cannam@154
|
1 /* Copyright (c) 2007-2008 CSIRO
|
cannam@154
|
2 Copyright (c) 2007-2010 Xiph.Org Foundation
|
cannam@154
|
3 Copyright (c) 2008 Gregory Maxwell
|
cannam@154
|
4 Written by Jean-Marc Valin and Gregory Maxwell */
|
cannam@154
|
5 /*
|
cannam@154
|
6 Redistribution and use in source and binary forms, with or without
|
cannam@154
|
7 modification, are permitted provided that the following conditions
|
cannam@154
|
8 are met:
|
cannam@154
|
9
|
cannam@154
|
10 - Redistributions of source code must retain the above copyright
|
cannam@154
|
11 notice, this list of conditions and the following disclaimer.
|
cannam@154
|
12
|
cannam@154
|
13 - Redistributions in binary form must reproduce the above copyright
|
cannam@154
|
14 notice, this list of conditions and the following disclaimer in the
|
cannam@154
|
15 documentation and/or other materials provided with the distribution.
|
cannam@154
|
16
|
cannam@154
|
17 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
|
cannam@154
|
18 ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
|
cannam@154
|
19 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
|
cannam@154
|
20 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
|
cannam@154
|
21 OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
cannam@154
|
22 EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
cannam@154
|
23 PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
cannam@154
|
24 PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
|
cannam@154
|
25 LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
|
cannam@154
|
26 NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
|
cannam@154
|
27 SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
cannam@154
|
28 */
|
cannam@154
|
29
|
cannam@154
|
30 #ifdef HAVE_CONFIG_H
|
cannam@154
|
31 #include "config.h"
|
cannam@154
|
32 #endif
|
cannam@154
|
33
|
cannam@154
|
34 #define CELT_ENCODER_C
|
cannam@154
|
35
|
cannam@154
|
36 #include "cpu_support.h"
|
cannam@154
|
37 #include "os_support.h"
|
cannam@154
|
38 #include "mdct.h"
|
cannam@154
|
39 #include <math.h>
|
cannam@154
|
40 #include "celt.h"
|
cannam@154
|
41 #include "pitch.h"
|
cannam@154
|
42 #include "bands.h"
|
cannam@154
|
43 #include "modes.h"
|
cannam@154
|
44 #include "entcode.h"
|
cannam@154
|
45 #include "quant_bands.h"
|
cannam@154
|
46 #include "rate.h"
|
cannam@154
|
47 #include "stack_alloc.h"
|
cannam@154
|
48 #include "mathops.h"
|
cannam@154
|
49 #include "float_cast.h"
|
cannam@154
|
50 #include <stdarg.h>
|
cannam@154
|
51 #include "celt_lpc.h"
|
cannam@154
|
52 #include "vq.h"
|
cannam@154
|
53
|
cannam@154
|
54
|
cannam@154
|
55 /** Encoder state
|
cannam@154
|
56 @brief Encoder state
|
cannam@154
|
57 */
|
cannam@154
|
58 struct OpusCustomEncoder {
|
cannam@154
|
59 const OpusCustomMode *mode; /**< Mode used by the encoder */
|
cannam@154
|
60 int channels;
|
cannam@154
|
61 int stream_channels;
|
cannam@154
|
62
|
cannam@154
|
63 int force_intra;
|
cannam@154
|
64 int clip;
|
cannam@154
|
65 int disable_pf;
|
cannam@154
|
66 int complexity;
|
cannam@154
|
67 int upsample;
|
cannam@154
|
68 int start, end;
|
cannam@154
|
69
|
cannam@154
|
70 opus_int32 bitrate;
|
cannam@154
|
71 int vbr;
|
cannam@154
|
72 int signalling;
|
cannam@154
|
73 int constrained_vbr; /* If zero, VBR can do whatever it likes with the rate */
|
cannam@154
|
74 int loss_rate;
|
cannam@154
|
75 int lsb_depth;
|
cannam@154
|
76 int lfe;
|
cannam@154
|
77 int disable_inv;
|
cannam@154
|
78 int arch;
|
cannam@154
|
79
|
cannam@154
|
80 /* Everything beyond this point gets cleared on a reset */
|
cannam@154
|
81 #define ENCODER_RESET_START rng
|
cannam@154
|
82
|
cannam@154
|
83 opus_uint32 rng;
|
cannam@154
|
84 int spread_decision;
|
cannam@154
|
85 opus_val32 delayedIntra;
|
cannam@154
|
86 int tonal_average;
|
cannam@154
|
87 int lastCodedBands;
|
cannam@154
|
88 int hf_average;
|
cannam@154
|
89 int tapset_decision;
|
cannam@154
|
90
|
cannam@154
|
91 int prefilter_period;
|
cannam@154
|
92 opus_val16 prefilter_gain;
|
cannam@154
|
93 int prefilter_tapset;
|
cannam@154
|
94 #ifdef RESYNTH
|
cannam@154
|
95 int prefilter_period_old;
|
cannam@154
|
96 opus_val16 prefilter_gain_old;
|
cannam@154
|
97 int prefilter_tapset_old;
|
cannam@154
|
98 #endif
|
cannam@154
|
99 int consec_transient;
|
cannam@154
|
100 AnalysisInfo analysis;
|
cannam@154
|
101 SILKInfo silk_info;
|
cannam@154
|
102
|
cannam@154
|
103 opus_val32 preemph_memE[2];
|
cannam@154
|
104 opus_val32 preemph_memD[2];
|
cannam@154
|
105
|
cannam@154
|
106 /* VBR-related parameters */
|
cannam@154
|
107 opus_int32 vbr_reservoir;
|
cannam@154
|
108 opus_int32 vbr_drift;
|
cannam@154
|
109 opus_int32 vbr_offset;
|
cannam@154
|
110 opus_int32 vbr_count;
|
cannam@154
|
111 opus_val32 overlap_max;
|
cannam@154
|
112 opus_val16 stereo_saving;
|
cannam@154
|
113 int intensity;
|
cannam@154
|
114 opus_val16 *energy_mask;
|
cannam@154
|
115 opus_val16 spec_avg;
|
cannam@154
|
116
|
cannam@154
|
117 #ifdef RESYNTH
|
cannam@154
|
118 /* +MAX_PERIOD/2 to make space for overlap */
|
cannam@154
|
119 celt_sig syn_mem[2][2*MAX_PERIOD+MAX_PERIOD/2];
|
cannam@154
|
120 #endif
|
cannam@154
|
121
|
cannam@154
|
122 celt_sig in_mem[1]; /* Size = channels*mode->overlap */
|
cannam@154
|
123 /* celt_sig prefilter_mem[], Size = channels*COMBFILTER_MAXPERIOD */
|
cannam@154
|
124 /* opus_val16 oldBandE[], Size = channels*mode->nbEBands */
|
cannam@154
|
125 /* opus_val16 oldLogE[], Size = channels*mode->nbEBands */
|
cannam@154
|
126 /* opus_val16 oldLogE2[], Size = channels*mode->nbEBands */
|
cannam@154
|
127 /* opus_val16 energyError[], Size = channels*mode->nbEBands */
|
cannam@154
|
128 };
|
cannam@154
|
129
|
cannam@154
|
130 int celt_encoder_get_size(int channels)
|
cannam@154
|
131 {
|
cannam@154
|
132 CELTMode *mode = opus_custom_mode_create(48000, 960, NULL);
|
cannam@154
|
133 return opus_custom_encoder_get_size(mode, channels);
|
cannam@154
|
134 }
|
cannam@154
|
135
|
cannam@154
|
136 OPUS_CUSTOM_NOSTATIC int opus_custom_encoder_get_size(const CELTMode *mode, int channels)
|
cannam@154
|
137 {
|
cannam@154
|
138 int size = sizeof(struct CELTEncoder)
|
cannam@154
|
139 + (channels*mode->overlap-1)*sizeof(celt_sig) /* celt_sig in_mem[channels*mode->overlap]; */
|
cannam@154
|
140 + channels*COMBFILTER_MAXPERIOD*sizeof(celt_sig) /* celt_sig prefilter_mem[channels*COMBFILTER_MAXPERIOD]; */
|
cannam@154
|
141 + 4*channels*mode->nbEBands*sizeof(opus_val16); /* opus_val16 oldBandE[channels*mode->nbEBands]; */
|
cannam@154
|
142 /* opus_val16 oldLogE[channels*mode->nbEBands]; */
|
cannam@154
|
143 /* opus_val16 oldLogE2[channels*mode->nbEBands]; */
|
cannam@154
|
144 /* opus_val16 energyError[channels*mode->nbEBands]; */
|
cannam@154
|
145 return size;
|
cannam@154
|
146 }
|
cannam@154
|
147
|
cannam@154
|
148 #ifdef CUSTOM_MODES
|
cannam@154
|
149 CELTEncoder *opus_custom_encoder_create(const CELTMode *mode, int channels, int *error)
|
cannam@154
|
150 {
|
cannam@154
|
151 int ret;
|
cannam@154
|
152 CELTEncoder *st = (CELTEncoder *)opus_alloc(opus_custom_encoder_get_size(mode, channels));
|
cannam@154
|
153 /* init will handle the NULL case */
|
cannam@154
|
154 ret = opus_custom_encoder_init(st, mode, channels);
|
cannam@154
|
155 if (ret != OPUS_OK)
|
cannam@154
|
156 {
|
cannam@154
|
157 opus_custom_encoder_destroy(st);
|
cannam@154
|
158 st = NULL;
|
cannam@154
|
159 }
|
cannam@154
|
160 if (error)
|
cannam@154
|
161 *error = ret;
|
cannam@154
|
162 return st;
|
cannam@154
|
163 }
|
cannam@154
|
164 #endif /* CUSTOM_MODES */
|
cannam@154
|
165
|
cannam@154
|
166 static int opus_custom_encoder_init_arch(CELTEncoder *st, const CELTMode *mode,
|
cannam@154
|
167 int channels, int arch)
|
cannam@154
|
