | * @defgroup struct_global Global values | ||
131 | * Global values, specified in the stream header / extradata or used | ||
132 | * 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 | int frame_lsp_bitsize; ///< size (in bits) of LSPs, when encoded | ||
162 | ///< per-frame (independent coding) | ||
163 | int sframe_lsp_bitsize; ///< size (in bits) of LSPs, when encoded | ||
164 | ///< per superframe (residual coding) | ||
165 | |||
166 | int min_pitch_val; ///< base value for pitch parsing code | ||
167 | int max_pitch_val; ///< max value + 1 for pitch parsing | ||
168 | int pitch_nbits; ///< number of bits used to specify the | ||
169 | ///< pitch value in the frame header | ||
170 | int block_pitch_nbits; ///< number of bits used to specify the | ||
171 | ///< first block's pitch value | ||
172 | int block_pitch_range; ///< range of the block pitch | ||
173 | int block_delta_pitch_nbits; ///< number of bits used to specify the | ||
174 | ///< delta pitch between this and the last | ||
175 | ///< block's pitch value, used in all but | ||
176 | ///< first block | ||
177 | int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is | ||
178 | ///< from -this to +this-1) | ||
179 | uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale | ||
180 | ///< conversion | ||
181 | |||
182 | /** | ||
183 | * @} | ||
184 | * @defgroup struct_packet Packet values | ||
185 | * Packet values, specified in the packet header or related to a packet. | ||
186 | * A packet is considered to be a single unit of data provided to this | ||
187 | * decoder by the demuxer. | ||
188 | * @{ | ||
189 | */ | ||
190 | int spillover_nbits; ///< number of bits of the previous packet's | ||
191 | ///< last superframe preceeding this | ||
192 | ///< packet's first full superframe (useful | ||
193 | ///< for re-synchronization also) | ||
194 | int has_residual_lsps; ///< if set, superframes contain one set of | ||
195 | ///< LSPs that cover all frames, encoded as | ||
196 | ///< independent and residual LSPs; if not | ||
197 | ///< set, each frame contains its own, fully | ||
198 | ///< independent, LSPs | ||
199 | int skip_bits_next; ///< number of bits to skip at the next call | ||
200 | ///< to #wmavoice_decode_packet() (since | ||
201 | ///< they're part of the previous superframe) | ||
202 | |||
203 | uint8_t sframe_cache[SFRAME_CACHE_MAXSIZE + FF_INPUT_BUFFER_PADDING_SIZE]; | ||
204 | ///< cache for superframe data split over | ||
205 | ///< multiple packets | ||
206 | int sframe_cache_size; ///< set to >0 if we have data from an | ||
207 | ///< (incomplete) superframe from a previous | ||
208 | ///< packet that spilled over in the current | ||
209 | ///< packet; specifies the amount of bits in | ||
210 | ///< #sframe_cache | ||
211 | PutBitContext pb; ///< bitstream writer for #sframe_cache | ||
212 | |||
213 | /** | ||
214 | * @} | ||
215 | * @defgroup struct_frame 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 | float gain_pred_err[6]; ///< cache for gain prediction | ||
251 | float excitation_history[MAX_SIGNAL_HISTORY]; | ||
252 | ///< cache of the signal of previous | ||
253 | ///< superframes, used as a history for | ||
254 | ///< signal generation | ||
255 | float synth_history[MAX_LSPS]; ///< see #excitation_history | ||
256 | /** | ||
257 | * @} | ||
258 | * @defgroup post_filter Postfilter values | ||
259 | * Varibales used for postfilter implementation, mostly history for | ||
260 | * smoothing and so on, and context variables for FFT/iFFT. | ||
261 | * @{ | ||
262 | */ | ||
263 | RDFTContext rdft, irdft; ///< contexts for FFT-calculation in the | ||
264 | ///< postfilter (for denoise filter) | ||
265 | DCTContext dct, dst; ///< contexts for phase shift (in Hilbert | ||
266 | ///< transform, part of postfilter) | ||
267 | float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi] | ||
268 | ///< range | ||
269 | float postfilter_agc; ///< gain control memory, used in | ||
270 | ///< #adaptive_gain_control() | ||
271 | float dcf_mem[2]; ///< DC filter history | ||
272 | float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE]; | ||
273 | ///< zero filter output (i.e. excitation) | ||
274 | ///< by postfilter | ||
275 | float denoise_filter_cache[MAX_FRAMESIZE]; | ||
276 | int denoise_filter_cache_size; ///< samples in #denoise_filter_cache | ||
277 | DECLARE_ALIGNED(16, float, tilted_lpcs_pf)[0x80]; | ||
278 | ///< aligned buffer for LPC tilting | ||
279 | DECLARE_ALIGNED(16, float, denoise_coeffs_pf)[0x80]; | ||
280 | ///< aligned buffer for denoise coefficients | ||
281 | DECLARE_ALIGNED(16, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16]; | ||
282 | ///< aligned buffer for postfilter speech | ||
283 | ///< synthesis | ||
284 | /** | ||
285 | * @} | ||
286 | */ | ||
287 | } WMAVoiceContext; | ||
288 | |||
289 | /* global decode context */ | ||
290 | static WMAVoiceContext globWMAVoiceCtx; | ||
291 | |||
292 | |||
293 | /** | ||
294 | * Set up the variable bit mode (VBM) tree from container extradata. | ||
295 | * @param gb bit I/O context. | ||
296 | * The bit context (s->gb) should be loaded with byte 23-46 of the | ||
297 | * container extradata (i.e. the ones containing the VBM tree). | ||
298 | * @param vbm_tree pointer to array to which the decoded VBM tree will be | ||
299 | * written. | ||
300 | * @return 0 on success, <0 on error. | ||
301 | */ | ||
302 | static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25]) | ||
303 | { | ||
304 | static const uint8_t bits[] = { | ||
305 | 2, 2, 2, 4, 4, 4, | ||
306 | 6, 6, 6, 8, 8, 8, | ||
307 | 10, 10, 10, 12, 12, 12, | ||
308 | 14, 14, 14, 14 | ||
309 | }; | ||
310 | static const uint16_t codes[] = { | ||
311 | 0x0000, 0x0001, 0x0002, // 00/01/10 | ||
312 | 0x000c, 0x000d, 0x000e, // 11+00/01/10 | ||
313 | 0x003c, 0x003d, 0x003e, // 1111+00/01/10 | ||
314 | 0x00fc, 0x00fd, 0x00fe, // 111111+00/01/10 | ||
315 | 0x03fc, 0x03fd, 0x03fe, // 11111111+00/01/10 | ||
316 | 0x0ffc, 0x0ffd, 0x0ffe, // 1111111111+00/01/10 | ||
317 | 0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx | ||
318 | }; | ||
319 | int cntr[8], n, res; | ||
320 | |||
321 | memset(vbm_tree, 0xff, sizeof(int8_t) * 25); | ||
322 | memset(cntr, 0, sizeof(cntr)); | ||
323 | for (n = 0; n < 17; n++) { | ||
324 | res = get_bits(gb, 3); | ||
325 | if (cntr[res] > 3) // should be >= 3 + (res == 7)) | ||
326 | return -1; | ||
327 | vbm_tree[res * 3 + cntr[res]++] = n; | ||
328 | } | ||
329 | INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits), | ||
330 | bits, 1, 1, codes, 2, 2, 132); | ||
331 | return 0; | ||
332 | } | ||
333 | |||
334 | /** | ||
335 | * Set up decoder with parameters from demuxer (extradata etc.). | ||
336 | */ | ||
337 | av_cold int wmavoice_decode_init(AVCodecContext *ctx) | ||
338 | { | ||
339 | int n, flags, pitch_range, lsp16_flag; | ||
340 | ctx->priv_data = &globWMAVoiceCtx; | ||
341 | WMAVoiceContext *s = ctx->priv_data; | ||
342 | |||
343 | /** | ||
344 | * Extradata layout: | ||
345 | * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c), | ||
346 | * - byte 19-22: flags field (annoyingly in LE; see below for known | ||
347 | * values), | ||
348 | * - byte 23-46: variable bitmode tree (really just 17 * 3 bits, | ||
349 | * rest is 0). | ||
350 | */ | ||
351 | if (ctx->extradata_size != 46) { | ||
352 | av_log(ctx, AV_LOG_ERROR, | ||
353 | "Invalid extradata size %d (should be 46)\n", | ||
354 | ctx->extradata_size); | ||
355 | return -1; | ||
356 | } | ||
357 | flags = AV_RL32(ctx->extradata + 18); | ||
358 | s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align); | ||
359 | s->do_apf = flags & 0x1; | ||
360 | if (s->do_apf) { | ||
361 | ff_rdft_init(&s->rdft, 7, DFT_R2C); | ||
362 | ff_rdft_init(&s->irdft, 7, IDFT_C2R); | ||
363 | ff_dct_init(&s->dct, 6, DCT_I); | ||
364 | ff_dct_init(&s->dst, 6, DST_I); | ||
365 | |||
366 | ff_sine_window_init(s->cos, 256); | ||
367 | memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0])); | ||
368 | for (n = 0; n < 255; n++) { | ||
369 | s->sin[n] = -s->sin[510 - n]; | ||
370 | s->cos[510 - n] = s->cos[n]; | ||
371 | } | ||
372 | } | ||
373 | s->denoise_strength = (flags >> 2) & 0xF; | ||
374 | if (s->denoise_strength >= 12) { | ||
375 | av_log(ctx, AV_LOG_ERROR, | ||
376 | "Invalid denoise filter strength %d (max=11)\n", | ||
377 | s->denoise_strength); | ||
378 | return -1; | ||
379 | } | ||
380 | s->denoise_tilt_corr = !!(flags & 0x40); | ||
381 | s->dc_level = (flags >> 7) & 0xF; | ||
382 | s->lsp_q_mode = !!(flags & 0x2000); | ||
383 | s->lsp_def_mode = !!(flags & 0x4000); | ||
384 | lsp16_flag = flags & 0x1000; | ||
385 | if (lsp16_flag) { | ||
386 | s->lsps = 16; | ||
387 | s->frame_lsp_bitsize = 34; | ||
388 | s->sframe_lsp_bitsize = 60; | ||
389 | } else { | ||
390 | s->lsps = 10; | ||
391 | s->frame_lsp_bitsize = 24; | ||
392 | s->sframe_lsp_bitsize = 48; | ||
393 | } | ||
394 | for (n = 0; n < s->lsps; n++) | ||
395 | s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0); | ||
396 | |||
397 | init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3); | ||
398 | if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) { | ||
399 | av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n"); | ||
400 | return -1; | ||
401 | } | ||
402 | |||
403 | s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8; | ||
404 | s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8; | ||
405 | pitch_range = s->max_pitch_val - s->min_pitch_val; | ||
406 | s->pitch_nbits = av_ceil_log2(pitch_range); | ||
407 | s->last_pitch_val = 40; | ||
408 | s->last_acb_type = ACB_TYPE_NONE; | ||
409 | s->history_nsamples = s->max_pitch_val + 8; | ||
410 | |||
411 | if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) { | ||
412 | int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8, | ||
413 | max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8; | ||
414 | |||
415 | av_log(ctx, AV_LOG_ERROR, | ||
416 | "Unsupported samplerate %d (min=%d, max=%d)\n", | ||
417 | ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz | ||
418 | |||
419 | return -1; | ||