168 {
|
cannam@154
|
169 if (channels < 0 || channels > 2)
|
cannam@154
|
170 return OPUS_BAD_ARG;
|
cannam@154
|
171
|
cannam@154
|
172 if (st==NULL || mode==NULL)
|
cannam@154
|
173 return OPUS_ALLOC_FAIL;
|
cannam@154
|
174
|
cannam@154
|
175 OPUS_CLEAR((char*)st, opus_custom_encoder_get_size(mode, channels));
|
cannam@154
|
176
|
cannam@154
|
177 st->mode = mode;
|
cannam@154
|
178 st->stream_channels = st->channels = channels;
|
cannam@154
|
179
|
cannam@154
|
180 st->upsample = 1;
|
cannam@154
|
181 st->start = 0;
|
cannam@154
|
182 st->end = st->mode->effEBands;
|
cannam@154
|
183 st->signalling = 1;
|
cannam@154
|
184 st->arch = arch;
|
cannam@154
|
185
|
cannam@154
|
186 st->constrained_vbr = 1;
|
cannam@154
|
187 st->clip = 1;
|
cannam@154
|
188
|
cannam@154
|
189 st->bitrate = OPUS_BITRATE_MAX;
|
cannam@154
|
190 st->vbr = 0;
|
cannam@154
|
191 st->force_intra = 0;
|
cannam@154
|
192 st->complexity = 5;
|
cannam@154
|
193 st->lsb_depth=24;
|
cannam@154
|
194
|
cannam@154
|
195 opus_custom_encoder_ctl(st, OPUS_RESET_STATE);
|
cannam@154
|
196
|
cannam@154
|
197 return OPUS_OK;
|
cannam@154
|
198 }
|
cannam@154
|
199
|
cannam@154
|
200 #ifdef CUSTOM_MODES
|
cannam@154
|
201 int opus_custom_encoder_init(CELTEncoder *st, const CELTMode *mode, int channels)
|
cannam@154
|
202 {
|
cannam@154
|
203 return opus_custom_encoder_init_arch(st, mode, channels, opus_select_arch());
|
cannam@154
|
204 }
|
cannam@154
|
205 #endif
|
cannam@154
|
206
|
cannam@154
|
207 int celt_encoder_init(CELTEncoder *st, opus_int32 sampling_rate, int channels,
|
cannam@154
|
208 int arch)
|
cannam@154
|
209 {
|
cannam@154
|
210 int ret;
|
cannam@154
|
211 ret = opus_custom_encoder_init_arch(st,
|
cannam@154
|
212 opus_custom_mode_create(48000, 960, NULL), channels, arch);
|
cannam@154
|
213 if (ret != OPUS_OK)
|
cannam@154
|
214 return ret;
|
cannam@154
|
215 st->upsample = resampling_factor(sampling_rate);
|
cannam@154
|
216 return OPUS_OK;
|
cannam@154
|
217 }
|
cannam@154
|
218
|
cannam@154
|
219 #ifdef CUSTOM_MODES
|
cannam@154
|
220 void opus_custom_encoder_destroy(CELTEncoder *st)
|
cannam@154
|
221 {
|
cannam@154
|
222 opus_free(st);
|
cannam@154
|
223 }
|
cannam@154
|
224 #endif /* CUSTOM_MODES */
|
cannam@154
|
225
|
cannam@154
|
226
|
cannam@154
|
227 static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int C,
|
cannam@154
|
228 opus_val16 *tf_estimate, int *tf_chan, int allow_weak_transients,
|
cannam@154
|
229 int *weak_transient)
|
cannam@154
|
230 {
|
cannam@154
|
231 int i;
|
cannam@154
|
232 VARDECL(opus_val16, tmp);
|
cannam@154
|
233 opus_val32 mem0,mem1;
|
cannam@154
|
234 int is_transient = 0;
|
cannam@154
|
235 opus_int32 mask_metric = 0;
|
cannam@154
|
236 int c;
|
cannam@154
|
237 opus_val16 tf_max;
|
cannam@154
|
238 int len2;
|
cannam@154
|
239 /* Forward masking: 6.7 dB/ms. */
|
cannam@154
|
240 #ifdef FIXED_POINT
|
cannam@154
|
241 int forward_shift = 4;
|
cannam@154
|
242 #else
|
cannam@154
|
243 opus_val16 forward_decay = QCONST16(.0625f,15);
|
cannam@154
|
244 #endif
|
cannam@154
|
245 /* Table of 6*64/x, trained on real data to minimize the average error */
|
cannam@154
|
246 static const unsigned char inv_table[128] = {
|
cannam@154
|
247 255,255,156,110, 86, 70, 59, 51, 45, 40, 37, 33, 31, 28, 26, 25,
|
cannam@154
|
248 23, 22, 21, 20, 19, 18, 17, 16, 16, 15, 15, 14, 13, 13, 12, 12,
|
cannam@154
|
249 12, 12, 11, 11, 11, 10, 10, 10, 9, 9, 9, 9, 9, 9, 8, 8,
|
cannam@154
|
250 8, 8, 8, 7, 7, 7, 7, 7, 7, 6, 6, 6, 6, 6, 6, 6,
|
cannam@154
|
251 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5, 5, 5, 5,
|
cannam@154
|
252 5, 5, 5, 5, 5, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4,
|
cannam@154
|
253 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 3, 3,
|
cannam@154
|
254 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 2,
|
cannam@154
|
255 };
|
cannam@154
|
256 SAVE_STACK;
|
cannam@154
|
257 ALLOC(tmp, len, opus_val16);
|
cannam@154
|
258
|
cannam@154
|
259 *weak_transient = 0;
|
cannam@154
|
260 /* For lower bitrates, let's be more conservative and have a forward masking
|
cannam@154
|
261 decay of 3.3 dB/ms. This avoids having to code transients at very low
|
cannam@154
|
262 bitrate (mostly for hybrid), which can result in unstable energy and/or
|
cannam@154
|
263 partial collapse. */
|
cannam@154
|
264 if (allow_weak_transients)
|
cannam@154
|
265 {
|
cannam@154
|
266 #ifdef FIXED_POINT
|
cannam@154
|
267 forward_shift = 5;
|
cannam@154
|
268 #else
|
cannam@154
|
269 forward_decay = QCONST16(.03125f,15);
|
cannam@154
|
270 #endif
|
cannam@154
|
271 }
|
cannam@154
|
272 len2=len/2;
|
cannam@154
|
273 for (c=0;c<C;c++)
|
cannam@154
|
274 {
|
cannam@154
|
275 opus_val32 mean;
|
cannam@154
|
276 opus_int32 unmask=0;
|
cannam@154
|
277 opus_val32 norm;
|
cannam@154
|
278 opus_val16 maxE;
|
cannam@154
|
279 mem0=0;
|
cannam@154
|
280 mem1=0;
|
cannam@154
|
281 /* High-pass filter: (1 - 2*z^-1 + z^-2) / (1 - z^-1 + .5*z^-2) */
|
cannam@154
|
282 for (i=0;i<len;i++)
|
cannam@154
|
283 {
|
cannam@154
|
284 opus_val32 x,y;
|
cannam@154
|
285 x = SHR32(in[i+c*len],SIG_SHIFT);
|
cannam@154
|
286 y = ADD32(mem0, x);
|
cannam@154
|
287 #ifdef FIXED_POINT
|
cannam@154
|
288 mem0 = mem1 + y - SHL32(x,1);
|
cannam@154
|
289 mem1 = x - SHR32(y,1);
|
cannam@154
|
290 #else
|
cannam@154
|
291 mem0 = mem1 + y - 2*x;
|
cannam@154
|
292 mem1 = x - .5f*y;
|
cannam@154
|
293 #endif
|
cannam@154
|
294 tmp[i] = SROUND16(y, 2);
|
cannam@154
|
295 /*printf("%f ", tmp[i]);*/
|
cannam@154
|
296 }
|
cannam@154
|
297 /*printf("\n");*/
|
cannam@154
|
298 /* First few samples are bad because we don't propagate the memory */
|
cannam@154
|
299 OPUS_CLEAR(tmp, 12);
|
cannam@154
|
300
|
cannam@154
|
301 #ifdef FIXED_POINT
|
cannam@154
|
302 /* Normalize tmp to max range */
|
cannam@154
|
303 {
|
cannam@154
|
304 int shift=0;
|
cannam@154
|
305 shift = 14-celt_ilog2(MAX16(1, celt_maxabs16(tmp, len)));
|
cannam@154
|
306 if (shift!=0)
|
cannam@154
|
307 {
|
cannam@154
|
308 for (i=0;i<len;i++)
|
cannam@154
|
309 tmp[i] = SHL16(tmp[i], shift);
|
cannam@154
|
310 }
|
cannam@154
|
311 }
|
cannam@154
|
312 #endif
|
cannam@154
|
313
|
cannam@154
|
314 mean=0;
|
cannam@154
|
315 mem0=0;
|
cannam@154
|
316 /* Grouping by two to reduce complexity */
|
cannam@154
|
317 /* Forward pass to compute the post-echo threshold*/
|
cannam@154
|
318 for (i=0;i<len2;i++)
|
cannam@154
|
319 {
|
cannam@154
|
320 opus_val16 x2 = PSHR32(MULT16_16(tmp[2*i],tmp[2*i]) + MULT16_16(tmp[2*i+1],tmp[2*i+1]),16);
|
cannam@154
|
321 mean += x2;
|
cannam@154
|
322 #ifdef FIXED_POINT
|
cannam@154
|
323 /* FIXME: Use PSHR16() instead */
|
cannam@154
|
324 tmp[i] = mem0 + PSHR32(x2-mem0,forward_shift);
|
cannam@154
|
325 #else
|
cannam@154
|
326 tmp[i] = mem0 + MULT16_16_P15(forward_decay,x2-mem0);
|
cannam@154
|
327 #endif
|
cannam@154
|
328 mem0 = tmp[i];
|
cannam@154
|
329 }
|
cannam@154
|
330
|
cannam@154
|
331 mem0=0;
|
cannam@154
|
332 maxE=0;
|
cannam@154
|
333 /* Backward pass to compute the pre-echo threshold */
|
cannam@154
|
334 for (i=len2-1;i>=0;i--)
|
cannam@154
|
335 {
|
cannam@154
|
336 /* Backward masking: 13.9 dB/ms. */
|
cannam@154
|
337 #ifdef FIXED_POINT
|
cannam@154
|
338 /* FIXME: Use PSHR16() instead */
|
cannam@154
|
339 tmp[i] = mem0 + PSHR32(tmp[i]-mem0,3);
|
cannam@154
|
340 #else
|
cannam@154
|
341 tmp[i] = mem0 + MULT16_16_P15(QCONST16(0.125f,15),tmp[i]-mem0);
|
cannam@154
|
342 #endif
|
cannam@154
|
343 mem0 = tmp[i];
|
cannam@154
|
344 maxE = MAX16(maxE, mem0);
|
cannam@154
|
345 }
|
cannam@154
|
346 /*for (i=0;i<len2;i++)printf("%f ", tmp[i]/mean);printf("\n");*/
|
cannam@154
|
347
|
cannam@154
|
348 /* Compute the ratio of the "frame energy" over the harmonic mean of the energy.
|
cannam@154
|
349 This essentially corresponds to a bitrate-normalized temporal noise-to-mask
|
cannam@154
|
350 ratio */
|
cannam@154
|
351
|
cannam@154
|
352 /* As a compromise with the old transient detector, frame energy is the
|
cannam@154
|
353 geometric mean of the energy and half the max */
|
cannam@154
|
354 #ifdef FIXED_POINT
|
cannam@154
|
355 /* Costs two sqrt() to avoid overflows */
|
cannam@154
|
356 mean = MULT16_16(celt_sqrt(mean), celt_sqrt(MULT16_16(maxE,len2>>1)));
|
cannam@154
|
357 #else
|
cannam@154
|
358 mean = celt_sqrt(mean * maxE*.5*len2);
|
cannam@154
|
359 #endif
|
cannam@154
|
360 /* Inverse of the mean energy in Q15+6 */
|
cannam@154
|
361 norm = SHL32(EXTEND32(len2),6+14)/ADD32(EPSILON,SHR32(mean,1));
|
cannam@154
|
362 /* Compute harmonic mean discarding the unreliable boundaries
|
cannam@154
|
363 The data is smooth, so we only take 1/4th of the samples */
|
cannam@154
|
364 unmask=0;
|
cannam@154
|
365 /* We should never see NaNs here. If we find any, then something really bad happened and we better abort
|
cannam@154
|
366 before it does any damage later on. If these asserts are disabled (no hardening), then the table
|
cannam@154
|
367 lookup a few lines below (id = ...) is likely to crash dur to an out-of-bounds read. DO NOT FIX
|
cannam@154
|
368 that crash on NaN since it could result in a worse issue later on. */
|
cannam@154
|
369 celt_assert(!celt_isnan(tmp[0]));
|
cannam@154
|
370 celt_assert(!celt_isnan(norm));
|
cannam@154
|
371 for (i=12;i<len2-5;i+=4)
|
cannam@154
|
372 {
|
cannam@154
|
373 int id;
|
cannam@154
|
374 #ifdef FIXED_POINT
|
cannam@154
|
375 id = MAX32(0,MIN32(127,MULT16_32_Q15(tmp[i]+EPSILON,norm))); /* Do not round to nearest */
|
cannam@154
|
376 #else
|
cannam@154
|
377 id = (int)MAX32(0,MIN32(127,floor(64*norm*(tmp[i]+EPSILON)))); /* Do not round to nearest */
|
cannam@154
|
378 #endif
|
cannam@154
|
379 unmask += inv_table[id];
|
cannam@154
|
380 }
|
cannam@154
|
381 /*printf("%d\n", unmask);*/
|
cannam@154
|
382 /* Normalize, compensate for the 1/4th of the sample and the factor of 6 in the inverse table */
|
cannam@154
|
383 unmask = 64*unmask*4/(6*(len2-17));
|
cannam@154
|
384 if (unmask>mask_metric)
|
cannam@154
|
385 {
|
cannam@154
|
386 *tf_chan = c;
|
cannam@154
|
387 mask_metric = unmask;
|
cannam@154
|
388 }
|
cannam@154
|
389 }
|
cannam@154
|
390 is_transient = mask_metric>200;
|
cannam@154
|
391 /* For low bitrates, define "weak transients" that need to be
|
cannam@154
|
392 handled differently to avoid partial collapse. */
|
cannam@154
|
393 if (allow_weak_transients && is_transient && mask_metric<600) {
|
cannam@154
|
394 is_transient = 0;
|
cannam@154
|
395 *weak_transient = 1;
|
cannam@154
|
396 }
|
cannam@154
|
397 /* Arbitrary metric for VBR boost */
|
cannam@154
|
398 tf_max = MAX16(0,celt_sqrt(27*mask_metric)-42);
|
cannam@154
|
399 /* *tf_estimate = 1 + MIN16(1, sqrt(MAX16(0, tf_max-30))/20); */
|
cannam@154
|
400 *tf_estimate = celt_sqrt(MAX32(0, SHL32(MULT16_16(QCONST16(0.0069,14),MIN16(163,tf_max)),14)-QCONST32(0.139,28)));
|
cannam@154
|
401 /*printf("%d %f\n", tf_max, mask_metric);*/
|
cannam@154
|
402 RESTORE_STACK;
|
cannam@154
|
403 #ifdef FUZZING
|
cannam@154
|
404 is_transient = rand()&0x1;
|
cannam@154
|
405 #endif
|
cannam@154
|
406 /*printf("%d %f %d\n", is_transient, (float)*tf_estimate, tf_max);*/
|
cannam@154
|
407 return is_transient;
|
cannam@154
|
408 }
|
cannam@154
|
409
|
cannam@154
|
410 /* Looks for sudden increases of energy to decide whether we need to patch
|
cannam@154
|
411 the transient decision */
|
cannam@154
|
412 static int patch_transient_decision(opus_val16 *newE, opus_val16 *oldE, int nbEBands,
|
cannam@154
|
413 int start, int end, int C)
|
cannam@154
|
414 {
|
cannam@154
|
415 int i, c;
|
cannam@154
|
416 opus_val32 mean_diff=0;
|
cannam@154
|
417 opus_val16 spread_old[26];
|
cannam@154
|
418 /* Apply an aggressive (-6 dB/Bark) spreading function to the old frame to
|
cannam@154
|
419 avoid false detection caused by irrelevant bands */
|
cannam@154
|
420 if (C==1)
|
cannam@154
|
421 {
|
cannam@154
|
422 spread_old[start] = oldE[start];
|
cannam@154
|
423 for (i=start+1;i<end;i++)
|
cannam@154
|
424 spread_old[i] = MAX16(spread_old[i-1]-QCONST16(1.0f, DB_SHIFT), oldE[i]);
|
cannam@154
|
425 } else {
|
cannam@154
|
426 spread_old[start] = MAX16(oldE[start],oldE[start+nbEBands]);
|
cannam@154
|
427 for (i=start+1;i<end;i++)
|
cannam@154
|
428 spread_old[i] = MAX16(spread_old[i-1]-QCONST16(1.0f, DB_SHIFT),
|
cannam@154
|
429 MAX16(oldE[i],oldE[i+nbEBands]));
|
cannam@154
|
430 }
|
cannam@154
|
431 for (i=end-2;i>=start;i--)
|
cannam@154
|
432 spread_old[i] = MAX16(spread_old[i], spread_old[i+1]-QCONST16(1.0f, DB_SHIFT));
|
cannam@154
|
433 /* Compute mean increase */
|
cannam@154
|
434 c=0; do {
|
cannam@154
|
435 for (i=IMAX(2,start);i<end-1;i++)
|
cannam@154
|
436 {
|
cannam@154
|
437 opus_val16 x1, x2;
|
cannam@154
|
438 x1 = MAX16(0, newE[i + c*nbEBands]);
|
cannam@154
|
439 x2 = MAX16(0, spread_old[i]);
|
cannam@154
|
440 mean_diff = ADD32(mean_diff, EXTEND32(MAX16(0, SUB16(x1, x2))));
|
cannam@154
|
441 }
|
cannam@154
|
442 } while (++c<C);
|
cannam@154
|
443 mean_diff = DIV32(mean_diff, C*(end-1-IMAX(2,start)));
|
cannam@154
|
444 /*printf("%f %f %d\n", mean_diff, max_diff, count);*/
|
cannam@154
|
445 return mean_diff > QCONST16(1.f, DB_SHIFT);
|
cannam@154
|
446 }
|
cannam@154
|
447
|
cannam@154
|
448 /** Apply window and compute the MDCT for all sub-frames and
|
cannam@154
|
449 all channels in a frame */
|
cannam@154
|
450 static void compute_mdcts(const CELTMode *mode, int shortBlocks, celt_sig * OPUS_RESTRICT in,
|
cannam@154
|
451 celt_sig * OPUS_RESTRICT out, int C, int CC, int LM, int upsample,
|
cannam@154
|
452 int arch)
|
cannam@154
|
453 {
|
cannam@154
|
454 const int overlap = mode->overlap;
|
cannam@154
|
455 int N;
|
cannam@154
|
456 int B;
|
cannam@154
|
457 int shift;
|
cannam@154
|
458 int i, b, c;
|
cannam@154
|
459 if (shortBlocks)
|
cannam@154
|
460 {
|
cannam@154
|
461 B = shortBlocks;
|
cannam@154
|
462 N = mode->shortMdctSize;
|
cannam@154
|
463 shift = mode->maxLM;
|
cannam@154
|
464 } else {
|
cannam@154
|
465 B = 1;
|
cannam@154
|
466 N = mode->shortMdctSize<<LM;
|
cannam@154
|
467 shift = mode->maxLM-LM;
|
cannam@154
|
468 }
|
cannam@154
|
469 c=0; do {
|
cannam@154
|
470 for (b=0;b<B;b++)
|
cannam@154
|
471 {
|
cannam@154
|
472 /* Interleaving the sub-frames while doing the MDCTs */
|
cannam@154
|
473 clt_mdct_forward(&mode->mdct, in+c*(B*N+overlap)+b*N,
|
cannam@154
|
474 &out[b+c*N*B], mode->window, overlap, shift, B,
|
cannam@154
|
475 arch);
|
cannam@154
|
476 }
|
cannam@154
|
477 } while (++c<CC);
|
cannam@154
|
478 if (CC==2&&C==1)
|
cannam@154
|
479 {
|
cannam@154
|
480 for (i=0;i<B*N;i++)
|
cannam@154
|
481 out[i] = ADD32(HALF32(out[i]), HALF32(out[B*N+i]));
|
cannam@154
|
482 }
|
cannam@154
|
483 if (upsample != 1)
|
cannam@154
|
484 {
|
cannam@154
|
485 c=0; do
|
cannam@154
|
486 {
|
cannam@154
|
487 int bound = B*N/upsample;
|
cannam@154
|
488 for (i=0;i<bound;i++)
|
cannam@154
|
489 out[c*B*N+i] *= upsample;
|
cannam@154
|
490 OPUS_CLEAR(&out[c*B*N+bound], B*N-bound);
|
cannam@154
|
491 } while (++c<C);
|
cannam@154
|
492 }
|
cannam@154
|
493 }
|
cannam@154
|
494
|
cannam@154
|
495
|
cannam@154
|
496 void celt_preemphasis(const opus_val16 * OPUS_RESTRICT pcmp, celt_sig * OPUS_RESTRICT inp,
|
cannam@154
|
497 int N, int CC, int upsample, const opus_val16 *coef, celt_sig *mem, int clip)
|
cannam@154
|
498 {
|
cannam@154
|
499 int i;
|
cannam@154
|
500 opus_val16 coef0;
|
cannam@154
|
501 celt_sig m;
|
cannam@154
|
502 int Nu;
|
cannam@154
|
503
|
cannam@154
|
504 coef0 = coef[0];
|
cannam@154
|
505 m = *mem;
|
cannam@154
|
506
|
cannam@154
|
507 /* Fast path for the normal 48kHz case and no clipping */
|
cannam@154
|
508 if (coef[1] == 0 && upsample == 1 && !clip)
|
cannam@154
|
509 {
|
cannam@154
|
510 for (i=0;i<N;i++)
|
cannam@154
|
511 {
|
cannam@154
|
512 opus_val16 x;
|
cannam@154
|
513 x = SCALEIN(pcmp[CC*i]);
|
cannam@154
|
514 /* Apply pre-emphasis */
|
cannam@154
|
515 inp[i] = SHL32(x, SIG_SHIFT) - m;
|
cannam@154
|
516 m = SHR32(MULT16_16(coef0, x), 15-SIG_SHIFT);
|
cannam@154
|
517 }
|
cannam@154
|
518 *mem = m;
|
cannam@154
|
519 return;
|
cannam@154
|
520 }
|
cannam@154
|
521
|
cannam@154
|
522 Nu = N/upsample;
|
cannam@154
|
523 if (upsample!=1)
|
cannam@154
|
524 {
|
cannam@154
|
525 OPUS_CLEAR(inp, N);
|
cannam@154
|
526 }
|
cannam@154
|
527 for (i=0;i<Nu;i++)
|
cannam@154
|
528 inp[i*upsample] = SCALEIN(pcmp[CC*i]);
|
cannam@154