420 | } | ||
421 | |||
422 | s->block_conv_table[0] = s->min_pitch_val; | ||
423 | s->block_conv_table[1] = (pitch_range * 25) >> 6; | ||
424 | s->block_conv_table[2] = (pitch_range * 44) >> 6; | ||
425 | s->block_conv_table[3] = s->max_pitch_val - 1; | ||
426 | s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF; | ||
427 | s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange); | ||
428 | s->block_pitch_range = s->block_conv_table[2] + | ||
429 | s->block_conv_table[3] + 1 + | ||
430 | 2 * (s->block_conv_table[1] - 2 * s->min_pitch_val); | ||
431 | s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range); | ||
432 | |||
433 | ctx->sample_fmt = SAMPLE_FMT_FLT; | ||
434 | |||
435 | return 0; | ||
436 | } | ||
437 | |||
438 | /** | ||
439 | * @defgroup postfilter Postfilter functions | ||
440 | * Postfilter functions (gain control, wiener denoise filter, DC filter, | ||
441 | * kalman smoothening, plus surrounding code to wrap it) | ||
442 | * @{ | ||
443 | */ | ||
444 | /** | ||
445 | * Adaptive gain control (as used in postfilter). | ||
446 | * | ||
447 | * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except | ||
448 | * that the energy here is calculated using sum(abs(...)), whereas the | ||
449 | * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)). | ||
450 | * | ||
451 | * @param out output buffer for filtered samples | ||
452 | * @param in input buffer containing the samples as they are after the | ||
453 | * postfilter steps so far | ||
454 | * @param speech_synth input buffer containing speech synth before postfilter | ||
455 | * @param size input buffer size | ||
456 | * @param alpha exponential filter factor | ||
457 | * @param gain_mem pointer to filter memory (single float) | ||
458 | */ | ||
459 | static void adaptive_gain_control(float *out, const float *in, | ||
460 | const float *speech_synth, | ||
461 | int size, float alpha, float *gain_mem) | ||
462 | { | ||
463 | int i; | ||
464 | float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor; | ||
465 | float mem = *gain_mem; | ||
466 | |||
467 | for (i = 0; i < size; i++) { | ||
468 | speech_energy += fabsf(speech_synth[i]); | ||
469 | postfilter_energy += fabsf(in[i]); | ||
470 | } | ||
471 | gain_scale_factor = (1.0 - alpha) * speech_energy / postfilter_energy; | ||
472 | |||
473 | for (i = 0; i < size; i++) { | ||
474 | mem = alpha * mem + gain_scale_factor; | ||
475 | out[i] = in[i] * mem; | ||
476 | } | ||
477 | |||
478 | *gain_mem = mem; | ||
479 | } | ||
480 | |||
481 | /** | ||
482 | * Kalman smoothing function. | ||
483 | * | ||
484 | * This function looks back pitch +/- 3 samples back into history to find | ||
485 | * the best fitting curve (that one giving the optimal gain of the two | ||
486 | * signals, i.e. the highest dot product between the two), and then | ||
487 | * uses that signal history to smoothen the output of the speech synthesis | ||
488 | * filter. | ||
489 | * | ||
490 | * @param s WMA Voice decoding context | ||
491 | * @param pitch pitch of the speech signal | ||
492 | * @param in input speech signal | ||
493 | * @param out output pointer for smoothened signal | ||
494 | * @param size input/output buffer size | ||
495 | * | ||
496 | * @returns -1 if no smoothening took place, e.g. because no optimal | ||
497 | * fit could be found, or 0 on success. | ||
498 | */ | ||
499 | static int kalman_smoothen(WMAVoiceContext *s, int pitch, | ||
500 | const float *in, float *out, int size) | ||
501 | { | ||
502 | int n; | ||
503 | float optimal_gain = 0, dot; | ||
504 | const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)], | ||
505 | *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)], | ||
506 | *best_hist_ptr; | ||
507 | |||
508 | /* find best fitting point in history */ | ||
509 | do { | ||
510 | dot = ff_dot_productf(in, ptr, size); | ||
511 | if (dot > optimal_gain) { | ||
512 | optimal_gain = dot; | ||
513 | best_hist_ptr = ptr; | ||
514 | } | ||
515 | } while (--ptr >= end); | ||
516 | |||
517 | if (optimal_gain <= 0) | ||
518 | return -1; | ||
519 | dot = ff_dot_productf(best_hist_ptr, best_hist_ptr, size); | ||
520 | if (dot <= 0) // would be 1.0 | ||
521 | return -1; | ||
522 | |||
523 | if (optimal_gain <= dot) { | ||
524 | dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000 | ||
525 | } else | ||
526 | dot = 0.625; | ||
527 | |||
528 | /* actual smoothing */ | ||
529 | for (n = 0; n < size; n++) | ||
530 | out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]); | ||
531 | |||
532 | return 0; | ||
533 | } | ||
534 | |||
535 | /** | ||
536 | * Get the tilt factor of a formant filter from its transfer function | ||
537 | * @see #tilt_factor() in amrnbdec.c, which does essentially the same, | ||
538 | * but somehow (??) it does a speech synthesis filter in the | ||
539 | * middle, which is missing here | ||
540 | * | ||
541 | * @param lpcs LPC coefficients | ||
542 | * @param n_lpcs Size of LPC buffer | ||
543 | * @returns the tilt factor | ||
544 | */ | ||
545 | static float tilt_factor(const float *lpcs, int n_lpcs) | ||
546 | { | ||
547 | float rh0, rh1; | ||
548 | |||
549 | rh0 = 1.0 + ff_dot_productf(lpcs, lpcs, n_lpcs); | ||
550 | rh1 = lpcs[0] + ff_dot_productf(lpcs, &lpcs[1], n_lpcs - 1); | ||
551 | |||
552 | return rh1 / rh0; | ||
553 | } | ||
554 | |||
555 | /** | ||
556 | * Derive denoise filter coefficients (in real domain) from the LPCs. | ||
557 | */ | ||
558 | static void calc_input_response(WMAVoiceContext *s, float *lpcs, | ||
559 | int fcb_type, float *coeffs, int remainder) | ||
560 | { | ||
561 | float last_coeff, min = 15.0, max = -15.0; | ||
562 | float irange, angle_mul, gain_mul, range, sq; | ||
563 | int n, idx; | ||
564 | |||
565 | /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */ | ||
566 | ff_rdft_calc(&s->rdft, lpcs); | ||
567 | #define log_range(var, assign) do { \ | ||
568 | float tmp = log10f(assign); var = tmp; \ | ||
569 | max = FFMAX(max, tmp); min = FFMIN(min, tmp); \ | ||
570 | } while (0) | ||
571 | log_range(last_coeff, lpcs[1] * lpcs[1]); | ||
572 | for (n = 1; n < 64; n++) | ||
573 | log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] + | ||
574 | lpcs[n * 2 + 1] * lpcs[n * 2 + 1]); | ||
575 | log_range(lpcs[0], lpcs[0] * lpcs[0]); | ||
576 | #undef log_range | ||
577 | range = max - min; | ||
578 | lpcs[64] = last_coeff; | ||
579 | |||
580 | /* Now, use this spectrum to pick out these frequencies with higher | ||
581 | * (relative) power/energy (which we then take to be "not noise"), | ||
582 | * and set up a table (still in lpc[]) of (relative) gains per frequency. | ||
583 | * These frequencies will be maintained, while others ("noise") will be | ||
584 | * decreased in the filter output. */ | ||
585 | irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63] | ||
586 | gain_mul = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) : | ||
587 | (5.0 / 14.7)); | ||
588 | angle_mul = gain_mul * (8.0 * M_LN10 / M_PI); | ||
589 | for (n = 0; n <= 64; n++) { | ||
590 | float pwr; | ||
591 | |||
592 | idx = FFMAX(0, lrint((max - lpcs[n]) * irange) - 1); | ||
593 | pwr = wmavoice_denoise_power_table[s->denoise_strength][idx]; | ||
594 | lpcs[n] = angle_mul * pwr; | ||
595 | |||
596 | /* 70.57 =~ 1/log10(1.0331663) */ | ||
597 | idx = (pwr * gain_mul - 0.0295) * 70.570526123; | ||
598 | if (idx > 127) { // fallback if index falls outside table range | ||
599 | coeffs[n] = wmavoice_energy_table[127] * | ||
600 | powf(1.0331663, idx - 127); | ||
601 | } else | ||
602 | coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)]; | ||
603 | } | ||
604 | |||
605 | /* calculate the Hilbert transform of the gains, which we do (since this | ||
606 | * is a sinus input) by doing a phase shift (in theory, H(sin())=cos()). | ||
607 | * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the | ||
608 | * "moment" of the LPCs in this filter. */ | ||
609 | ff_dct_calc(&s->dct, lpcs); | ||
610 | ff_dct_calc(&s->dst, lpcs); | ||
611 | |||
612 | /* Split out the coefficient indexes into phase/magnitude pairs */ | ||
613 | idx = 255 + av_clip(lpcs[64], -255, 255); | ||
614 | coeffs[0] = coeffs[0] * s->cos[idx]; | ||
615 | idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255); | ||
616 | last_coeff = coeffs[64] * s->cos[idx]; | ||
617 | for (n = 63;; n--) { | ||
618 | idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255); | ||
619 | coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx]; | ||
620 | coeffs[n * 2] = coeffs[n] * s->cos[idx]; | ||
621 | |||
622 | if (!--n) break; | ||
623 | |||
624 | idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255); | ||
625 | coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx]; | ||
626 | coeffs[n * 2] = coeffs[n] * s->cos[idx]; | ||
627 | } | ||
628 | coeffs[1] = last_coeff; | ||
629 | |||
630 | /* move into real domain */ | ||
631 | ff_rdft_calc(&s->irdft, coeffs); | ||
632 | |||
633 | /* tilt correction and normalize scale */ | ||
634 | memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder)); | ||
635 | if (s->denoise_tilt_corr) { | ||
636 | float tilt_mem = 0; | ||
637 | |||
638 | coeffs[remainder - 1] = 0; | ||
639 | ff_tilt_compensation(&tilt_mem, | ||
640 | -1.8 * tilt_factor(coeffs, remainder - 1), | ||
641 | coeffs, remainder); | ||
642 | } | ||
643 | sq = (1.0 / 64.0) * sqrtf(1 / ff_dot_productf(coeffs, coeffs, remainder)); | ||
644 | for (n = 0; n < remainder; n++) | ||
645 | coeffs[n] *= sq; | ||
646 | } | ||
647 | |||
648 | /** | ||
649 | * This function applies a Wiener filter on the (noisy) speech signal as | ||
650 | * a means to denoise it. | ||
651 | * | ||
652 | * - take RDFT of LPCs to get the power spectrum of the noise + speech; | ||
653 | * - using this power spectrum, calculate (for each frequency) the Wiener | ||
654 | * filter gain, which depends on the frequency power and desired level | ||
655 | * of noise subtraction (when set too high, this leads to artifacts) | ||
656 | * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse | ||
657 | * of 4-8kHz); | ||