|
529
|
cannam@154
|
530 #ifndef FIXED_POINT
|
cannam@154
|
531 if (clip)
|
cannam@154
|
532 {
|
cannam@154
|
533 /* Clip input to avoid encoding non-portable files */
|
cannam@154
|
534 for (i=0;i<Nu;i++)
|
cannam@154
|
535 inp[i*upsample] = MAX32(-65536.f, MIN32(65536.f,inp[i*upsample]));
|
cannam@154
|
536 }
|
cannam@154
|
537 #else
|
cannam@154
|
538 (void)clip; /* Avoids a warning about clip being unused. */
|
cannam@154
|
539 #endif
|
cannam@154
|
540 #ifdef CUSTOM_MODES
|
cannam@154
|
541 if (coef[1] != 0)
|
cannam@154
|
542 {
|
cannam@154
|
543 opus_val16 coef1 = coef[1];
|
cannam@154
|
544 opus_val16 coef2 = coef[2];
|
cannam@154
|
545 for (i=0;i<N;i++)
|
cannam@154
|
546 {
|
cannam@154
|
547 celt_sig x, tmp;
|
cannam@154
|
548 x = inp[i];
|
cannam@154
|
549 /* Apply pre-emphasis */
|
cannam@154
|
550 tmp = MULT16_16(coef2, x);
|
cannam@154
|
551 inp[i] = tmp + m;
|
cannam@154
|
552 m = MULT16_32_Q15(coef1, inp[i]) - MULT16_32_Q15(coef0, tmp);
|
cannam@154
|
553 }
|
cannam@154
|
554 } else
|
cannam@154
|
555 #endif
|
cannam@154
|
556 {
|
cannam@154
|
557 for (i=0;i<N;i++)
|
cannam@154
|
558 {
|
cannam@154
|
559 opus_val16 x;
|
cannam@154
|
560 x = inp[i];
|
cannam@154
|
561 /* Apply pre-emphasis */
|
cannam@154
|
562 inp[i] = SHL32(x, SIG_SHIFT) - m;
|
cannam@154
|
563 m = SHR32(MULT16_16(coef0, x), 15-SIG_SHIFT);
|
cannam@154
|
564 }
|
cannam@154
|
565 }
|
cannam@154
|
566 *mem = m;
|
cannam@154
|
567 }
|
cannam@154
|
568
|
cannam@154
|
569
|
cannam@154
|
570
|
cannam@154
|
571 static opus_val32 l1_metric(const celt_norm *tmp, int N, int LM, opus_val16 bias)
|
cannam@154
|
572 {
|
cannam@154
|
573 int i;
|
cannam@154
|
574 opus_val32 L1;
|
cannam@154
|
575 L1 = 0;
|
cannam@154
|
576 for (i=0;i<N;i++)
|
cannam@154
|
577 L1 += EXTEND32(ABS16(tmp[i]));
|
cannam@154
|
578 /* When in doubt, prefer good freq resolution */
|
cannam@154
|
579 L1 = MAC16_32_Q15(L1, LM*bias, L1);
|
cannam@154
|
580 return L1;
|
cannam@154
|
581
|
cannam@154
|
582 }
|
cannam@154
|
583
|
cannam@154
|
584 static int tf_analysis(const CELTMode *m, int len, int isTransient,
|
cannam@154
|
585 int *tf_res, int lambda, celt_norm *X, int N0, int LM,
|
cannam@154
|
586 opus_val16 tf_estimate, int tf_chan, int *importance)
|
cannam@154
|
587 {
|
cannam@154
|
588 int i;
|
cannam@154
|
589 VARDECL(int, metric);
|
cannam@154
|
590 int cost0;
|
cannam@154
|
591 int cost1;
|
cannam@154
|
592 VARDECL(int, path0);
|
cannam@154
|
593 VARDECL(int, path1);
|
cannam@154
|
594 VARDECL(celt_norm, tmp);
|
cannam@154
|
595 VARDECL(celt_norm, tmp_1);
|
cannam@154
|
596 int sel;
|
cannam@154
|
597 int selcost[2];
|
cannam@154
|
598 int tf_select=0;
|
cannam@154
|
599 opus_val16 bias;
|
cannam@154
|
600
|
cannam@154
|
601 SAVE_STACK;
|
cannam@154
|
602 bias = MULT16_16_Q14(QCONST16(.04f,15), MAX16(-QCONST16(.25f,14), QCONST16(.5f,14)-tf_estimate));
|
cannam@154
|
603 /*printf("%f ", bias);*/
|
cannam@154
|
604
|
cannam@154
|
605 ALLOC(metric, len, int);
|
cannam@154
|
606 ALLOC(tmp, (m->eBands[len]-m->eBands[len-1])<<LM, celt_norm);
|
cannam@154
|
607 ALLOC(tmp_1, (m->eBands[len]-m->eBands[len-1])<<LM, celt_norm);
|
cannam@154
|
608 ALLOC(path0, len, int);
|
cannam@154
|
609 ALLOC(path1, len, int);
|
cannam@154
|
610
|
cannam@154
|
611 for (i=0;i<len;i++)
|
cannam@154
|
612 {
|
cannam@154
|
613 int k, N;
|
cannam@154
|
614 int narrow;
|
cannam@154
|
615 opus_val32 L1, best_L1;
|
cannam@154
|
616 int best_level=0;
|
cannam@154
|
617 N = (m->eBands[i+1]-m->eBands[i])<<LM;
|
cannam@154
|
618 /* band is too narrow to be split down to LM=-1 */
|
cannam@154
|
619 narrow = (m->eBands[i+1]-m->eBands[i])==1;
|
cannam@154
|
620 OPUS_COPY(tmp, &X[tf_chan*N0 + (m->eBands[i]<<LM)], N);
|
cannam@154
|
621 /* Just add the right channel if we're in stereo */
|
cannam@154
|
622 /*if (C==2)
|
cannam@154
|
623 for (j=0;j<N;j++)
|
cannam@154
|
624 tmp[j] = ADD16(SHR16(tmp[j], 1),SHR16(X[N0+j+(m->eBands[i]<<LM)], 1));*/
|
cannam@154
|
625 L1 = l1_metric(tmp, N, isTransient ? LM : 0, bias);
|
cannam@154
|
626 best_L1 = L1;
|
cannam@154
|
627 /* Check the -1 case for transients */
|
cannam@154
|
628 if (isTransient && !narrow)
|
cannam@154
|
629 {
|
cannam@154
|
630 OPUS_COPY(tmp_1, tmp, N);
|
cannam@154
|
631 haar1(tmp_1, N>>LM, 1<<LM);
|
cannam@154
|
632 L1 = l1_metric(tmp_1, N, LM+1, bias);
|
cannam@154
|
633 if (L1<best_L1)
|
cannam@154
|
634 {
|
cannam@154
|
635 best_L1 = L1;
|
cannam@154
|
636 best_level = -1;
|
cannam@154
|
637 }
|
cannam@154
|
638 }
|
cannam@154
|
639 /*printf ("%f ", L1);*/
|
cannam@154
|
640 for (k=0;k<LM+!(isTransient||narrow);k++)
|
cannam@154
|
641 {
|
cannam@154
|
642 int B;
|
cannam@154
|
643
|
cannam@154
|
644 if (isTransient)
|
cannam@154
|
645 B = (LM-k-1);
|
cannam@154
|
646 else
|
cannam@154
|
647 B = k+1;
|
cannam@154
|
648
|
cannam@154
|
649 haar1(tmp, N>>k, 1<<k);
|
cannam@154
|
650
|
cannam@154
|
651 L1 = l1_metric(tmp, N, B, bias);
|
cannam@154
|
652
|
cannam@154
|
653 if (L1 < best_L1)
|
cannam@154
|
654 {
|
cannam@154
|
655 best_L1 = L1;
|
cannam@154
|
656 best_level = k+1;
|
cannam@154
|
657 }
|
cannam@154
|
658 }
|
cannam@154
|
659 /*printf ("%d ", isTransient ? LM-best_level : best_level);*/
|
cannam@154
|
660 /* metric is in Q1 to be able to select the mid-point (-0.5) for narrower bands */
|
cannam@154
|
661 if (isTransient)
|
cannam@154
|
662 metric[i] = 2*best_level;
|
cannam@154
|
663 else
|
cannam@154
|
664 metric[i] = -2*best_level;
|
cannam@154
|
665 /* For bands that can't be split to -1, set the metric to the half-way point to avoid
|
cannam@154
|
666 biasing the decision */
|
cannam@154
|
667 if (narrow && (metric[i]==0 || metric[i]==-2*LM))
|
cannam@154
|
668 metric[i]-=1;
|
cannam@154
|
669 /*printf("%d ", metric[i]/2 + (!isTransient)*LM);*/
|
cannam@154
|
670 }
|
cannam@154
|
671 /*printf("\n");*/
|
cannam@154
|
672 /* Search for the optimal tf resolution, including tf_select */
|
cannam@154
|
673 tf_select = 0;
|
cannam@154
|
674 for (sel=0;sel<2;sel++)
|
cannam@154
|
675 {
|
cannam@154
|
676 cost0 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*sel+0]);
|
cannam@154
|
677 cost1 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*sel+1]) + (isTransient ? 0 : lambda);
|
cannam@154
|
678 for (i=1;i<len;i++)
|
cannam@154
|
679 {
|
cannam@154
|
680 int curr0, curr1;
|
cannam@154
|
681 curr0 = IMIN(cost0, cost1 + lambda);
|
cannam@154
|
682 curr1 = IMIN(cost0 + lambda, cost1);
|
cannam@154
|
683 cost0 = curr0 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel+0]);
|
cannam@154
|
684 cost1 = curr1 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel+1]);
|
cannam@154
|
685 }
|
cannam@154
|
686 cost0 = IMIN(cost0, cost1);
|
cannam@154
|
687 selcost[sel]=cost0;
|
cannam@154
|
688 }
|
cannam@154
|
689 /* For now, we're conservative and only allow tf_select=1 for transients.
|
cannam@154
|
690 * If tests confirm it's useful for non-transients, we could allow it. */
|
cannam@154
|
691 if (selcost[1]<selcost[0] && isTransient)
|
cannam@154
|
692 tf_select=1;
|
cannam@154
|
693 cost0 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*tf_select+0]);
|
cannam@154
|
694 cost1 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*tf_select+1]) + (isTransient ? 0 : lambda);
|
cannam@154
|
695 /* Viterbi forward pass */
|
cannam@154
|
696 for (i=1;i<len;i++)
|
cannam@154
|
697 {
|
cannam@154
|
698 int curr0, curr1;
|
cannam@154
|
699 int from0, from1;
|
cannam@154
|
700
|
cannam@154
|
701 from0 = cost0;
|
cannam@154
|
702 from1 = cost1 + lambda;
|
cannam@154
|
703 if (from0 < from1)
|
cannam@154
|
704 {
|
cannam@154
|
705 curr0 = from0;
|
cannam@154
|
706 path0[i]= 0;
|
cannam@154
|
707 } else {
|
cannam@154
|
708 curr0 = from1;
|
cannam@154
|
709 path0[i]= 1;
|
cannam@154
|
710 }
|
cannam@154
|
711
|
cannam@154
|
712 from0 = cost0 + lambda;
|
cannam@154
|
713 from1 = cost1;
|
cannam@154
|
714 if (from0 < from1)
|
cannam@154
|
715 {
|
cannam@154
|
716 curr1 = from0;
|
cannam@154
|
717 path1[i]= 0;
|
cannam@154
|
718 } else {
|
cannam@154
|
719 curr1 = from1;
|
cannam@154
|
720 path1[i]= 1;
|
cannam@154
|
721 }
|
cannam@154
|
722 cost0 = curr0 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_select+0]);
|
cannam@154
|
723 cost1 = curr1 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_select+1]);
|
cannam@154
|
724 }
|
cannam@154
|
725 tf_res[len-1] = cost0 < cost1 ? 0 : 1;
|
cannam@154
|
726 /* Viterbi backward pass to check the decisions */
|
cannam@154
|
727 for (i=len-2;i>=0;i--)
|
cannam@154
|
728 {
|
cannam@154
|
729 if (tf_res[i+1] == 1)
|
cannam@154
|
730 tf_res[i] = path1[i+1];
|
cannam@154
|
731 else
|
cannam@154
|
732 tf_res[i] = path0[i+1];
|
cannam@154
|
733 }
|
cannam@154
|
734 /*printf("%d %f\n", *tf_sum, tf_estimate);*/
|
cannam@154
|
735 RESTORE_STACK;
|
cannam@154
|
736 #ifdef FUZZING
|
cannam@154
|
737 tf_select = rand()&0x1;
|
cannam@154
|
738 tf_res[0] = rand()&0x1;
|
cannam@154
|
739 for (i=1;i<len;i++)
|
cannam@154
|
740 tf_res[i] = tf_res[i-1] ^ ((rand()&0xF) == 0);
|
cannam@154
|
741 #endif
|
cannam@154
|
742 return tf_select;
|
cannam@154
|
743 }
|
cannam@154
|
744
|
cannam@154
|
745 static void tf_encode(int start, int end, int isTransient, int *tf_res, int LM, int tf_select, ec_enc *enc)
|
cannam@154
|
746 {
|
cannam@154
|
747 int curr, i;
|
cannam@154
|
748 int tf_select_rsv;
|
cannam@154
|
749 int tf_changed;
|
cannam@154
|
750 int logp;
|
cannam@154
|
751 opus_uint32 budget;
|
cannam@154
|
752 opus_uint32 tell;
|
cannam@154
|
753 budget = enc->storage*8;
|
cannam@154
|
754 tell = ec_tell(enc);
|
cannam@154
|
755 logp = isTransient ? 2 : 4;
|
cannam@154
|
756 /* Reserve space to code the tf_select decision. */
|
cannam@154
|
757 tf_select_rsv = LM>0 && tell+logp+1 <= budget;
|
cannam@154
|
758 budget -= tf_select_rsv;
|
cannam@154
|
759 curr = tf_changed = 0;
|
cannam@154
|
760 for (i=start;i<end;i++)
|
cannam@154
|
761 {
|
cannam@154
|
762 if (tell+logp<=budget)
|
cannam@154
|
763 {
|
cannam@154
|
764 ec_enc_bit_logp(enc, tf_res[i] ^ curr, logp);
|
cannam@154
|
765 tell = ec_tell(enc);
|
cannam@154
|
766 curr = tf_res[i];
|
cannam@154
|
767 tf_changed |= curr;
|
cannam@154
|
768 }
|
cannam@154
|
769 else
|
cannam@154
|
770 tf_res[i] = curr;
|
cannam@154
|
771 logp = isTransient ? 4 : 5;
|
cannam@154
|
772 }
|
cannam@154
|
773 /* Only code tf_select if it would actually make a difference. */
|
cannam@154
|
774 if (tf_select_rsv &&
|
cannam@154
|
775 tf_select_table[LM][4*isTransient+0+tf_changed]!=
|
cannam@154
|
776 tf_select_table[LM][4*isTransient+2+tf_changed])
|
cannam@154
|
777 ec_enc_bit_logp(enc, tf_select, 1);
|
cannam@154
|
778 else
|
cannam@154
|
779 tf_select = 0;
|
cannam@154
|
780 for (i=start;i<end;i++)
|
cannam@154
|
781 tf_res[i] = tf_select_table[LM][4*isTransient+2*tf_select+tf_res[i]];
|
cannam@154
|
782 /*for(i=0;i<end;i++)printf("%d ", isTransient ? tf_res[i] : LM+tf_res[i]);printf("\n");*/
|
cannam@154
|
783 }
|
cannam@154
|
784
|
cannam@154
|
785
|
cannam@154
|
786 static int alloc_trim_analysis(const CELTMode *m, const celt_norm *X,
|
cannam@154
|
787 const opus_val16 *bandLogE, int end, int LM, int C, int N0,
|
cannam@154
|
788 AnalysisInfo *analysis, opus_val16 *stereo_saving, opus_val16 tf_estimate,
|
cannam@154
|
789 int intensity, opus_val16 surround_trim, opus_int32 equiv_rate, int arch)
|
cannam@154
|
790 {
|
cannam@154
|
791 int i;
|
cannam@154
|
792 opus_val32 diff=0;
|
cannam@154
|
793 int c;
|
cannam@154
|
794 int trim_index;
|
cannam@154
|
795 opus_val16 trim = QCONST16(5.f, 8);
|
cannam@154
|
796 opus_val16 logXC, logXC2;
|
cannam@154
|
797 /* At low bitrate, reducing the trim seems to help. At higher bitrates, it's less
|
cannam@154
|
798 clear what's best, so we're keeping it as it was before, at least for now. */
|
cannam@154
|
799 if (equiv_rate < 64000) {
|
cannam@154
|
800 trim = QCONST16(4.f, 8);
|
cannam@154
|
801 } else if (equiv_rate < 80000) {
|
cannam@154
|
802 opus_int32 frac = (equiv_rate-64000) >> 10;
|
cannam@154
|
803 trim = QCONST16(4.f, 8) + QCONST16(1.f/16.f, 8)*frac;
|
cannam@154
|
804 }
|
cannam@154
|
805 if (C==2)
|
cannam@154
|
806 {
|
cannam@154
|
807 opus_val16 sum = 0; /* Q10 */
|
cannam@154
|
808 opus_val16 minXC; /* Q10 */
|
cannam@154
|
809 /* Compute inter-channel correlation for low frequencies */
|
cannam@154
|
810 for (i=0;i<8;i++)
|
cannam@154
|
811 {
|
cannam@154
|
812 opus_val32 partial;
|
cannam@154
|
813 partial = celt_inner_prod(&X[m->eBands[i]<<LM], &X[N0+(m->eBands[i]<<LM)],
|
cannam@154
|
814 (m->eBands[i+1]-m->eBands[i])<<LM, arch);
|
cannam@154
|
815 sum = ADD16(sum, EXTRACT16(SHR32(partial, 18)));
|
cannam@154
|
816 }
|
cannam@154
|
817 sum = MULT16_16_Q15(QCONST16(1.f/8, 15), sum);
|
cannam@154
|
818 sum = MIN16(QCONST16(1.f, 10), ABS16(sum));
|
cannam@154
|
819 minXC = sum;
|
cannam@154
|
820 for (i=8;i<intensity;i++)
|
cannam@154
|
821 {
|
cannam@154
|
822 opus_val32 partial;
|
cannam@154
|
823 partial = celt_inner_prod(&X[m->eBands[i]<<LM], &X[N0+(m->eBands[i]<<LM)],
|
cannam@154
|
824 (m->eBands[i+1]-m->eBands[i])<<LM, arch);
|
cannam@154
|
825 minXC = MIN16(minXC, ABS16(EXTRACT16(SHR32(partial, 18))));
|
cannam@154
|
826 }
|
cannam@154
|
827 minXC = MIN16(QCONST16(1.f, 10), ABS16(minXC));
|
cannam@154
|
828 /*printf ("%f\n", sum);*/
|
cannam@154
|
829 /* mid-side savings estimations based on the LF average*/
|
cannam@154
|
830 logXC = celt_log2(QCONST32(1.001f, 20)-MULT16_16(sum, sum));
|
cannam@154
|
831 /* mid-side savings estimations based on min correlation */
|
cannam@154
|
832 logXC2 = MAX16(HALF16(logXC), celt_log2(QCONST32(1.001f, 20)-MULT16_16(minXC, minXC)));
|
cannam@154
|
833 #ifdef FIXED_POINT
|
cannam@154
|
834 /* Compensate for Q20 vs Q14 input and convert output to Q8 */
|
cannam@154
|
835 logXC = PSHR32(logXC-QCONST16(6.f, DB_SHIFT),DB_SHIFT-8);
|
cannam@154
|
836 logXC2 = PSHR32(logXC2-QCONST16(6.f, DB_SHIFT),DB_SHIFT-8);
|
cannam@154
|
837 #endif
|
cannam@154
|
838
|
cannam@154
|
839 trim += MAX16(-QCONST16(4.f, 8), MULT16_16_Q15(QCONST16(.75f,15),logXC));
|
cannam@154
|
840 *stereo_saving = MIN16(*stereo_saving + QCONST16(0.25f, 8), -HALF16(logXC2));
|
cannam@154
|
841 }
|
cannam@154
|
842
|
cannam@154
|
843 /* Estimate spectral tilt */
|
cannam@154
|
844 c=0; do {
|
cannam@154
|
845 for (i=0;i<end-1;i++)
|
cannam@154
|
846 {
|
cannam@154
|
847 diff += bandLogE[i+c*m->nbEBands]*(opus_int32)(2+2*i-end);
|
cannam@154
|
848 }
|
cannam@154
|
849 } while (++c<C);
|
cannam@154
|
850 diff /= C*(end-1);
|
cannam@154
|
851 /*printf("%f\n", diff);*/
|
cannam@154
|
852 trim -= MAX32(-QCONST16(2.f, 8), MIN32(QCONST16(2.f, 8), SHR32(diff+QCONST16(1.f, DB_SHIFT),DB_SHIFT-8)/6 ));
|
cannam@154
|
853 trim -= SHR16(surround_trim, DB_SHIFT-8);
|
cannam@154
|
854 trim -= 2*SHR16(tf_estimate, 14-8);
|
cannam@154
|
855 #ifndef DISABLE_FLOAT_API
|
cannam@154
|
856 if (analysis->valid)
|
cannam@154
|
857 {
|
cannam@154
|
858 trim -= MAX16(-QCONST16(2.f, 8), MIN16(QCONST16(2.f, 8),
|
cannam@154
|
859 (opus_val16)(QCONST16(2.f, 8)*(analysis->tonality_slope+.05f))));
|
cannam@154
|
860 }
|
cannam@154
|
861 #else
|
cannam@154
|
862 (void)analysis;
|
cannam@154
|
863 #endif
|
cannam@154
|
864
|
cannam@154
|
865 #ifdef FIXED_POINT
|
cannam@154
|
866 trim_index = PSHR32(trim, 8);
|
cannam@154
|
867 #else
|
cannam@154
|
868 trim_index = (int)floor(.5f+trim);
|
cannam@154
|
869 #endif
|
cannam@154
|
870 trim_index = IMAX(0, IMIN(10, trim_index));
|
cannam@154
|
871 /*printf("%d\n", trim_index);*/
|
cannam@154
|
872 #ifdef FUZZING
|
cannam@154
|
873 trim_index = rand()%11;
|
cannam@154
|
874 #endif
|
cannam@154
|
875 return trim_index;
|
cannam@154
|
876 }
|
cannam@154
|
877
|
cannam@154
|
878 static int stereo_analysis(const CELTMode *m, const celt_norm *X,
|
cannam@154
|
879 int LM, int N0)
|
cannam@154
|
880 {
|
cannam@154
|
881 int i;
|
cannam@154
|
882 int thetas;
|
cannam@154
|
883 opus_val32 sumLR = EPSILON, sumMS = EPSILON;
|
cannam@154
|
884
|
cannam@154
|
885 /* Use the L1 norm to model the entropy of the L/R signal vs the M/S signal */
|
cannam@154
|
886 for (i=0;i<13;i++)
|
cannam@154
|
887 {
|
cannam@154
|
888 int j;
|
cannam@154
|
889 for (j=m->eBands[i]<<LM;j<m->eBands[i+1]<<LM;j++)
|
cannam@154
|
890 {
|
cannam@154
|
891 opus_val32 L, R, M, S;
|
cannam@154
|
892 /* We cast to 32-bit first because of the -32768 case */
|
cannam@154
|
893 L = EXTEND32(X[j]);
|
cannam@154
|
894 R = EXTEND32(X[N0+j]);
|
cannam@154
|
895 M = ADD32(L, R);
|
cannam@154
|
896 S = SUB32(L, R);
|
cannam@154
|
897 sumLR = ADD32(sumLR, ADD32(ABS32(L), ABS32(R)));
|
cannam@154
|
898 sumMS = ADD32(sumMS, ADD32(ABS32(M), ABS32(S)));
|
cannam@154
|
899 }
|
cannam@154
|
900 }
|
cannam@154
|
901 sumMS = MULT16_32_Q15(QCONST16(0.707107f, 15), sumMS);
|
cannam@154
|
902 thetas = 13;
|
cannam@154
|
903 /* We don't need thetas for lower bands with LM<=1 */
|
cannam@154
|
904 if (LM<=1)
|
cannam@154
|
905 thetas -= 8;
|
cannam@154
|
906 return MULT16_32_Q15((m->eBands[13]<<(LM+1))+thetas, sumMS)
|
cannam@154
|
907 > MULT16_32_Q15(m->eBands[13]<<(LM+1), sumLR);
|
cannam@154
|
908 }
|
cannam@154
|
909
|
cannam@154
|
910 #define MSWAP(a,b) do {opus_val16 tmp = a;a=b;b=tmp;} while(0)
|
cannam@154
|
911 static opus_val16 median_of_5(const opus_val16 *x)