658 | * - by doing a phase shift, calculate the Hilbert transform of this array | ||
659 | * of per-frequency filter-gains to get the filtering coefficients; | ||
660 | * - smoothen/normalize/de-tilt these filter coefficients as desired; | ||
661 | * - take RDFT of noisy sound, apply the coefficients and take its IRDFT | ||
662 | * to get the denoised speech signal; | ||
663 | * - the leftover (i.e. output of the IRDFT on denoised speech data beyond | ||
664 | * the frame boundary) are saved and applied to subsequent frames by an | ||
665 | * overlap-add method (otherwise you get clicking-artifacts). | ||
666 | * | ||
667 | * @param s WMA Voice decoding context | ||
668 | * @param fcb_type Frame (codebook) type | ||
669 | * @param synth_pf input: the noisy speech signal, output: denoised speech | ||
670 | * data; should be 16-byte aligned (for ASM purposes) | ||
671 | * @param size size of the speech data | ||
672 | * @param lpcs LPCs used to synthesize this frame's speech data | ||
673 | */ | ||
674 | static void wiener_denoise(WMAVoiceContext *s, int fcb_type, | ||
675 | float *synth_pf, int size, | ||
676 | const float *lpcs) | ||
677 | { | ||
678 | int remainder, lim, n; | ||
679 | |||
680 | if (fcb_type != FCB_TYPE_SILENCE) { | ||
681 | float *tilted_lpcs = s->tilted_lpcs_pf, | ||
682 | *coeffs = s->denoise_coeffs_pf, tilt_mem = 0; | ||
683 | |||
684 | tilted_lpcs[0] = 1.0; | ||
685 | memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps); | ||
686 | memset(&tilted_lpcs[s->lsps + 1], 0, | ||
687 | sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1)); | ||
688 | ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps), | ||
689 | tilted_lpcs, s->lsps + 2); | ||
690 | |||
691 | /* The IRDFT output (127 samples for 7-bit filter) beyond the frame | ||
692 | * size is applied to the next frame. All input beyond this is zero, | ||
693 | * and thus all output beyond this will go towards zero, hence we can | ||
694 | * limit to min(size-1, 127-size) as a performance consideration. */ | ||
695 | remainder = FFMIN(127 - size, size - 1); | ||
696 | calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder); | ||
697 | |||
698 | /* apply coefficients (in frequency spectrum domain), i.e. complex | ||
699 | * number multiplication */ | ||
700 | memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size)); | ||
701 | ff_rdft_calc(&s->rdft, synth_pf); | ||
702 | ff_rdft_calc(&s->rdft, coeffs); | ||
703 | synth_pf[0] *= coeffs[0]; | ||
704 | synth_pf[1] *= coeffs[1]; | ||
705 | for (n = 1; n < 64; n++) { | ||
706 | float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1]; | ||
707 | synth_pf[n * 2] = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1]; | ||
708 | synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1]; | ||
709 | } | ||
710 | ff_rdft_calc(&s->irdft, synth_pf); | ||
711 | } | ||
712 | |||
713 | /* merge filter output with the history of previous runs */ | ||
714 | if (s->denoise_filter_cache_size) { | ||
715 | lim = FFMIN(s->denoise_filter_cache_size, size); | ||
716 | for (n = 0; n < lim; n++) | ||
717 | synth_pf[n] += s->denoise_filter_cache[n]; | ||
718 | s->denoise_filter_cache_size -= lim; | ||
719 | memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size], | ||
720 | sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size); | ||
721 | } | ||
722 | |||
723 | /* move remainder of filter output into a cache for future runs */ | ||
724 | if (fcb_type != FCB_TYPE_SILENCE) { | ||
725 | lim = FFMIN(remainder, s->denoise_filter_cache_size); | ||
726 | for (n = 0; n < lim; n++) | ||
727 | s->denoise_filter_cache[n] += synth_pf[size + n]; | ||
728 | if (lim < remainder) { | ||
729 | memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim], | ||
730 | sizeof(s->denoise_filter_cache[0]) * (remainder - lim)); | ||
731 | s->denoise_filter_cache_size = remainder; | ||
732 | } | ||
733 | } | ||
734 | } | ||
735 | |||
736 | /** | ||
737 | * Averaging projection filter, the postfilter used in WMAVoice. | ||
738 | * | ||
739 | * This uses the following steps: | ||
740 | * - A zero-synthesis filter (generate excitation from synth signal) | ||
741 | * - Kalman smoothing on excitation, based on pitch | ||
742 | * - Re-synthesized smoothened output | ||
743 | * - Iterative Wiener denoise filter | ||
744 | * - Adaptive gain filter | ||
745 | * - DC filter | ||
746 | * | ||
747 | * @param s WMAVoice decoding context | ||
748 | * @param synth Speech synthesis output (before postfilter) | ||
749 | * @param samples Output buffer for filtered samples | ||
750 | * @param size Buffer size of synth & samples | ||
751 | * @param lpcs Generated LPCs used for speech synthesis | ||
752 | * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned) | ||
753 | * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses) | ||
754 | * @param pitch Pitch of the input signal | ||
755 | */ | ||
756 | static void postfilter(WMAVoiceContext *s, const float *synth, | ||
757 | float *samples, int size, | ||
758 | const float *lpcs, float *zero_exc_pf, | ||
759 | int fcb_type, int pitch) | ||
760 | { | ||
761 | float synth_filter_in_buf[MAX_FRAMESIZE / 2], | ||
762 | *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16], | ||
763 | *synth_filter_in = zero_exc_pf; | ||
764 | |||
765 | assert(size <= MAX_FRAMESIZE / 2); | ||
766 | |||
767 | /* generate excitation from input signal */ | ||
768 | ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps); | ||
769 | |||
770 | if (fcb_type >= FCB_TYPE_AW_PULSES && | ||
771 | !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size)) | ||
772 | synth_filter_in = synth_filter_in_buf; | ||
773 | |||
774 | /* re-synthesize speech after smoothening, and keep history */ | ||
775 | ff_celp_lp_synthesis_filterf(synth_pf, lpcs, | ||
776 | synth_filter_in, size, s->lsps); | ||
777 | memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps], | ||
778 | sizeof(synth_pf[0]) * s->lsps); | ||
779 | |||
780 | wiener_denoise(s, fcb_type, synth_pf, size, lpcs); | ||
781 | |||
782 | adaptive_gain_control(samples, synth_pf, synth, size, 0.99, | ||
783 | &s->postfilter_agc); | ||
784 | |||
785 | if (s->dc_level > 8) { | ||
786 | /* remove ultra-low frequency DC noise / highpass filter; | ||
787 | * coefficients are identical to those used in SIPR decoding, | ||
788 | * and very closely resemble those used in AMR-NB decoding. */ | ||
789 | ff_acelp_apply_order_2_transfer_function(samples, samples, | ||
790 | (const float[2]) { -1.99997, 1.0 }, | ||
791 | (const float[2]) { -1.9330735188, 0.93589198496 }, | ||
792 | 0.93980580475, s->dcf_mem, size); | ||
793 | } | ||
794 | } | ||
795 | /** | ||
796 | * @} | ||
797 | */ | ||
798 | |||
799 | /** | ||
800 | * Dequantize LSPs | ||
801 | * @param lsps output pointer to the array that will hold the LSPs | ||
802 | * @param num number of LSPs to be dequantized | ||
803 | * @param values quantized values, contains n_stages values | ||
804 | * @param sizes range (i.e. max value) of each quantized value | ||
805 | * @param n_stages number of dequantization runs | ||
806 | * @param table dequantization table to be used | ||
807 | * @param mul_q LSF multiplier | ||
808 | * @param base_q base (lowest) LSF values | ||
809 | */ | ||
810 | static void dequant_lsps(double *lsps, int num, | ||
811 | const uint16_t *values, | ||
812 | const uint16_t *sizes, | ||
813 | int n_stages, const uint8_t *table, | ||
814 | const double *mul_q, | ||
815 | const double *base_q) | ||
816 | { | ||
817 | int n, m; | ||
818 | |||
819 | memset(lsps, 0, num * sizeof(*lsps)); | ||
820 | for (n = 0; n < n_stages; n++) { | ||
821 | const uint8_t *t_off = &table[values[n] * num]; | ||
822 | double base = base_q[n], mul = mul_q[n]; | ||
823 | |||
824 | for (m = 0; m < num; m++) | ||
825 | lsps[m] += base + mul * t_off[m]; | ||
826 | |||
827 | table += sizes[n] * num; | ||
828 | } | ||
829 | } | ||
830 | |||
831 | /** | ||
832 | * @defgroup lsp_dequant LSP dequantization routines | ||
833 | * LSP dequantization routines, for 10/16LSPs and independent/residual coding. | ||
834 | * @note we assume enough bits are available, caller should check. | ||
835 | * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits; | ||
836 | * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits. | ||
837 | * @{ | ||
838 | */ | ||
839 | /** | ||
840 | * Parse 10 independently-coded LSPs. | ||
841 | */ | ||
842 | static void dequant_lsp10i(GetBitContext *gb, double *lsps) | ||
843 | { | ||
844 | static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 }; | ||
845 | static const double mul_lsf[4] = { | ||
846 | 5.2187144800e-3, 1.4626986422e-3, | ||
847 | 9.6179549166e-4, 1.1325736225e-3 | ||
848 | }; | ||
849 | static const double base_lsf[4] = { | ||
850 | M_PI * -2.15522e-1, M_PI * -6.1646e-2, | ||
851 | M_PI * -3.3486e-2, M_PI * -5.7408e-2 | ||
852 | }; | ||
853 | uint16_t v[4]; | ||
854 | |||
855 | v[0] = get_bits(gb, 8); | ||
856 | v[1] = get_bits(gb, 6); | ||
857 | v[2] = get_bits(gb, 5); | ||
858 | v[3] = get_bits(gb, 5); | ||
859 | |||
860 | dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i, | ||
861 | mul_lsf, base_lsf); | ||
862 | } | ||
863 | |||
864 | /** | ||
865 | * Parse 10 independently-coded LSPs, and then derive the tables to | ||
866 | * generate LSPs for the other frames from them (residual coding). | ||
867 | */ | ||
868 | static void dequant_lsp10r(GetBitContext *gb, | ||
869 | double *i_lsps, const double *old, | ||
870 | double *a1, double *a2, int q_mode) | ||
871 | { | ||
872 | static const uint16_t vec_sizes[3] = { 128, 64, 64 }; | ||
873 | static const double mul_lsf[3] = { | ||
874 | 2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3 | ||
875 | }; | ||
876 | static const double base_lsf[3] = { | ||
877 | M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2 | ||
878 | }; | ||
879 | const float (*ipol_tab)[2][10] = q_mode ? | ||
880 | wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a; | ||
881 | uint16_t interpol, v[3]; | ||
882 | int n; | ||
883 | |||
884 | dequant_lsp10i(gb, i_lsps); | ||
885 | |||
886 | interpol = get_bits(gb, 5); | ||
887 | v[0] = get_bits(gb, 7); | ||
888 | v[1] = get_bits(gb, 6); | ||
889 | v[2] = get_bits(gb, 6); | ||
890 | |||
891 | for (n = 0; n < 10; n++) { | ||
892 | double delta = old[n] - i_lsps[n]; | ||
893 | a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n]; | ||
894 | a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n]; | ||
895 | } | ||
896 | |||
897 | dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r, | ||
898 | mul_lsf, base_lsf); | ||
899 | } | ||
900 | |||
901 | /** | ||
902 | * Parse 16 independently-coded LSPs. | ||
903 | */ | ||
904 | static void dequant_lsp16i(GetBitContext *gb, double *lsps) | ||
905 | { | ||
906 | static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 }; | ||
907 | static const double mul_lsf[5] = { | ||
908 | 3.3439586280e-3, 6.9908173703e-4, | ||
909 | 3.3216608306e-3, 1.0334960326e-3, | ||
910 | 3.1899104283e-3 | ||
911 | }; | ||
912 | static const double base_lsf[5] = { | ||
913 | M_PI * -1.27576e-1, M_PI * -2.4292e-2, | ||
914 | M_PI * -1.28094e-1, M_PI * -3.2128e-2, | ||
915 | M_PI * -1.29816e-1 | ||
916 | }; | ||
917 | uint16_t v[5]; | ||
918 | |||
919 | v[0] = get_bits(gb, 8); | ||
920 | v[1] = get_bits(gb, 6); | ||
921 | v[2] = get_bits(gb, 7); | ||
922 | v[3] = get_bits(gb, 6); | ||
923 | v[4] = get_bits(gb, 7); | ||
924 | |||
925 | dequant_lsps( lsps, 5, v, vec_sizes, 2, | ||
926 | wmavoice_dq_lsp16i1, mul_lsf, base_lsf); | ||
927 | dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2, | ||
928 | wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]); | ||
929 | dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1, | ||
930 | wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]); | ||
931 | } | ||
932 | |||
933 | /** | ||
934 | * Parse 16 independently-coded LSPs, and then derive the tables to | ||
935 | * generate LSPs for the other frames from them (residual coding). | ||
936 | */ | ||
937 | static void dequant_lsp16r(GetBitContext *gb, | ||
938 | double *i_lsps, const double *old, | ||
939 | double *a1, double *a2, int q_mode) | ||
940 | { | ||
941 | static const uint16_t vec_sizes[3] = { 128, 128, 128 }; | ||
942 | static const double mul_lsf[3] = { | ||
943 | 1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3 | ||
944 | }; | ||
945 | static const double base_lsf[3] = { | ||
946 | M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2 | ||
947 | }; | ||
948 | const float (*ipol_tab)[2][16] = q_mode ? | ||
949 | wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a; | ||
950 | uint16_t interpol, v[3]; | ||
951 | int n; | ||
952 | |||
953 | dequant_lsp16i(gb, i_lsps); | ||
954 | |||
955 | interpol = get_bits(gb, 5); | ||
956 | v[0] = get_bits(gb, 7); | ||
957 | v[1] = get_bits(gb, 7); | ||
958 | v[2] = get_bits(gb, 7); | ||
959 | |||
960 | for (n = 0; n < 16; n++) { | ||
961 | double delta = old[n] - i_lsps[n]; | ||
962 | a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n]; | ||
963 | a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n]; | ||
964 | } | ||
965 | |||
966 | dequant_lsps( a2, 10, v, vec_sizes, 1, | ||
967 | wmavoice_dq_lsp16r1, mul_lsf, base_lsf); | ||
968 | dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1, | ||
969 | wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]); | ||
970 | dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1, | ||
971 | wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]); | ||
972 | } | ||
973 | |||
974 | /** | ||
975 | * @} | ||
976 | * @defgroup aw Pitch-adaptive window coding functions | ||
977 | * The next few functions are for pitch-adaptive window coding. | ||
978 | * @{ | ||
979 | */ | ||
980 | /** | ||
981 | * Parse the offset of the first pitch-adaptive window pulses, and | ||
982 | * the distribution of pulses between the two blocks in this frame. | ||
983 | * @param s WMA Voice decoding context private data | ||
984 | * @param gb bit I/O context | ||
985 | * @param pitch pitch for each block in this frame | ||
986 | */ | ||
987 | static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb, | ||
988 | const int *pitch) | ||
989 | { | ||
990 | static const int16_t start_offset[94] = { | ||
991 | -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11, | ||
992 | 13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26, | ||
993 | 27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43, | ||
994 | 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67, | ||
995 | 69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91, | ||
996 | 93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115, | ||
997 | 117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139, | ||
998 | 141, 143, 145, 147, 149, 151, 153, 155, 157, 159 | ||
999 | }; | ||
1000 | int bits, offset; | ||
1001 | |||
1002 | /* position of pulse */ | ||
1003 | s->aw_idx_is_ext = 0; | ||
1004 | if ((bits = get_bits(gb, 6)) >= 54) { | ||
1005 | s->aw_idx_is_ext = 1; | ||
1006 | bits += (bits - 54) * 3 + get_bits(gb, 2); | ||
1007 | } | ||
1008 | |||
1009 | /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count | ||
1010 | * the distribution of the pulses in each block contained in this frame. */ | ||
1011 | s->aw_pulse_range = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16; | ||
1012 | for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ; | ||
1013 | s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0]; | ||
1014 | s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2; | ||
1015 | offset += s->aw_n_pulses[0] * pitch[0]; | ||
1016 | s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1]; | ||
1017 | s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2; | ||
1018 | |||
1019 | /* if continuing from a position before the block, reset position to | ||
1020 | * start of block (when corrected for the range over which it can be | ||
1021 | * spread in aw_pulse_set1()). */ | ||
1022 | if (start_offset[bits] < MAX_FRAMESIZE / 2) { | ||
1023 | while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0) | ||
1024 | s->aw_first_pulse_off[1] -= pitch[1]; | ||
1025 | if (start_offset[bits] < 0) | ||
1026 | while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0) | ||
1027 | s->aw_first_pulse_off[0] -= pitch[0]; | ||
1028 | } | ||
1029 | } | ||
1030 | |||
1031 | /** | ||
1032 | * Apply second set of pitch-adaptive window pulses. | ||
1033 | * @param s WMA Voice decoding context private data | ||
1034 | * @param gb bit I/O context | ||
1035 | * @param block_idx block index in frame [0, 1] | ||
1036 | * @param fcb structure containing fixed codebook vector info | ||
1037 | */ | ||
1038 | static void aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb, | ||
1039 | int block_idx, AMRFixed *fcb) | ||
1040 | { | ||
1041 | uint16_t use_mask[7]; // only 5 are used, rest is padding | ||
1042 | /* in this function, idx is the index in the 80-bit (+ padding) use_mask | ||
1043 | * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits | ||
1044 | * of idx are the position of the bit within a particular item in the | ||
1045 | * array (0 being the most significant bit, and 15 being the least | ||
1046 | * significant bit), and the remainder (>> 4) is the index in the | ||
1047 | * use_mask[]-array. This is faster and uses less memory than using a | ||
1048 | * 80-byte/80-int array. */ | ||
1049 | int pulse_off = s->aw_first_pulse_off[block_idx], | ||
1050 | pulse_start, n, idx, range, aidx, start_off = 0; | ||
1051 | |||
1052 | /* set offset of first pulse to within this block */ | ||
1053 | if (s->aw_n_pulses[block_idx] > 0) | ||
1054 | while (pulse_off + s->aw_pulse_range < 1) | ||
1055 | pulse_off += fcb->pitch_lag; | ||
1056 | |||
1057 | /* find range per pulse */ | ||
1058 | if (s->aw_n_pulses[0] > 0) { | ||
1059 | if (block_idx == 0) { | ||
1060 | range = 32; | ||
1061 | } else /* block_idx = 1 */ { | ||
1062 | range = 8; | ||
1063 | if (s->aw_n_pulses[block_idx] > 0) | ||
1064 | pulse_off = s->aw_next_pulse_off_cache; | ||
1065 | } | ||
1066 | } else | ||
1067 | range = 16; | ||
1068 | pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0; | ||
1069 | |||
1070 | /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly, | ||
1071 | * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus | ||
1072 | * we exclude that range from being pulsed again in this function. */ | ||
1073 | memset( use_mask, -1, 5 * sizeof(use_mask[0])); | ||
1074 | memset(&use_mask[5], 0, 2 * sizeof(use_mask[0])); | ||
1075 | if (s->aw_n_pulses[block_idx] > 0) | ||
1076 | for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) { | ||
1077 | int excl_range = s->aw_pulse_range; // always 16 or 24 | ||
1078 | uint16_t *use_mask_ptr = &use_mask[idx >> 4]; | ||
1079 | int first_sh = 16 - (idx & 15); | ||
1080 | *use_mask_ptr++ &= 0xFFFF << first_sh; | ||
1081 | excl_range -= first_sh; | ||
1082 | if (excl_range >= 16) { | ||
1083 | *use_mask_ptr++ = 0; | ||
1084 | *use_mask_ptr &= 0xFFFF >> (excl_range - 16); | ||
1085 | } else | ||
1086 | *use_mask_ptr &= 0xFFFF >> excl_range; | ||
1087 | } | ||
1088 | |||
1089 | /* find the 'aidx'th offset that is not excluded */ | ||
1090 | aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4); | ||
1091 | for (n = 0; n <= aidx; pulse_start++) { | ||
1092 | for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ; | ||
1093 | if (idx >= MAX_FRAMESIZE / 2) { // find from zero | ||
1094 | if (use_mask[0]) idx = 0x0F; | ||
1095 | else if (use_mask[1]) idx = 0x1F; | ||
1096 | else if (use_mask[2]) idx = 0x2F; | ||
1097 | else if (use_mask[3]) idx = 0x3F; | ||
1098 | else if (use_mask[4]) idx = 0x4F; | ||
1099 | else return; | ||
1100 | idx -= av_log2_16bit(use_mask[idx >> 4]); | ||
1101 | } | ||
1102 | if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) { | ||
1103 | use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15)); | ||
1104 | n++; | ||
1105 | start_off = idx; | ||
1106 | } | ||
1107 | } | ||
1108 | |||
1109 | fcb->x[fcb->n] = start_off; | ||
1110 | fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0; | ||
1111 | fcb->n++; | ||
1112 | |||