|
cannam@154
|
912 {
|
cannam@154
|
913 opus_val16 t0, t1, t2, t3, t4;
|
cannam@154
|
914 t2 = x[2];
|
cannam@154
|
915 if (x[0] > x[1])
|
cannam@154
|
916 {
|
cannam@154
|
917 t0 = x[1];
|
cannam@154
|
918 t1 = x[0];
|
cannam@154
|
919 } else {
|
cannam@154
|
920 t0 = x[0];
|
cannam@154
|
921 t1 = x[1];
|
cannam@154
|
922 }
|
cannam@154
|
923 if (x[3] > x[4])
|
cannam@154
|
924 {
|
cannam@154
|
925 t3 = x[4];
|
cannam@154
|
926 t4 = x[3];
|
cannam@154
|
927 } else {
|
cannam@154
|
928 t3 = x[3];
|
cannam@154
|
929 t4 = x[4];
|
cannam@154
|
930 }
|
cannam@154
|
931 if (t0 > t3)
|
cannam@154
|
932 {
|
cannam@154
|
933 MSWAP(t0, t3);
|
cannam@154
|
934 MSWAP(t1, t4);
|
cannam@154
|
935 }
|
cannam@154
|
936 if (t2 > t1)
|
cannam@154
|
937 {
|
cannam@154
|
938 if (t1 < t3)
|
cannam@154
|
939 return MIN16(t2, t3);
|
cannam@154
|
940 else
|
cannam@154
|
941 return MIN16(t4, t1);
|
cannam@154
|
942 } else {
|
cannam@154
|
943 if (t2 < t3)
|
cannam@154
|
944 return MIN16(t1, t3);
|
cannam@154
|
945 else
|
cannam@154
|
946 return MIN16(t2, t4);
|
cannam@154
|
947 }
|
cannam@154
|
948 }
|
cannam@154
|
949
|
cannam@154
|
950 static opus_val16 median_of_3(const opus_val16 *x)
|
cannam@154
|
951 {
|
cannam@154
|
952 opus_val16 t0, t1, t2;
|
cannam@154
|
953 if (x[0] > x[1])
|
cannam@154
|
954 {
|
cannam@154
|
955 t0 = x[1];
|
cannam@154
|
956 t1 = x[0];
|
cannam@154
|
957 } else {
|
cannam@154
|
958 t0 = x[0];
|
cannam@154
|
959 t1 = x[1];
|
cannam@154
|
960 }
|
cannam@154
|
961 t2 = x[2];
|
cannam@154
|
962 if (t1 < t2)
|
cannam@154
|
963 return t1;
|
cannam@154
|
964 else if (t0 < t2)
|
cannam@154
|
965 return t2;
|
cannam@154
|
966 else
|
cannam@154
|
967 return t0;
|
cannam@154
|
968 }
|
cannam@154
|
969
|
cannam@154
|
970 static opus_val16 dynalloc_analysis(const opus_val16 *bandLogE, const opus_val16 *bandLogE2,
|
cannam@154
|
971 int nbEBands, int start, int end, int C, int *offsets, int lsb_depth, const opus_int16 *logN,
|
cannam@154
|
972 int isTransient, int vbr, int constrained_vbr, const opus_int16 *eBands, int LM,
|
cannam@154
|
973 int effectiveBytes, opus_int32 *tot_boost_, int lfe, opus_val16 *surround_dynalloc,
|
cannam@154
|
974 AnalysisInfo *analysis, int *importance, int *spread_weight)
|
cannam@154
|
975 {
|
cannam@154
|
976 int i, c;
|
cannam@154
|
977 opus_int32 tot_boost=0;
|
cannam@154
|
978 opus_val16 maxDepth;
|
cannam@154
|
979 VARDECL(opus_val16, follower);
|
cannam@154
|
980 VARDECL(opus_val16, noise_floor);
|
cannam@154
|
981 SAVE_STACK;
|
cannam@154
|
982 ALLOC(follower, C*nbEBands, opus_val16);
|
cannam@154
|
983 ALLOC(noise_floor, C*nbEBands, opus_val16);
|
cannam@154
|
984 OPUS_CLEAR(offsets, nbEBands);
|
cannam@154
|
985 /* Dynamic allocation code */
|
cannam@154
|
986 maxDepth=-QCONST16(31.9f, DB_SHIFT);
|
cannam@154
|
987 for (i=0;i<end;i++)
|
cannam@154
|
988 {
|
cannam@154
|
989 /* Noise floor must take into account eMeans, the depth, the width of the bands
|
cannam@154
|
990 and the preemphasis filter (approx. square of bark band ID) */
|
cannam@154
|
991 noise_floor[i] = MULT16_16(QCONST16(0.0625f, DB_SHIFT),logN[i])
|
cannam@154
|
992 +QCONST16(.5f,DB_SHIFT)+SHL16(9-lsb_depth,DB_SHIFT)-SHL16(eMeans[i],6)
|
cannam@154
|
993 +MULT16_16(QCONST16(.0062,DB_SHIFT),(i+5)*(i+5));
|
cannam@154
|
994 }
|
cannam@154
|
995 c=0;do
|
cannam@154
|
996 {
|
cannam@154
|
997 for (i=0;i<end;i++)
|
cannam@154
|
998 maxDepth = MAX16(maxDepth, bandLogE[c*nbEBands+i]-noise_floor[i]);
|
cannam@154
|
999 } while (++c<C);
|
cannam@154
|
1000 {
|
cannam@154
|
1001 /* Compute a really simple masking model to avoid taking into account completely masked
|
cannam@154
|
1002 bands when computing the spreading decision. */
|
cannam@154
|
1003 VARDECL(opus_val16, mask);
|
cannam@154
|
1004 VARDECL(opus_val16, sig);
|
cannam@154
|
1005 ALLOC(mask, nbEBands, opus_val16);
|
cannam@154
|
1006 ALLOC(sig, nbEBands, opus_val16);
|
cannam@154
|
1007 for (i=0;i<end;i++)
|
cannam@154
|
1008 mask[i] = bandLogE[i]-noise_floor[i];
|
cannam@154
|
1009 if (C==2)
|
cannam@154
|
1010 {
|
cannam@154
|
1011 for (i=0;i<end;i++)
|
cannam@154
|
1012 mask[i] = MAX16(mask[i], bandLogE[nbEBands+i]-noise_floor[i]);
|
cannam@154
|
1013 }
|
cannam@154
|
1014 OPUS_COPY(sig, mask, end);
|
cannam@154
|
1015 for (i=1;i<end;i++)
|
cannam@154
|
1016 mask[i] = MAX16(mask[i], mask[i-1] - QCONST16(2.f, DB_SHIFT));
|
cannam@154
|
1017 for (i=end-2;i>=0;i--)
|
cannam@154
|
1018 mask[i] = MAX16(mask[i], mask[i+1] - QCONST16(3.f, DB_SHIFT));
|
cannam@154
|
1019 for (i=0;i<end;i++)
|
cannam@154
|
1020 {
|
cannam@154
|
1021 /* Compute SMR: Mask is never more than 72 dB below the peak and never below the noise floor.*/
|
cannam@154
|
1022 opus_val16 smr = sig[i]-MAX16(MAX16(0, maxDepth-QCONST16(12.f, DB_SHIFT)), mask[i]);
|
cannam@154
|
1023 /* Clamp SMR to make sure we're not shifting by something negative or too large. */
|
cannam@154
|
1024 #ifdef FIXED_POINT
|
cannam@154
|
1025 /* FIXME: Use PSHR16() instead */
|
cannam@154
|
1026 int shift = -PSHR32(MAX16(-QCONST16(5.f, DB_SHIFT), MIN16(0, smr)), DB_SHIFT);
|
cannam@154
|
1027 #else
|
cannam@154
|
1028 int shift = IMIN(5, IMAX(0, -(int)floor(.5f + smr)));
|
cannam@154
|
1029 #endif
|
cannam@154
|
1030 spread_weight[i] = 32 >> shift;
|
cannam@154
|
1031 }
|
cannam@154
|
1032 /*for (i=0;i<end;i++)
|
cannam@154
|
1033 printf("%d ", spread_weight[i]);
|
cannam@154
|
1034 printf("\n");*/
|
cannam@154
|
1035 }
|
cannam@154
|
1036 /* Make sure that dynamic allocation can't make us bust the budget */
|
cannam@154
|
1037 if (effectiveBytes > 50 && LM>=1 && !lfe)
|
cannam@154
|
1038 {
|
cannam@154
|
1039 int last=0;
|
cannam@154
|
1040 c=0;do
|
cannam@154
|
1041 {
|
cannam@154
|
1042 opus_val16 offset;
|
cannam@154
|
1043 opus_val16 tmp;
|
cannam@154
|
1044 opus_val16 *f;
|
cannam@154
|
1045 f = &follower[c*nbEBands];
|
cannam@154
|
1046 f[0] = bandLogE2[c*nbEBands];
|
cannam@154
|
1047 for (i=1;i<end;i++)
|
cannam@154
|
1048 {
|
cannam@154
|
1049 /* The last band to be at least 3 dB higher than the previous one
|
cannam@154
|
1050 is the last we'll consider. Otherwise, we run into problems on
|
cannam@154
|
1051 bandlimited signals. */
|
cannam@154
|
1052 if (bandLogE2[c*nbEBands+i] > bandLogE2[c*nbEBands+i-1]+QCONST16(.5f,DB_SHIFT))
|
cannam@154
|
1053 last=i;
|
cannam@154
|
1054 f[i] = MIN16(f[i-1]+QCONST16(1.5f,DB_SHIFT), bandLogE2[c*nbEBands+i]);
|
cannam@154
|
1055 }
|
cannam@154
|
1056 for (i=last-1;i>=0;i--)
|
cannam@154
|
1057 f[i] = MIN16(f[i], MIN16(f[i+1]+QCONST16(2.f,DB_SHIFT), bandLogE2[c*nbEBands+i]));
|
cannam@154
|
1058
|
cannam@154
|
1059 /* Combine with a median filter to avoid dynalloc triggering unnecessarily.
|
cannam@154
|
1060 The "offset" value controls how conservative we are -- a higher offset
|
cannam@154
|
1061 reduces the impact of the median filter and makes dynalloc use more bits. */
|
cannam@154
|
1062 offset = QCONST16(1.f, DB_SHIFT);
|
cannam@154
|
1063 for (i=2;i<end-2;i++)
|
cannam@154
|
1064 f[i] = MAX16(f[i], median_of_5(&bandLogE2[c*nbEBands+i-2])-offset);
|
cannam@154
|
1065 tmp = median_of_3(&bandLogE2[c*nbEBands])-offset;
|
cannam@154
|
1066 f[0] = MAX16(f[0], tmp);
|
cannam@154
|
1067 f[1] = MAX16(f[1], tmp);
|
cannam@154
|
1068 tmp = median_of_3(&bandLogE2[c*nbEBands+end-3])-offset;
|
cannam@154
|
1069 f[end-2] = MAX16(f[end-2], tmp);
|
cannam@154
|
1070 f[end-1] = MAX16(f[end-1], tmp);
|
cannam@154
|
1071
|
cannam@154
|
1072 for (i=0;i<end;i++)
|
cannam@154
|
1073 f[i] = MAX16(f[i], noise_floor[i]);
|
cannam@154
|
1074 } while (++c<C);
|
cannam@154
|
1075 if (C==2)
|
cannam@154
|
1076 {
|
cannam@154
|
1077 for (i=start;i<end;i++)
|
cannam@154
|
1078 {
|
cannam@154
|
1079 /* Consider 24 dB "cross-talk" */
|
cannam@154
|
1080 follower[nbEBands+i] = MAX16(follower[nbEBands+i], follower[ i]-QCONST16(4.f,DB_SHIFT));
|
cannam@154
|
1081 follower[ i] = MAX16(follower[ i], follower[nbEBands+i]-QCONST16(4.f,DB_SHIFT));
|
cannam@154
|
1082 follower[i] = HALF16(MAX16(0, bandLogE[i]-follower[i]) + MAX16(0, bandLogE[nbEBands+i]-follower[nbEBands+i]));
|
cannam@154
|
1083 }
|
cannam@154
|
1084 } else {
|
cannam@154
|
1085 for (i=start;i<end;i++)
|
cannam@154
|
1086 {
|
cannam@154
|
1087 follower[i] = MAX16(0, bandLogE[i]-follower[i]);
|
cannam@154
|
1088 }
|
cannam@154
|
1089 }
|
cannam@154
|
1090 for (i=start;i<end;i++)
|
cannam@154
|
1091 follower[i] = MAX16(follower[i], surround_dynalloc[i]);
|
cannam@154
|
1092 for (i=start;i<end;i++)
|
cannam@154
|
1093 {
|
cannam@154
|
1094 #ifdef FIXED_POINT
|
cannam@154
|
1095 importance[i] = PSHR32(13*celt_exp2(MIN16(follower[i], QCONST16(4.f, DB_SHIFT))), 16);