1113 | /* set offset for next block, relative to start of that block */ | ||
1114 | n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag; | ||
1115 | s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0; | ||
1116 | } | ||
1117 | |||
1118 | /** | ||
1119 | * Apply first set of pitch-adaptive window pulses. | ||
1120 | * @param s WMA Voice decoding context private data | ||
1121 | * @param gb bit I/O context | ||
1122 | * @param block_idx block index in frame [0, 1] | ||
1123 | * @param fcb storage location for fixed codebook pulse info | ||
1124 | */ | ||
1125 | static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb, | ||
1126 | int block_idx, AMRFixed *fcb) | ||
1127 | { | ||
1128 | int val = get_bits(gb, 12 - 2 * (s->aw_idx_is_ext && !block_idx)); | ||
1129 | float v; | ||
1130 | |||
1131 | if (s->aw_n_pulses[block_idx] > 0) { | ||
1132 | int n, v_mask, i_mask, sh, n_pulses; | ||
1133 | |||
1134 | if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each | ||
1135 | n_pulses = 3; | ||
1136 | v_mask = 8; | ||
1137 | i_mask = 7; | ||
1138 | sh = 4; | ||
1139 | } else { // 4 pulses, 1:sign + 2:index each | ||
1140 | n_pulses = 4; | ||
1141 | v_mask = 4; | ||
1142 | i_mask = 3; | ||
1143 | sh = 3; | ||
1144 | } | ||
1145 | |||
1146 | for (n = n_pulses - 1; n >= 0; n--, val >>= sh) { | ||
1147 | fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0; | ||
1148 | fcb->x[fcb->n] = (val & i_mask) * n_pulses + n + | ||
1149 | s->aw_first_pulse_off[block_idx]; | ||
1150 | while (fcb->x[fcb->n] < 0) | ||
1151 | fcb->x[fcb->n] += fcb->pitch_lag; | ||
1152 | if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2) | ||
1153 | fcb->n++; | ||
1154 | } | ||
1155 | } else { | ||
1156 | int num2 = (val & 0x1FF) >> 1, delta, idx; | ||
1157 | |||
1158 | if (num2 < 1 * 79) { delta = 1; idx = num2 + 1; } | ||
1159 | else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; } | ||
1160 | else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; } | ||
1161 | else { delta = 7; idx = num2 + 1 - 3 * 75; } | ||
1162 | v = (val & 0x200) ? -1.0 : 1.0; | ||
1163 | |||
1164 | fcb->no_repeat_mask |= 3 << fcb->n; | ||
1165 | fcb->x[fcb->n] = idx - delta; | ||
1166 | fcb->y[fcb->n] = v; | ||
1167 | fcb->x[fcb->n + 1] = idx; | ||
1168 | fcb->y[fcb->n + 1] = (val & 1) ? -v : v; | ||
1169 | fcb->n += 2; | ||
1170 | } | ||
1171 | } | ||
1172 | |||
1173 | /** | ||
1174 | * @} | ||
1175 | * | ||
1176 | * Generate a random number from frame_cntr and block_idx, which will lief | ||
1177 | * in the range [0, 1000 - block_size] (so it can be used as an index in a | ||
1178 | * table of size 1000 of which you want to read block_size entries). | ||
1179 | * | ||
1180 | * @param frame_cntr current frame number | ||
1181 | * @param block_num current block index | ||
1182 | * @param block_size amount of entries we want to read from a table | ||
1183 | * that has 1000 entries | ||
1184 | * @return a (non-)random number in the [0, 1000 - block_size] range. | ||
1185 | */ | ||
1186 | static int pRNG(int frame_cntr, int block_num, int block_size) | ||
1187 | { | ||
1188 | /* array to simplify the calculation of z: | ||
1189 | * y = (x % 9) * 5 + 6; | ||
1190 | * z = (49995 * x) / y; | ||
1191 | * Since y only has 9 values, we can remove the division by using a | ||
1192 | * LUT and using FASTDIV-style divisions. For each of the 9 values | ||
1193 | * of y, we can rewrite z as: | ||
1194 | * z = x * (49995 / y) + x * ((49995 % y) / y) | ||
1195 | * In this table, each col represents one possible value of y, the | ||
1196 | * first number is 49995 / y, and the second is the FASTDIV variant | ||
1197 | * of 49995 % y / y. */ | ||
1198 | static const unsigned int div_tbl[9][2] = { | ||
1199 | { 8332, 3 * 715827883U }, // y = 6 | ||
1200 | { 4545, 0 * 390451573U }, // y = 11 | ||
1201 | { 3124, 11 * 268435456U }, // y = 16 | ||
1202 | { 2380, 15 * 204522253U }, // y = 21 | ||
1203 | { 1922, 23 * 165191050U }, // y = 26 | ||
1204 | { 1612, 23 * 138547333U }, // y = 31 | ||
1205 | { 1388, 27 * 119304648U }, // y = 36 | ||
1206 | { 1219, 16 * 104755300U }, // y = 41 | ||
1207 | { 1086, 39 * 93368855U } // y = 46 | ||
1208 | }; | ||
1209 | unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr; | ||
1210 | if (x >= 0xFFFF) x -= 0xFFFF; // max value of x is 8*1877+0xFFFE=0x13AA6, | ||
1211 | // so this is effectively a modulo (%) | ||
1212 | y = x - 9 * MULH(477218589, x); // x % 9 | ||
1213 | z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1])); | ||
1214 | // z = x * 49995 / (y * 5 + 6) | ||
1215 | return z % (1000 - block_size); | ||
1216 | } | ||
1217 | |||
1218 | /** | ||
1219 | * Parse hardcoded signal for a single block. | ||
1220 | * @note see #synth_block(). | ||
1221 | */ | ||
1222 | static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb, | ||
1223 | int block_idx, int size, | ||
1224 | const struct frame_type_desc *frame_desc, | ||
1225 | float *excitation) | ||
1226 | { | ||
1227 | float gain; | ||
1228 | int n, r_idx; | ||
1229 | |||
1230 | assert(size <= MAX_FRAMESIZE); | ||
1231 | |||
1232 | /* Set the offset from which we start reading wmavoice_std_codebook */ | ||
1233 | if (frame_desc->fcb_type == FCB_TYPE_SILENCE) { | ||
1234 | r_idx = pRNG(s->frame_cntr, block_idx, size); | ||
1235 | gain = s->silence_gain; | ||
1236 | } else /* FCB_TYPE_HARDCODED */ { | ||
1237 | r_idx = get_bits(gb, 8); | ||
1238 | gain = wmavoice_gain_universal[get_bits(gb, 6)]; | ||
1239 | } | ||
1240 | |||
1241 | /* Clear gain prediction parameters */ | ||
1242 | memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err)); | ||
1243 | |||
1244 | /* Apply gain to hardcoded codebook and use that as excitation signal */ | ||
1245 | for (n = 0; n < size; n++) | ||
1246 | excitation[n] = wmavoice_std_codebook[r_idx + n] * gain; | ||
1247 | } | ||
1248 | |||
1249 | /** | ||
1250 | * Parse FCB/ACB signal for a single block. | ||
1251 | * @note see #synth_block(). | ||
1252 | */ | ||
1253 | static void synth_block_fcb_acb(WMAVoiceContext *s, GetBitContext *gb, | ||
1254 | int block_idx, int size, | ||
1255 | int block_pitch_sh2, | ||
1256 | const struct frame_type_desc *frame_desc, | ||
1257 | float *excitation) | ||
1258 | { | ||
1259 | static const float gain_coeff[6] = { | ||
1260 | 0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458 | ||
1261 | }; | ||
1262 | float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain; | ||
1263 | int n, idx, gain_weight; | ||
1264 | AMRFixed fcb; | ||
1265 | |||
1266 | assert(size <= MAX_FRAMESIZE / 2); | ||
1267 | memset(pulses, 0, sizeof(*pulses) * size); | ||
1268 | |||
1269 | fcb.pitch_lag = block_pitch_sh2 >> 2; | ||
1270 | fcb.pitch_fac = 1.0; | ||
1271 | fcb.no_repeat_mask = 0; | ||
1272 | fcb.n = 0; | ||
1273 | |||
1274 | /* For the other frame types, this is where we apply the innovation | ||
1275 | * (fixed) codebook pulses of the speech signal. */ | ||
1276 | if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) { | ||
1277 | aw_pulse_set1(s, gb, block_idx, &fcb); | ||
1278 | aw_pulse_set2(s, gb, block_idx, &fcb); | ||
1279 | } else /* FCB_TYPE_EXC_PULSES */ { | ||
1280 | int offset_nbits = 5 - frame_desc->log_n_blocks; | ||
1281 | |||
1282 | fcb.no_repeat_mask = -1; | ||
1283 | /* similar to ff_decode_10_pulses_35bits(), but with single pulses | ||
1284 | * (instead of double) for a subset of pulses */ | ||
1285 | for (n = 0; n < 5; n++) { | ||
1286 | float sign; | ||
1287 | int pos1, pos2; | ||
1288 | |||
1289 | sign = get_bits1(gb) ? 1.0 : -1.0; | ||
1290 | pos1 = get_bits(gb, offset_nbits); | ||
1291 | fcb.x[fcb.n] = n + 5 * pos1; | ||
1292 | fcb.y[fcb.n++] = sign; | ||
1293 | if (n < frame_desc->dbl_pulses) { | ||
1294 | pos2 = get_bits(gb, offset_nbits); | ||
1295 | fcb.x[fcb.n] = n + 5 * pos2; | ||
1296 | fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign; | ||
1297 | } | ||
1298 | } | ||
1299 | } | ||
1300 | ff_set_fixed_vector(pulses, &fcb, 1.0, size); | ||
1301 | |||
1302 | /* Calculate gain for adaptive & fixed codebook signal. | ||
1303 | * see ff_amr_set_fixed_gain(). */ | ||
1304 | idx = get_bits(gb, 7); | ||
1305 | fcb_gain = expf(ff_dot_productf(s->gain_pred_err, gain_coeff, 6) - | ||
1306 | 5.2409161640 + wmavoice_gain_codebook_fcb[idx]); | ||
1307 | acb_gain = wmavoice_gain_codebook_acb[idx]; | ||
1308 | pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx], | ||
1309 | -2.9957322736 /* log(0.05) */, | ||
1310 | 1.6094379124 /* log(5.0) */); | ||
1311 | |||
1312 | gain_weight = 8 >> frame_desc->log_n_blocks; | ||
1313 | memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err, | ||
1314 | sizeof(*s->gain_pred_err) * (6 - gain_weight)); | ||
1315 | for (n = 0; n < gain_weight; n++) | ||
1316 | s->gain_pred_err[n] = pred_err; | ||
1317 | |||
1318 | /* Calculation of adaptive codebook */ | ||
1319 | if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) { | ||
1320 | int len; | ||
1321 | for (n = 0; n < size; n += len) { | ||
1322 | int next_idx_sh16; | ||
1323 | int abs_idx = block_idx * size + n; | ||
1324 | int pitch_sh16 = (s->last_pitch_val << 16) + | ||
1325 | s->pitch_diff_sh16 * abs_idx; | ||
1326 | int pitch = (pitch_sh16 + 0x6FFF) >> 16; | ||
1327 | int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000; | ||
1328 | idx = idx_sh16 >> 16; | ||
1329 | if (s->pitch_diff_sh16) { | ||
1330 | if (s->pitch_diff_sh16 > 0) { | ||
1331 | next_idx_sh16 = (idx_sh16) &~ 0xFFFF; | ||
1332 | } else | ||
1333 | next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF; | ||
1334 | len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8, | ||
1335 | 1, size - n); | ||
1336 | } else | ||
1337 | len = size; | ||
1338 | |||
1339 | ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch], | ||
1340 | wmavoice_ipol1_coeffs, 17, | ||
1341 | idx, 9, len); | ||
1342 | } | ||
1343 | } else /* ACB_TYPE_HAMMING */ { | ||
1344 | int block_pitch = block_pitch_sh2 >> 2; | ||
1345 | idx = block_pitch_sh2 & 3; | ||
1346 | if (idx) { | ||
1347 | ff_acelp_interpolatef(excitation, &excitation[-block_pitch], | ||
1348 | wmavoice_ipol2_coeffs, 4, | ||
1349 | idx, 8, size); | ||
1350 | } else | ||
1351 | av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch, | ||
1352 | sizeof(float) * size); | ||
1353 | } | ||
1354 | |||
1355 | /* Interpolate ACB/FCB and use as excitation signal */ | ||
1356 | ff_weighted_vector_sumf(excitation, excitation, pulses, | ||
1357 | acb_gain, fcb_gain, size); | ||
1358 | } | ||
1359 | |||
1360 | /** | ||
1361 | * Parse data in a single block. | ||
1362 | * @note we assume enough bits are available, caller should check. | ||
1363 | * | ||
1364 | * @param s WMA Voice decoding context private data | ||
1365 | * @param gb bit I/O context | ||
1366 | * @param block_idx index of the to-be-read block | ||
1367 | * @param size amount of samples to be read in this block | ||
1368 | * @param block_pitch_sh2 pitch for this block << 2 | ||
1369 | * @param lsps LSPs for (the end of) this frame | ||
1370 | * @param prev_lsps LSPs for the last frame | ||
1371 | * @param frame_desc frame type descriptor | ||
1372 | * @param excitation target memory for the ACB+FCB interpolated signal | ||
1373 | * @param synth target memory for the speech synthesis filter output | ||
1374 | * @return 0 on success, <0 on error. | ||
1375 | */ | ||
1376 | static void synth_block(WMAVoiceContext *s, GetBitContext *gb, | ||
1377 | int block_idx, int size, | ||
1378 | int block_pitch_sh2, | ||
1379 | const double *lsps, const double *prev_lsps, | ||
1380 | const struct frame_type_desc *frame_desc, | ||
1381 | float *excitation, float *synth) | ||
1382 | { | ||
1383 | double i_lsps[MAX_LSPS]; | ||
1384 | float lpcs[MAX_LSPS]; | ||
1385 | float fac; | ||
1386 | int n; | ||
1387 | |||
1388 | if (frame_desc->acb_type == ACB_TYPE_NONE) | ||
1389 | synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation); | ||
1390 | else | ||
1391 | synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2, | ||
1392 | frame_desc, excitation); | ||
1393 | |||
1394 | /* convert interpolated LSPs to LPCs */ | ||
1395 | fac = (block_idx + 0.5) / frame_desc->n_blocks; | ||
1396 | for (n = 0; n < s->lsps; n++) // LSF -> LSP | ||
1397 | i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n])); | ||
1398 | ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1); | ||
1399 | |||
1400 | /* Speech synthesis */ | ||
1401 | ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps); | ||
1402 | } | ||
1403 | |||
1404 | /** | ||
1405 | * Synthesize output samples for a single frame. | ||
1406 | * @note we assume enough bits are available, caller should check. | ||
1407 | * | ||
1408 | * @param ctx WMA Voice decoder context | ||
1409 | * @param gb bit I/O context (s->gb or one for cross-packet superframes) | ||
1410 | * @param frame_idx Frame number within superframe [0-2] | ||
1411 | * @param samples pointer to output sample buffer, has space for at least 160 | ||
1412 | * samples | ||
1413 | * @param lsps LSP array | ||
1414 | * @param prev_lsps array of previous frame's LSPs | ||
1415 | * @param excitation target buffer for excitation signal | ||
1416 | * @param synth target buffer for synthesized speech data | ||
1417 | * @return 0 on success, <0 on error. | ||
1418 | */ | ||
1419 | static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx, | ||
1420 | float *samples, | ||
1421 | const double *lsps, const double *prev_lsps, | ||
1422 | float *excitation, float *synth) | ||
1423 | { | ||
1424 | WMAVoiceContext *s = ctx->priv_data; | ||
1425 | int n, n_blocks_x2, log_n_blocks_x2, cur_pitch_val; | ||
1426 | int pitch[MAX_BLOCKS], last_block_pitch; | ||
1427 | |||
1428 | /* Parse frame type ("frame header"), see frame_descs */ | ||
1429 | int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)], | ||
1430 | block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks; | ||
1431 | |||
1432 | if (bd_idx < 0) { | ||
1433 | av_log(ctx, AV_LOG_ERROR, | ||
1434 | "Invalid frame type VLC code, skipping\n"); | ||
1435 | return -1; | ||
1436 | } | ||
1437 | |||
1438 | /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */ | ||
1439 | if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) { | ||
1440 | /* Pitch is provided per frame, which is interpreted as the pitch of | ||
1441 | * the last sample of the last block of this frame. We can interpolate | ||
1442 | * the pitch of other blocks (and even pitch-per-sample) by gradually | ||
1443 | * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */ | ||
1444 | n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1; | ||
1445 | log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1; | ||
1446 | cur_pitch_val = s->min_pitch_val + get_bits(gb, s->pitch_nbits); | ||
1447 | cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1); | ||
1448 | if (s->last_acb_type == ACB_TYPE_NONE || | ||
1449 | 20 * abs(cur_pitch_val - s->last_pitch_val) > | ||
1450 | (cur_pitch_val + s->last_pitch_val)) | ||
1451 | s->last_pitch_val = cur_pitch_val; | ||
1452 | |||
1453 | /* pitch per block */ | ||
1454 | for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) { | ||
1455 | int fac = n * 2 + 1; | ||
1456 | |||
1457 | pitch[n] = (MUL16(fac, cur_pitch_val) + | ||
1458 | MUL16((n_blocks_x2 - fac), s->last_pitch_val) + | ||
1459 | frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2; | ||
1460 | } | ||
1461 | |||
1462 | /* "pitch-diff-per-sample" for calculation of pitch per sample */ | ||
1463 | s->pitch_diff_sh16 = | ||
1464 | ((cur_pitch_val - s->last_pitch_val) << 16) / MAX_FRAMESIZE; | ||
1465 | } | ||
1466 | |||
1467 | /* Global gain (if silence) and pitch-adaptive window coordinates */ | ||
1468 | switch (frame_descs[bd_idx].fcb_type) { | ||
1469 | case FCB_TYPE_SILENCE: | ||
1470 | s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)]; | ||
1471 | break; | ||
1472 | case FCB_TYPE_AW_PULSES: | ||
1473 | aw_parse_coords(s, gb, pitch); | ||
1474 | break; | ||
1475 | } | ||
1476 | |||
1477 | for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) { | ||
1478 | int bl_pitch_sh2; | ||
1479 | |||
1480 | /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */ | ||
1481 | switch (frame_descs[bd_idx].acb_type) { | ||
1482 | case ACB_TYPE_HAMMING: { | ||
1483 | /* Pitch is given per block. Per-block pitches are encoded as an | ||
1484 | * absolute value for the first block, and then delta values | ||
1485 | * relative to this value) for all subsequent blocks. The scale of | ||
1486 | * this pitch value is semi-logaritmic compared to its use in the | ||
1487 | * decoder, so we convert it to normal scale also. */ | ||
1488 | int block_pitch, | ||
1489 | t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2, | ||
1490 | t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1, | ||
1491 | t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1; | ||
1492 | |||
1493 | if (n == 0) { | ||
1494 | block_pitch = get_bits(gb, s->block_pitch_nbits); | ||
1495 | } else | ||
1496 | block_pitch = last_block_pitch - s->block_delta_pitch_hrange + | ||
1497 | get_bits(gb, s->block_delta_pitch_nbits); | ||
1498 | /* Convert last_ so that any next delta is within _range */ | ||
1499 | last_block_pitch = av_clip(block_pitch, | ||
1500 | s->block_delta_pitch_hrange, | ||
1501 | s->block_pitch_range - | ||
1502 | s->block_delta_pitch_hrange); | ||
1503 | |||
1504 | /* Convert semi-log-style scale back to normal scale */ | ||
1505 | if (block_pitch < t1) { | ||
1506 | bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch; | ||
1507 | } else { | ||
1508 | block_pitch -= t1; | ||
1509 | if (block_pitch < t2) { | ||
1510 | bl_pitch_sh2 = | ||
1511 | (s->block_conv_table[1] << 2) + (block_pitch << 1); | ||
1512 | } else { | ||
1513 | block_pitch -= t2; | ||
1514 | if (block_pitch < t3) { | ||
1515 | bl_pitch_sh2 = | ||
1516 | (s->block_conv_table[2] + block_pitch) << 2; | ||
1517 | } else | ||
1518 | bl_pitch_sh2 = s->block_conv_table[3] << 2; | ||
1519 | } | ||
1520 | } | ||
1521 | pitch[n] = bl_pitch_sh2 >> 2; | ||
1522 | break; | ||
1523 | } | ||
1524 | |||
1525 | case ACB_TYPE_ASYMMETRIC: { | ||
1526 | bl_pitch_sh2 = pitch[n] << 2; | ||
1527 | break; | ||
1528 | } | ||
1529 | |||
1530 | default: // ACB_TYPE_NONE has no pitch | ||
1531 | bl_pitch_sh2 = 0; | ||
1532 | break; | ||
1533 | } | ||
1534 | |||
1535 | synth_block(s, gb, n, block_nsamples, bl_pitch_sh2, | ||
1536 | lsps, prev_lsps, &frame_descs[bd_idx], | ||
1537 | &excitation[n * block_nsamples], | ||
1538 | &synth[n * block_nsamples]); | ||
1539 | } | ||
1540 | |||
1541 | /* Averaging projection filter, if applicable. Else, just copy samples | ||
1542 | * from synthesis buffer */ | ||
1543 | if (s->do_apf) { | ||
1544 | double i_lsps[MAX_LSPS]; | ||
1545 | float lpcs[MAX_LSPS]; | ||
1546 | |||
1547 | for (n = 0; n < s->lsps; n++) // LSF -> LSP | ||
1548 | i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n])); | ||
1549 | ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1); | ||
1550 | postfilter(s, synth, samples, 80, lpcs, | ||
1551 | &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx], | ||
1552 | frame_descs[bd_idx].fcb_type, pitch[0]); | ||
1553 | |||
1554 | for (n = 0; n < s->lsps; n++) // LSF -> LSP | ||
1555 | i_lsps[n] = cos(lsps[n]); | ||
1556 | ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1); | ||
1557 | postfilter(s, &synth[80], &samples[80], 80, lpcs, | ||
1558 | &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80], | ||
1559 | frame_descs[bd_idx].fcb_type, pitch[0]); | ||
1560 | } else | ||
1561 | memcpy(samples, synth, 160 * sizeof(synth[0])); | ||
1562 | |||
1563 | /* Cache values for next frame */ | ||
1564 | s->frame_cntr++; | ||
1565 | if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%) | ||
1566 | s->last_acb_type = frame_descs[bd_idx].acb_type; | ||
1567 | switch (frame_descs[bd_idx].acb_type) { | ||
1568 | case ACB_TYPE_NONE: | ||
1569 | s->last_pitch_val = 0; | ||
1570 | break; | ||
1571 | case ACB_TYPE_ASYMMETRIC: | ||
1572 | s->last_pitch_val = cur_pitch_val; | ||
1573 | break; | ||
1574 | case ACB_TYPE_HAMMING: | ||
1575 | s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1]; | ||
1576 | break; | ||