|
cannam@154
|
1096 #else
|
cannam@154
|
1097 importance[i] = (int)floor(.5f+13*celt_exp2(MIN16(follower[i], QCONST16(4.f, DB_SHIFT))));
|
cannam@154
|
1098 #endif
|
cannam@154
|
1099 }
|
cannam@154
|
1100 /* For non-transient CBR/CVBR frames, halve the dynalloc contribution */
|
cannam@154
|
1101 if ((!vbr || constrained_vbr)&&!isTransient)
|
cannam@154
|
1102 {
|
cannam@154
|
1103 for (i=start;i<end;i++)
|
cannam@154
|
1104 follower[i] = HALF16(follower[i]);
|
cannam@154
|
1105 }
|
cannam@154
|
1106 for (i=start;i<end;i++)
|
cannam@154
|
1107 {
|
cannam@154
|
1108 if (i<8)
|
cannam@154
|
1109 follower[i] *= 2;
|
cannam@154
|
1110 if (i>=12)
|
cannam@154
|
1111 follower[i] = HALF16(follower[i]);
|
cannam@154
|
1112 }
|
cannam@154
|
1113 #ifdef DISABLE_FLOAT_API
|
cannam@154
|
1114 (void)analysis;
|
cannam@154
|
1115 #else
|
cannam@154
|
1116 if (analysis->valid)
|
cannam@154
|
1117 {
|
cannam@154
|
1118 for (i=start;i<IMIN(LEAK_BANDS, end);i++)
|
cannam@154
|
1119 follower[i] = follower[i] + QCONST16(1.f/64.f, DB_SHIFT)*analysis->leak_boost[i];
|
cannam@154
|
1120 }
|
cannam@154
|
1121 #endif
|
cannam@154
|
1122 for (i=start;i<end;i++)
|
cannam@154
|
1123 {
|
cannam@154
|
1124 int width;
|
cannam@154
|
1125 int boost;
|
cannam@154
|
1126 int boost_bits;
|
cannam@154
|
1127
|
cannam@154
|
1128 follower[i] = MIN16(follower[i], QCONST16(4, DB_SHIFT));
|
cannam@154
|
1129
|
cannam@154
|
1130 width = C*(eBands[i+1]-eBands[i])<<LM;
|
cannam@154
|
1131 if (width<6)
|
cannam@154
|
1132 {
|
cannam@154
|
1133 boost = (int)SHR32(EXTEND32(follower[i]),DB_SHIFT);
|
cannam@154
|
1134 boost_bits = boost*width<<BITRES;
|
cannam@154
|
1135 } else if (width > 48) {
|
cannam@154
|
1136 boost = (int)SHR32(EXTEND32(follower[i])*8,DB_SHIFT);
|
cannam@154
|
1137 boost_bits = (boost*width<<BITRES)/8;
|
cannam@154
|
1138 } else {
|
cannam@154
|
1139 boost = (int)SHR32(EXTEND32(follower[i])*width/6,DB_SHIFT);
|
cannam@154
|
1140 boost_bits = boost*6<<BITRES;
|
cannam@154
|
1141 }
|
cannam@154
|
1142 /* For CBR and non-transient CVBR frames, limit dynalloc to 2/3 of the bits */
|
cannam@154
|
1143 if ((!vbr || (constrained_vbr&&!isTransient))
|
cannam@154
|
1144 && (tot_boost+boost_bits)>>BITRES>>3 > 2*effectiveBytes/3)
|
cannam@154
|
1145 {
|
cannam@154
|
1146 opus_int32 cap = ((2*effectiveBytes/3)<<BITRES<<3);
|
cannam@154
|
1147 offsets[i] = cap-tot_boost;
|
cannam@154
|
1148 tot_boost = cap;
|
cannam@154
|
1149 break;
|
cannam@154
|
1150 } else {
|
cannam@154
|
1151 offsets[i] = boost;
|
cannam@154
|
1152 tot_boost += boost_bits;
|
cannam@154
|
1153 }
|
cannam@154
|
1154 }
|
cannam@154
|
1155 } else {
|
cannam@154
|
1156 for (i=start;i<end;i++)
|
cannam@154
|
1157 importance[i] = 13;
|
cannam@154
|
1158 }
|
cannam@154
|
1159 *tot_boost_ = tot_boost;
|
cannam@154
|
1160 RESTORE_STACK;
|
cannam@154
|
1161 return maxDepth;
|
cannam@154
|
1162 }
|
cannam@154
|
1163
|
cannam@154
|
1164
|
cannam@154
|
1165 static int run_prefilter(CELTEncoder *st, celt_sig *in, celt_sig *prefilter_mem, int CC, int N,
|
cannam@154
|
1166 int prefilter_tapset, int *pitch, opus_val16 *gain, int *qgain, int enabled, int nbAvailableBytes, AnalysisInfo *analysis)
|
cannam@154
|
1167 {
|
cannam@154
|
1168 int c;
|
cannam@154
|
1169 VARDECL(celt_sig, _pre);
|
cannam@154
|
1170 celt_sig *pre[2];
|
cannam@154
|
1171 const CELTMode *mode;
|
cannam@154
|
1172 int pitch_index;
|
cannam@154
|
1173 opus_val16 gain1;
|
cannam@154
|
1174 opus_val16 pf_threshold;
|
cannam@154
|
1175 int pf_on;
|
cannam@154
|
1176 int qg;
|
cannam@154
|
1177 int overlap;
|
cannam@154
|
1178 SAVE_STACK;
|
cannam@154
|
1179
|
cannam@154
|
1180 mode = st->mode;
|
cannam@154
|
1181 overlap = mode->overlap;
|
cannam@154
|
1182 ALLOC(_pre, CC*(N+COMBFILTER_MAXPERIOD), celt_sig);
|
cannam@154
|
1183
|
cannam@154
|
1184 pre[0] = _pre;
|
cannam@154
|
1185 pre[1] = _pre + (N+COMBFILTER_MAXPERIOD);
|
cannam@154
|
1186
|
cannam@154
|
1187
|
cannam@154
|
1188 c=0; do {
|
cannam@154
|
1189 OPUS_COPY(pre[c], prefilter_mem+c*COMBFILTER_MAXPERIOD, COMBFILTER_MAXPERIOD);
|
cannam@154
|
1190 OPUS_COPY(pre[c]+COMBFILTER_MAXPERIOD, in+c*(N+overlap)+overlap, N);
|
cannam@154
|
1191 } while (++c<CC);
|
cannam@154
|
1192
|
cannam@154
|
1193 if (enabled)
|
cannam@154
|
1194 {
|
cannam@154
|
1195 VARDECL(opus_val16, pitch_buf);
|
cannam@154
|
1196 ALLOC(pitch_buf, (COMBFILTER_MAXPERIOD+N)>>1, opus_val16);
|
cannam@154
|
1197
|
cannam@154
|
1198 pitch_downsample(pre, pitch_buf, COMBFILTER_MAXPERIOD+N, CC, st->arch);
|
cannam@154
|
1199 /* Don't search for the fir last 1.5 octave of the range because
|
cannam@154
|
1200 there's too many false-positives due to short-term correlation */
|
cannam@154
|
1201 pitch_search(pitch_buf+(COMBFILTER_MAXPERIOD>>1), pitch_buf, N,
|
cannam@154
|
1202 COMBFILTER_MAXPERIOD-3*COMBFILTER_MINPERIOD, &pitch_index,
|
cannam@154
|
1203 st->arch);
|
cannam@154
|
1204 pitch_index = COMBFILTER_MAXPERIOD-pitch_index;
|
cannam@154
|
1205
|
cannam@154
|
1206 gain1 = remove_doubling(pitch_buf, COMBFILTER_MAXPERIOD, COMBFILTER_MINPERIOD,
|
cannam@154
|
1207 N, &pitch_index, st->prefilter_period, st->prefilter_gain, st->arch);
|
cannam@154
|
1208 if (pitch_index > COMBFILTER_MAXPERIOD-2)
|
cannam@154
|
1209 pitch_index = COMBFILTER_MAXPERIOD-2;
|
cannam@154
|
1210 gain1 = MULT16_16_Q15(QCONST16(.7f,15),gain1);
|
cannam@154
|
1211 /*printf("%d %d %f %f\n", pitch_change, pitch_index, gain1, st->analysis.tonality);*/
|
cannam@154
|
1212 if (st->loss_rate>2)
|
cannam@154
|
1213 gain1 = HALF32(gain1);
|
cannam@154
|
1214 if (st->loss_rate>4)
|
cannam@154
|
1215 gain1 = HALF32(gain1);
|
cannam@154
|
1216 if (st->loss_rate>8)
|
cannam@154
|
1217 gain1 = 0;
|
cannam@154
|
1218 } else {
|
cannam@154
|
1219 gain1 = 0;
|
cannam@154
|
1220 pitch_index = COMBFILTER_MINPERIOD;
|
cannam@154
|
1221 }
|
cannam@154
|
1222 #ifndef DISABLE_FLOAT_API
|
cannam@154
|
1223 if (analysis->valid)
|
cannam@154
|
1224 gain1 = (opus_val16)(gain1 * analysis->max_pitch_ratio);
|
cannam@154
|
1225 #else
|
cannam@154
|
1226 (void)analysis;
|
cannam@154
|
1227 #endif
|
cannam@154
|
1228 /* Gain threshold for enabling the prefilter/postfilter */
|
cannam@154
|
1229 pf_threshold = QCONST16(.2f,15);
|
cannam@154
|
1230
|
cannam@154
|
1231 /* Adjusting the threshold based on rate and continuity */
|
cannam@154
|
1232 if (abs(pitch_index-st->prefilter_period)*10>pitch_index)
|
cannam@154
|
1233 pf_threshold += QCONST16(.2f,15);
|
cannam@154
|
1234 if (nbAvailableBytes<25)
|
cannam@154
|
1235 pf_threshold += QCONST16(.1f,15);
|
cannam@154
|
1236 if (nbAvailableBytes<35)
|
cannam@154
|
1237 pf_threshold += QCONST16(.1f,15);
|
cannam@154
|
1238 if (st->prefilter_gain > QCONST16(.4f,15))
|
cannam@154
|
1239 pf_threshold -= QCONST16(.1f,15);
|
cannam@154
|
1240 if (st->prefilter_gain > QCONST16(.55f,15))
|
cannam@154
|
1241 pf_threshold -= QCONST16(.1f,15);
|
cannam@154
|
1242
|
cannam@154
|
1243 /* Hard threshold at 0.2 */
|
cannam@154
|
1244 pf_threshold = MAX16(pf_threshold, QCONST16(.2f,15));
|
cannam@154
|
1245 if (gain1<pf_threshold)
|
cannam@154
|
1246 {
|
cannam@154
|
1247 gain1 = 0;
|
cannam@154
|
1248 pf_on = 0;
|
cannam@154
|
1249 qg = 0;
|
cannam@154
|
1250 } else {
|
cannam@154
|
1251 /*This block is not gated by a total bits check only because
|
cannam@154
|
1252 of the nbAvailableBytes check above.*/
|
cannam@154
|
1253 if (ABS16(gain1-st->prefilter_gain)<QCONST16(.1f,15))
|
cannam@154
|
1254 gain1=st->prefilter_gain;
|
cannam@154
|
1255
|
cannam@154
|
1256 #ifdef FIXED_POINT
|
cannam@154
|
1257 qg = ((gain1+1536)>>10)/3-1;
|
cannam@154
|
1258 #else
|
cannam@154
|
1259 qg = (int)floor(.5f+gain1*32/3)-1;
|
cannam@154
|
1260 #endif
|
cannam@154
|
1261 qg = IMAX(0, IMIN(7, qg));
|
cannam@154
|
1262 gain1 = QCONST16(0.09375f,15)*(qg+1);
|
cannam@154
|
1263 pf_on = 1;
|
cannam@154
|
1264 }
|
cannam@154
|
1265 /*printf("%d %f\n", pitch_index, gain1);*/
|
cannam@154
|
1266
|
cannam@154
|
1267 c=0; do {
|
cannam@154
|
1268 int offset = mode->shortMdctSize-overlap;
|
cannam@154
|
1269 st->prefilter_period=IMAX(st->prefilter_period, COMBFILTER_MINPERIOD);
|
cannam@154
|
1270 OPUS_COPY(in+c*(N+overlap), st->in_mem+c*(overlap), overlap);
|
cannam@154
|
1271 if (offset)
|
cannam@154
|
1272 comb_filter(in+c*(N+overlap)+overlap, pre[c]+COMBFILTER_MAXPERIOD,
|
cannam@154
|
1273 st->prefilter_period, st->prefilter_period, offset, -st->prefilter_gain, -st->prefilter_gain,
|
cannam@154
|
1274 st->prefilter_tapset, st->prefilter_tapset, NULL, 0, st->arch);
|
cannam@154
|
1275
|
cannam@154
|