1577 | } | ||
1578 | |||
1579 | return 0; | ||
1580 | } | ||
1581 | |||
1582 | /** | ||
1583 | * Ensure minimum value for first item, maximum value for last value, | ||
1584 | * proper spacing between each value and proper ordering. | ||
1585 | * | ||
1586 | * @param lsps array of LSPs | ||
1587 | * @param num size of LSP array | ||
1588 | * | ||
1589 | * @note basically a double version of #ff_acelp_reorder_lsf(), might be | ||
1590 | * useful to put in a generic location later on. Parts are also | ||
1591 | * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(), | ||
1592 | * which is in float. | ||
1593 | */ | ||
1594 | static void stabilize_lsps(double *lsps, int num) | ||
1595 | { | ||
1596 | int n, m, l; | ||
1597 | |||
1598 | /* set minimum value for first, maximum value for last and minimum | ||
1599 | * spacing between LSF values. | ||
1600 | * Very similar to ff_set_min_dist_lsf(), but in double. */ | ||
1601 | lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI); | ||
1602 | for (n = 1; n < num; n++) | ||
1603 | lsps[n] = FFMAX(lsps[n], lsps[n - 1] + 0.0125 * M_PI); | ||
1604 | lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI); | ||
1605 | |||
1606 | /* reorder (looks like one-time / non-recursed bubblesort). | ||
1607 | * Very similar to ff_sort_nearly_sorted_floats(), but in double. */ | ||
1608 | for (n = 1; n < num; n++) { | ||
1609 | if (lsps[n] < lsps[n - 1]) { | ||
1610 | for (m = 1; m < num; m++) { | ||
1611 | double tmp = lsps[m]; | ||
1612 | for (l = m - 1; l >= 0; l--) { | ||
1613 | if (lsps[l] <= tmp) break; | ||
1614 | lsps[l + 1] = lsps[l]; | ||
1615 | } | ||
1616 | lsps[l + 1] = tmp; | ||
1617 | } | ||
1618 | break; | ||
1619 | } | ||
1620 | } | ||
1621 | } | ||
1622 | |||
1623 | /** | ||
1624 | * Test if there's enough bits to read 1 superframe. | ||
1625 | * | ||
1626 | * @param orig_gb bit I/O context used for reading. This function | ||
1627 | * does not modify the state of the bitreader; it | ||
1628 | * only uses it to copy the current stream position | ||
1629 | * @param s WMA Voice decoding context private data | ||
1630 | * @return -1 if unsupported, 1 on not enough bits or 0 if OK. | ||
1631 | */ | ||
1632 | static int check_bits_for_superframe(GetBitContext *orig_gb, | ||
1633 | WMAVoiceContext *s) | ||
1634 | { | ||
1635 | GetBitContext s_gb, *gb = &s_gb; | ||
1636 | int n, need_bits, bd_idx; | ||
1637 | const struct frame_type_desc *frame_desc; | ||
1638 | |||
1639 | /* initialize a copy */ | ||
1640 | init_get_bits(gb, orig_gb->buffer, orig_gb->size_in_bits); | ||
1641 | skip_bits_long(gb, get_bits_count(orig_gb)); | ||
1642 | assert(get_bits_left(gb) == get_bits_left(orig_gb)); | ||
1643 | |||
1644 | /* superframe header */ | ||
1645 | if (get_bits_left(gb) < 14) | ||
1646 | return 1; | ||
1647 | if (!get_bits1(gb)) | ||
1648 | return -1; // WMAPro-in-WMAVoice superframe | ||
1649 | if (get_bits1(gb)) skip_bits(gb, 12); // number of samples in superframe | ||
1650 | if (s->has_residual_lsps) { // residual LSPs (for all frames) | ||
1651 | if (get_bits_left(gb) < s->sframe_lsp_bitsize) | ||
1652 | return 1; | ||
1653 | skip_bits_long(gb, s->sframe_lsp_bitsize); | ||
1654 | } | ||
1655 | |||
1656 | /* frames */ | ||
1657 | for (n = 0; n < MAX_FRAMES; n++) { | ||
1658 | int aw_idx_is_ext = 0; | ||
1659 | |||
1660 | if (!s->has_residual_lsps) { // independent LSPs (per-frame) | ||
1661 | if (get_bits_left(gb) < s->frame_lsp_bitsize) return 1; | ||
1662 | skip_bits_long(gb, s->frame_lsp_bitsize); | ||
1663 | } | ||
1664 | bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)]; | ||
1665 | if (bd_idx < 0) | ||
1666 | return -1; // invalid frame type VLC code | ||
1667 | frame_desc = &frame_descs[bd_idx]; | ||
1668 | if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) { | ||
1669 | if (get_bits_left(gb) < s->pitch_nbits) | ||
1670 | return 1; | ||
1671 | skip_bits_long(gb, s->pitch_nbits); | ||
1672 | } | ||
1673 | if (frame_desc->fcb_type == FCB_TYPE_SILENCE) { | ||
1674 | skip_bits(gb, 8); | ||
1675 | } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) { | ||
1676 | int tmp = get_bits(gb, 6); | ||
1677 | if (tmp >= 0x36) { | ||
1678 | skip_bits(gb, 2); | ||
1679 | aw_idx_is_ext = 1; | ||
1680 | } | ||
1681 | } | ||
1682 | |||
1683 | /* blocks */ | ||
1684 | if (frame_desc->acb_type == ACB_TYPE_HAMMING) { | ||
1685 | need_bits = s->block_pitch_nbits + | ||
1686 | (frame_desc->n_blocks - 1) * s->block_delta_pitch_nbits; | ||
1687 | } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) { | ||
1688 | need_bits = 2 * !aw_idx_is_ext; | ||
1689 | } else | ||
1690 | need_bits = 0; | ||
1691 | need_bits += frame_desc->frame_size; | ||
1692 | if (get_bits_left(gb) < need_bits) | ||
1693 | return 1; | ||
1694 | skip_bits_long(gb, need_bits); | ||
1695 | } | ||
1696 | |||
1697 | return 0; | ||
1698 | } | ||
1699 | |||
1700 | /** | ||
1701 | * Synthesize output samples for a single superframe. If we have any data | ||
1702 | * cached in s->sframe_cache, that will be used instead of whatever is loaded | ||
1703 | * in s->gb. | ||
1704 | * | ||
1705 | * WMA Voice superframes contain 3 frames, each containing 160 audio samples, | ||
1706 | * to give a total of 480 samples per frame. See #synth_frame() for frame | ||
1707 | * parsing. In addition to 3 frames, superframes can also contain the LSPs | ||
1708 | * (if these are globally specified for all frames (residually); they can | ||
1709 | * also be specified individually per-frame. See the s->has_residual_lsps | ||
1710 | * option), and can specify the number of samples encoded in this superframe | ||
1711 | * (if less than 480), usually used to prevent blanks at track boundaries. | ||
1712 | * | ||
1713 | * @param ctx WMA Voice decoder context | ||
1714 | * @param samples pointer to output buffer for voice samples | ||
1715 | * @param data_size pointer containing the size of #samples on input, and the | ||
1716 | * amount of #samples filled on output | ||
1717 | * @return 0 on success, <0 on error or 1 if there was not enough data to | ||
1718 | * fully parse the superframe | ||
1719 | */ | ||
1720 | static int synth_superframe(AVCodecContext *ctx, | ||
1721 | float *samples, int *data_size) | ||
1722 | { | ||
1723 | WMAVoiceContext *s = ctx->priv_data; | ||
1724 | GetBitContext *gb = &s->gb, s_gb; | ||
1725 | int n, res, n_samples = 480; | ||
1726 | double lsps[MAX_FRAMES][MAX_LSPS]; | ||
1727 | const double *mean_lsf = s->lsps == 16 ? | ||
1728 | wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode]; | ||
1729 | float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12]; | ||
1730 | float synth[MAX_LSPS + MAX_SFRAMESIZE]; | ||
1731 | |||
1732 | memcpy(synth, s->synth_history, | ||
1733 | s->lsps * sizeof(*synth)); | ||
1734 | memcpy(excitation, s->excitation_history, | ||
1735 | s->history_nsamples * sizeof(*excitation)); | ||
1736 | |||
1737 | if (s->sframe_cache_size > 0) { | ||
1738 | gb = &s_gb; | ||
1739 | init_get_bits(gb, s->sframe_cache, s->sframe_cache_size); | ||
1740 | s->sframe_cache_size = 0; | ||
1741 | } | ||
1742 | |||
1743 | if ((res = check_bits_for_superframe(gb, s)) == 1) return 1; | ||
1744 | |||
1745 | /* First bit is speech/music bit, it differentiates between WMAVoice | ||
1746 | * speech samples (the actual codec) and WMAVoice music samples, which | ||
1747 | * are really WMAPro-in-WMAVoice-superframes. I've never seen those in | ||
1748 | * the wild yet. */ | ||
1749 | if (!get_bits1(gb)) { | ||
1750 | av_log_missing_feature(ctx, "WMAPro-in-WMAVoice support", 1); | ||
1751 | return ERROR_WMAPRO_IN_WMAVOICE; | ||
1752 | } | ||
1753 | |||
1754 | /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */ | ||
1755 | if (get_bits1(gb)) { | ||
1756 | if ((n_samples = get_bits(gb, 12)) > 480) { | ||
1757 | av_log(ctx, AV_LOG_ERROR, | ||
1758 | "Superframe encodes >480 samples (%d), not allowed\n", | ||
1759 | n_samples); | ||
1760 | return -1; | ||
1761 | } | ||
1762 | } | ||
1763 | /* Parse LSPs, if global for the superframe (can also be per-frame). */ | ||
1764 | if (s->has_residual_lsps) { | ||
1765 | double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2]; | ||
1766 | |||
1767 | for (n = 0; n < s->lsps; n++) | ||
1768 | prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n]; | ||
1769 | |||
1770 | if (s->lsps == 10) { | ||
1771 | dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode); | ||
1772 | } else /* s->lsps == 16 */ | ||
1773 | dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode); | ||
1774 | |||
1775 | for (n = 0; n < s->lsps; n++) { | ||
1776 | lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]); | ||
1777 | lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]); | ||
1778 | lsps[2][n] += mean_lsf[n]; | ||
1779 | } | ||
1780 | for (n = 0; n < 3; n++) | ||
1781 | stabilize_lsps(lsps[n], s->lsps); | ||
1782 | } | ||
1783 | |||
1784 | /* Parse frames, optionally preceeded by per-frame (independent) LSPs. */ | ||
1785 | for (n = 0; n < 3; n++) { | ||
1786 | if (!s->has_residual_lsps) { | ||
1787 | int m; | ||
1788 | |||
1789 | if (s->lsps == 10) { | ||
1790 | dequant_lsp10i(gb, lsps[n]); | ||
1791 | } else /* s->lsps == 16 */ | ||
1792 | dequant_lsp16i(gb, lsps[n]); | ||
1793 | |||
1794 | for (m = 0; m < s->lsps; m++) | ||
1795 | lsps[n][m] += mean_lsf[m]; | ||
1796 | stabilize_lsps(lsps[n], s->lsps); | ||
1797 | } | ||
1798 | |||
1799 | if ((res = synth_frame(ctx, gb, n, | ||
1800 | &samples[n * MAX_FRAMESIZE], | ||
1801 | lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1], | ||
1802 | &excitation[s->history_nsamples + n * MAX_FRAMESIZE], | ||
1803 | &synth[s->lsps + n * MAX_FRAMESIZE]))) | ||
1804 | return res; | ||
1805 | } | ||
1806 | |||
1807 | /* Statistics? FIXME - we don't check for length, a slight overrun | ||
1808 | * will be caught by internal buffer padding, and anything else | ||
1809 | * will be skipped, not read. */ | ||
1810 | if (get_bits1(gb)) { | ||
1811 | res = get_bits(gb, 4); | ||
1812 | skip_bits(gb, 10 * (res + 1)); | ||
1813 | } | ||
1814 | |||
1815 | /* Specify nr. of output samples */ | ||
1816 | *data_size = n_samples * sizeof(float); | ||
1817 | |||
1818 | /* Update history */ | ||
1819 | memcpy(s->prev_lsps, lsps[2], | ||
1820 | s->lsps * sizeof(*s->prev_lsps)); | ||
1821 | memcpy(s->synth_history, &synth[MAX_SFRAMESIZE], | ||
1822 | s->lsps * sizeof(*synth)); | ||
1823 | memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE], | ||
1824 | s->history_nsamples * sizeof(*excitation)); | ||
1825 | if (s->do_apf) | ||
1826 | memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE], | ||
1827 | s->history_nsamples * sizeof(*s->zero_exc_pf)); | ||
1828 | |||
1829 | return 0; | ||
1830 | } | ||
1831 | |||
1832 | /** | ||
1833 | * Parse the packet header at the start of each packet (input data to this | ||
1834 | * decoder). | ||
1835 | * | ||
1836 | * @param s WMA Voice decoding context private data | ||
1837 | * @return 1 if not enough bits were available, or 0 on success. | ||
1838 | */ | ||
1839 | static int parse_packet_header(WMAVoiceContext *s) | ||
1840 | { | ||
1841 | GetBitContext *gb = &s->gb; | ||
1842 | unsigned int res; | ||
1843 | |||
1844 | if (get_bits_left(gb) < 11) | ||
1845 | return 1; | ||
1846 | skip_bits(gb, 4); // packet sequence number | ||
1847 | s->has_residual_lsps = get_bits1(gb); | ||
1848 | do { | ||
1849 | res = get_bits(gb, 6); // number of superframes per packet | ||
1850 | // (minus first one if there is spillover) | ||
1851 | if (get_bits_left(gb) < 6 * (res == 0x3F) + s->spillover_bitsize) | ||
1852 | return 1; | ||
1853 | } while (res == 0x3F); | ||
1854 | s->spillover_nbits = get_bits(gb, s->spillover_bitsize); | ||
1855 | |||
1856 | return 0; | ||
1857 | } | ||
1858 | |||
1859 | /** | ||
1860 | * Copy (unaligned) bits from gb/data/size to pb. | ||
1861 | * | ||
1862 | * @param pb target buffer to copy bits into | ||
1863 | * @param data source buffer to copy bits from | ||
1864 | * @param size size of the source data, in bytes | ||
1865 | * @param gb bit I/O context specifying the current position in the source. | ||
1866 | * data. This function might use this to align the bit position to | ||
1867 | * a whole-byte boundary before calling #ff_copy_bits() on aligned | ||
1868 | * source data | ||
1869 | * @param nbits the amount of bits to copy from source to target | ||
1870 | * | ||
1871 | * @note after calling this function, the current position in the input bit | ||
1872 | * I/O context is undefined. | ||
1873 | */ | ||
1874 | static void copy_bits(PutBitContext *pb, | ||
1875 | const uint8_t *data, int size, | ||
1876 | GetBitContext *gb, int nbits) | ||
1877 | { | ||
1878 | int rmn_bytes, rmn_bits; | ||
1879 | |||
1880 | rmn_bits = rmn_bytes = get_bits_left(gb); | ||
1881 | if (rmn_bits < nbits) | ||
1882 | return; | ||
1883 | rmn_bits &= 7; rmn_bytes >>= 3; | ||
1884 | if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0) | ||
1885 | put_bits(pb, rmn_bits, get_bits(gb, rmn_bits)); | ||
1886 | ff_copy_bits(pb, data + size - rmn_bytes, | ||
1887 | FFMIN(nbits - rmn_bits, rmn_bytes << 3)); | ||
1888 | } | ||
1889 | |||
1890 | /** | ||
1891 | * Packet decoding: a packet is anything that the (ASF) demuxer contains, | ||
1892 | * and we expect that the demuxer / application provides it to us as such | ||
1893 | * (else you'll probably get garbage as output). Every packet has a size of | ||
1894 | * ctx->block_align bytes, starts with a packet header (see | ||
1895 | * #parse_packet_header()), and then a series of superframes. Superframe | ||
1896 | * boundaries may exceed packets, i.e. superframes can split data over | ||
1897 | * multiple (two) packets. | ||
1898 | * | ||
1899 | * For more information about frames, see #synth_superframe(). | ||
1900 | */ | ||
1901 | int wmavoice_decode_packet(AVCodecContext *ctx, void *data, | ||
1902 | int *data_size, AVPacket *avpkt) | ||
1903 | { | ||
1904 | WMAVoiceContext *s = ctx->priv_data; | ||
1905 | GetBitContext *gb = &s->gb; | ||
1906 | int size, res, pos; | ||
1907 | |||
1908 | if (*data_size < 480 * sizeof(float)) { | ||
1909 | av_log(ctx, AV_LOG_ERROR, | ||
1910 | "Output buffer too small (%d given - %zu needed)\n", | ||
1911 | *data_size, 480 * sizeof(float)); | ||
1912 | return -1; | ||
1913 | } | ||
1914 | *data_size = 0; | ||
1915 | |||
1916 | /* Packets are sometimes a multiple of ctx->block_align, with a packet | ||
1917 | * header at each ctx->block_align bytes. However, FFmpeg's ASF demuxer | ||
1918 | * feeds us ASF packets, which may concatenate multiple "codec" packets | ||
1919 | * in a single "muxer" packet, so we artificially emulate that by | ||
1920 | * capping the packet size at ctx->block_align. */ | ||
1921 | for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align); | ||
1922 | if (!size) | ||
1923 | return 0; | ||
1924 | init_get_bits(&s->gb, avpkt->data, size << 3); | ||
1925 | |||
1926 | /* size == ctx->block_align is used to indicate whether we are dealing with | ||
1927 | * a new packet or a packet of which we already read the packet header | ||
1928 | * previously. */ | ||
1929 | if (size == ctx->block_align) { // new packet header | ||
1930 | if ((res = parse_packet_header(s)) < 0) | ||
1931 | return res; | ||
1932 | |||
1933 | /* If the packet header specifies a s->spillover_nbits, then we want | ||
1934 | * to push out all data of the previous packet (+ spillover) before | ||
1935 | * continuing to parse new superframes in the current packet. */ | ||
1936 | if (s->spillover_nbits > 0) { | ||
1937 | if (s->sframe_cache_size > 0) { | ||
1938 | int cnt = get_bits_count(gb); | ||
1939 | copy_bits(&s->pb, avpkt->data, size, gb, s->spillover_nbits); | ||
1940 | flush_put_bits(&s->pb); | ||
1941 | s->sframe_cache_size += s->spillover_nbits; | ||
1942 | if ((res = synth_superframe(ctx, data, data_size)) == 0 && | ||
1943 | *data_size > 0) { | ||
1944 | /* convert the float values to int32 for rockbox */ | ||
1945 | int i; | ||
1946 | int32_t *iptr = data; | ||
1947 | float *fptr = data; | ||
1948 | for(i = 0; i < *data_size/sizeof(float); i++) | ||
1949 | { | ||
1950 | fptr[i] *= (float)(INT32_MAX); | ||
1951 | iptr[i] = (int32_t)fptr[i]; | ||
1952 | } | ||
1953 | cnt += s->spillover_nbits; | ||
1954 | s->skip_bits_next = cnt & 7; | ||
1955 | return cnt >> 3; | ||
1956 | } else | ||
1957 | skip_bits_long (gb, s->spillover_nbits - cnt + | ||
1958 | get_bits_count(gb)); // resync | ||
1959 | } else | ||
1960 | skip_bits_long(gb, s->spillover_nbits); // resync | ||
1961 | } | ||
1962 | } else if (s->skip_bits_next) | ||
1963 | skip_bits(gb, s->skip_bits_next); | ||
1964 | |||
1965 | /* Try parsing superframes in current packet */ | ||
1966 | s->sframe_cache_size = 0; | ||
1967 | s->skip_bits_next = 0; | ||
1968 | pos = get_bits_left(gb); | ||
1969 | if ((res = synth_superframe(ctx, data, data_size)) < 0) { | ||
1970 | return res; | ||
1971 | } else if (*data_size > 0) { | ||
1972 | int cnt = get_bits_count(gb); | ||
1973 | s->skip_bits_next = cnt & 7; | ||
1974 | /* convert the float values to int32 for rockbox */ | ||
1975 | int i; | ||
1976 | int32_t *iptr = data; | ||
1977 | float *fptr = data; | ||
1978 | for(i = 0; i < *data_size/sizeof(float); i++) | ||
1979 | { | ||
1980 | fptr[i] *= (float)(INT32_MAX); | ||
1981 | iptr[i] = (int32_t)fptr[i]; | ||
1982 | } | ||
1983 | return cnt >> 3; | ||
1984 | } else if ((s->sframe_cache_size = pos) > 0) { | ||
1985 | /* rewind bit reader to start of last (incomplete) superframe... */ | ||
1986 | init_get_bits(gb, avpkt->data, size << 3); | ||
1987 | skip_bits_long(gb, (size << 3) - pos); | ||
1988 | //assert(get_bits_left(gb) == pos); | ||
1989 | |||
1990 | /* ...and cache it for spillover in next packet */ | ||
1991 | init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE); | ||
1992 | copy_bits(&s->pb, avpkt->data, size, gb, s->sframe_cache_size); | ||
1993 | // FIXME bad - just copy bytes as whole and add use the | ||
1994 | // skip_bits_next field | ||
1995 | } | ||
1996 | |||
1997 | return size; | ||
1998 | } | ||
1999 | |||
2000 | static av_cold int wmavoice_decode_end(AVCodecContext *ctx) | ||
2001 | { | ||
2002 | WMAVoiceContext *s = ctx->priv_data; | ||
2003 | |||
2004 | if (s->do_apf) { | ||
2005 | ff_rdft_end(&s->rdft); | ||
2006 | ff_rdft_end(&s->irdft); | ||
2007 | ff_dct_end(&s->dct); | ||
2008 | ff_dct_end(&s->dst); | ||
2009 | } | ||
2010 | |||
2011 | return 0; | ||
2012 | } | ||
2013 | |||
2014 | static av_cold void wmavoice_flush(AVCodecContext *ctx) | ||
2015 | { | ||
2016 | WMAVoiceContext *s = ctx->priv_data; | ||
2017 | int n; | ||
2018 | |||
2019 | s->postfilter_agc = 0; | ||
2020 | s->sframe_cache_size = 0; | ||
2021 | s->skip_bits_next = 0; | ||
2022 | for (n = 0; n < s->lsps; n++) | ||
2023 | s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0); | ||
2024 | memset(s->excitation_history, 0, | ||
2025 | sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY); | ||
2026 | memset(s->synth_history, 0, | ||
2027 | sizeof(*s->synth_history) * MAX_LSPS); | ||
2028 | memset(s->gain_pred_err, 0, | ||
2029 | sizeof(s->gain_pred_err)); | ||
2030 | |||
2031 | if (s->do_apf) { | ||
2032 | memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0, | ||
2033 | sizeof(*s->synth_filter_out_buf) * s->lsps); | ||
2034 | memset(s->dcf_mem, 0, | ||
2035 | sizeof(*s->dcf_mem) * 2); | ||
2036 | memset(s->zero_exc_pf, 0, | ||
2037 | sizeof(*s->zero_exc_pf) * s->history_nsamples); | ||
2038 | memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache)); | ||
2039 | } | ||
2040 | } | ||
2041 | #if 0 | ||
2042 | AVCodec wmavoice_decoder = { | ||
2043 | "wmavoice", | ||
2044 | AVMEDIA_TYPE_AUDIO, | ||
2045 | CODEC_ID_WMAVOICE, | ||
2046 | sizeof(WMAVoiceContext), | ||
2047 | wmavoice_decode_init, | ||
2048 | NULL, | ||
2049 | wmavoice_decode_end, | ||
2050 | wmavoice_decode_packet, | ||
2051 | CODEC_CAP_SUBFRAMES, | ||
2052 | .flush = wmavoice_flush, | ||
2053 | .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"), | ||
2054 | }; | ||
2055 | #endif | ||
2056 | |||
2057 | int main(void) | ||
2058 | { | ||
2059 | return 0; | ||
2060 | } | ||