1276 comb_filter(in+c*(N+overlap)+overlap+offset, pre[c]+COMBFILTER_MAXPERIOD+offset,
|
cannam@154
|
1277 st->prefilter_period, pitch_index, N-offset, -st->prefilter_gain, -gain1,
|
cannam@154
|
1278 st->prefilter_tapset, prefilter_tapset, mode->window, overlap, st->arch);
|
cannam@154
|
1279 OPUS_COPY(st->in_mem+c*(overlap), in+c*(N+overlap)+N, overlap);
|
cannam@154
|
1280
|
cannam@154
|
1281 if (N>COMBFILTER_MAXPERIOD)
|
cannam@154
|
1282 {
|
cannam@154
|
1283 OPUS_COPY(prefilter_mem+c*COMBFILTER_MAXPERIOD, pre[c]+N, COMBFILTER_MAXPERIOD);
|
cannam@154
|
1284 } else {
|
cannam@154
|
1285 OPUS_MOVE(prefilter_mem+c*COMBFILTER_MAXPERIOD, prefilter_mem+c*COMBFILTER_MAXPERIOD+N, COMBFILTER_MAXPERIOD-N);
|
cannam@154
|
1286 OPUS_COPY(prefilter_mem+c*COMBFILTER_MAXPERIOD+COMBFILTER_MAXPERIOD-N, pre[c]+COMBFILTER_MAXPERIOD, N);
|
cannam@154
|
1287 }
|
cannam@154
|
1288 } while (++c<CC);
|
cannam@154
|
1289
|
cannam@154
|
1290 RESTORE_STACK;
|
cannam@154
|
1291 *gain = gain1;
|
cannam@154
|
1292 *pitch = pitch_index;
|
cannam@154
|
1293 *qgain = qg;
|
cannam@154
|
1294 return pf_on;
|
cannam@154
|
1295 }
|
cannam@154
|
1296
|
cannam@154
|
1297 static int compute_vbr(const CELTMode *mode, AnalysisInfo *analysis, opus_int32 base_target,
|
cannam@154
|
1298 int LM, opus_int32 bitrate, int lastCodedBands, int C, int intensity,
|
cannam@154
|
1299 int constrained_vbr, opus_val16 stereo_saving, int tot_boost,
|
cannam@154
|
1300 opus_val16 tf_estimate, int pitch_change, opus_val16 maxDepth,
|
cannam@154
|
1301 int lfe, int has_surround_mask, opus_val16 surround_masking,
|
cannam@154
|
1302 opus_val16 temporal_vbr)
|
cannam@154
|
1303 {
|
cannam@154
|
1304 /* The target rate in 8th bits per frame */
|
cannam@154
|
1305 opus_int32 target;
|
cannam@154
|
1306 int coded_bins;
|
cannam@154
|
1307 int coded_bands;
|
cannam@154
|
1308 opus_val16 tf_calibration;
|
cannam@154
|
1309 int nbEBands;
|
cannam@154
|
1310 const opus_int16 *eBands;
|
cannam@154
|
1311
|
cannam@154
|
1312 nbEBands = mode->nbEBands;
|
cannam@154
|
1313 eBands = mode->eBands;
|
cannam@154
|
1314
|
cannam@154
|
1315 coded_bands = lastCodedBands ? lastCodedBands : nbEBands;
|
cannam@154
|
1316 coded_bins = eBands[coded_bands]<<LM;
|
cannam@154
|
1317 if (C==2)
|
cannam@154
|
1318 coded_bins += eBands[IMIN(intensity, coded_bands)]<<LM;
|
cannam@154
|
1319
|
cannam@154
|
1320 target = base_target;
|
cannam@154
|
1321
|
cannam@154
|
1322 /*printf("%f %f %f %f %d %d ", st->analysis.activity, st->analysis.tonality, tf_estimate, st->stereo_saving, tot_boost, coded_bands);*/
|
cannam@154
|
1323 #ifndef DISABLE_FLOAT_API
|
cannam@154
|
1324 if (analysis->valid && analysis->activity<.4)
|
cannam@154
|
1325 target -= (opus_int32)((coded_bins<<BITRES)*(.4f-analysis->activity));
|
cannam@154
|
1326 #endif
|
cannam@154
|
1327 /* Stereo savings */
|
cannam@154
|
1328 if (C==2)
|
cannam@154
|
1329 {
|
cannam@154
|
1330 int coded_stereo_bands;
|
cannam@154
|
1331 int coded_stereo_dof;
|
cannam@154
|
1332 opus_val16 max_frac;
|
cannam@154
|
1333 coded_stereo_bands = IMIN(intensity, coded_bands);
|
cannam@154
|
1334 coded_stereo_dof = (eBands[coded_stereo_bands]<<LM)-coded_stereo_bands;
|
cannam@154
|
1335 /* Maximum fraction of the bits we can save if the signal is mono. */
|
cannam@154
|
1336 max_frac = DIV32_16(MULT16_16(QCONST16(0.8f, 15), coded_stereo_dof), coded_bins);
|
cannam@154
|
1337 stereo_saving = MIN16(stereo_saving, QCONST16(1.f, 8));
|
cannam@154
|
1338 /*printf("%d %d %d ", coded_stereo_dof, coded_bins, tot_boost);*/
|
cannam@154
|
1339 target -= (opus_int32)MIN32(MULT16_32_Q15(max_frac,target),
|
cannam@154
|
1340 SHR32(MULT16_16(stereo_saving-QCONST16(0.1f,8),(coded_stereo_dof<<BITRES)),8));
|
cannam@154
|
1341 }
|
cannam@154
|
1342 /* Boost the rate according to dynalloc (minus the dynalloc average for calibration). */
|
cannam@154
|
1343 target += tot_boost-(19<<LM);
|
cannam@154
|
1344 /* Apply transient boost, compensating for average boost. */
|
cannam@154
|
1345 tf_calibration = QCONST16(0.044f,14);
|
cannam@154
|
1346 target += (opus_int32)SHL32(MULT16_32_Q15(tf_estimate-tf_calibration, target),1);
|
cannam@154
|
1347
|
cannam@154
|
1348 #ifndef DISABLE_FLOAT_API
|
cannam@154
|
1349 /* Apply tonality boost */
|
cannam@154
|
1350 if (analysis->valid && !lfe)
|
cannam@154
|
1351 {
|
cannam@154
|
1352 opus_int32 tonal_target;
|
cannam@154
|
1353 float tonal;
|
cannam@154
|
1354
|
cannam@154
|
1355 /* Tonality boost (compensating for the average). */
|
cannam@154
|
1356 tonal = MAX16(0.f,analysis->tonality-.15f)-0.12f;
|
cannam@154
|
1357 tonal_target = target + (opus_int32)((coded_bins<<BITRES)*1.2f*tonal);
|
cannam@154
|
1358 if (pitch_change)
|
cannam@154
|
1359 tonal_target += (opus_int32)((coded_bins<<BITRES)*.8f);
|
cannam@154
|
1360 /*printf("%f %f ", analysis->tonality, tonal);*/
|
cannam@154
|
1361 target = tonal_target;
|
cannam@154
|
1362 }
|
cannam@154
|
1363 #else
|
cannam@154
|
1364 (void)analysis;
|
cannam@154
|
1365 (void)pitch_change;
|
cannam@154
|
1366 #endif
|
cannam@154
|
1367
|
cannam@154
|
1368 if (has_surround_mask&&!lfe)
|
cannam@154
|
1369 {
|
cannam@154
|
1370 opus_int32 surround_target = target + (opus_int32)SHR32(MULT16_16(surround_masking,coded_bins<<BITRES), DB_SHIFT);
|
cannam@154
|
1371 /*printf("%f %d %d %d %d %d %d ", surround_masking, coded_bins, st->end, st->intensity, surround_target, target, st->bitrate);*/
|
cannam@154
|
1372 target = IMAX(target/4, surround_target);
|
cannam@154
|
1373 }
|
cannam@154
|
1374
|
cannam@154
|
1375 {
|
cannam@154
|
1376 opus_int32 floor_depth;
|
cannam@154
|
1377 int bins;
|
cannam@154
|
1378 bins = eBands[nbEBands-2]<<LM;
|
cannam@154
|
1379 /*floor_depth = SHR32(MULT16_16((C*bins<<BITRES),celt_log2(SHL32(MAX16(1,sample_max),13))), DB_SHIFT);*/
|
cannam@154
|
1380 floor_depth = (opus_int32)SHR32(MULT16_16((C*bins<<BITRES),maxDepth), DB_SHIFT);
|
cannam@154
|
1381 floor_depth = IMAX(floor_depth, target>>2);
|
cannam@154
|
1382 target = IMIN(target, floor_depth);
|
cannam@154
|
1383 /*printf("%f %d\n", maxDepth, floor_depth);*/
|
cannam@154
|
1384 }
|
cannam@154
|
1385
|
cannam@154
|
1386 /* Make VBR less aggressive for constrained VBR because we can't keep a higher bitrate
|
cannam@154
|
1387 for long. Needs tuning. */
|
cannam@154
|
1388 if ((!has_surround_mask||lfe) && constrained_vbr)
|
cannam@154
|
1389 {
|
cannam@154
|
1390 target = base_target + (opus_int32)MULT16_32_Q15(QCONST16(0.67f, 15), target-base_target);
|
cannam@154
|
1391 }
|
cannam@154
|
1392
|
cannam@154
|
1393 if (!has_surround_mask && tf_estimate < QCONST16(.2f, 14))
|
cannam@154
|
1394 {
|
cannam@154
|
1395 opus_val16 amount;
|
cannam@154
|
1396 opus_val16 tvbr_factor;
|
cannam@154
|
1397 amount = MULT16_16_Q15(QCONST16(.0000031f, 30), IMAX(0, IMIN(32000, 96000-bitrate)));
|
cannam@154
|
1398 tvbr_factor = SHR32(MULT16_16(temporal_vbr, amount), DB_SHIFT);
|
cannam@154
|
1399 target += (opus_int32)MULT16_32_Q15(tvbr_factor, target);
|
cannam@154
|
1400 }
|
cannam@154
|
1401
|
cannam@154
|
1402 /* Don't allow more than doubling the rate */
|
cannam@154
|
1403 target = IMIN(2*base_target, target);
|
cannam@154
|
1404
|
cannam@154
|
1405 return target;
|
cannam@154
|
1406 }
|
cannam@154
|
1407
|
cannam@154
|
1408 int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes, ec_enc *enc)
|
cannam@154
|
1409 {
|
cannam@154
|
1410 int i, c, N;
|
cannam@154
|
1411 opus_int32 bits;
|
cannam@154
|
1412 ec_enc _enc;
|
cannam@154
|
1413 VARDECL(celt_sig, in);
|
cannam@154
|
1414 VARDECL(celt_sig, freq);
|
cannam@154
|
1415 VARDECL(celt_norm, X);
|
cannam@154
|
1416 VARDECL(celt_ener, bandE);
|
cannam@154
|
1417 VARDECL(opus_val16, bandLogE);
|
cannam@154
|
1418 VARDECL(opus_val16, bandLogE2);
|
cannam@154
|
1419 VARDECL(int, fine_quant);
|
cannam@154
|
1420 VARDECL(opus_val16, error);
|
cannam@154
|
1421 VARDECL(int, pulses);
|
cannam@154
|
1422 VARDECL(int, cap);
|
cannam@154
|
1423 VARDECL(int, offsets);
|
cannam@154
|
1424 VARDECL(int, importance);
|
cannam@154
|
1425 VARDECL(int, spread_weight);
|
cannam@154
|
1426 VARDECL(int, fine_priority);
|
cannam@154
|
1427 VARDECL(int, tf_res);
|
cannam@154
|
1428 VARDECL(unsigned char, collapse_masks);
|
cannam@154
|
1429 celt_sig *prefilter_mem;
|
cannam@154
|
1430 opus_val16 *oldBandE, *oldLogE, *oldLogE2, *energyError;
|
cannam@154
|
1431 int shortBlocks=0;
|
cannam@154
|
1432 int isTransient=0;
|
cannam@154
|
1433 const int CC = st->channels;
|
cannam@154
|
1434 const int C = st->stream_channels;
|
cannam@154
|
1435 int LM, M;
|
cannam@154
|
1436 int tf_select;
|
cannam@154
|
1437 int nbFilledBytes, nbAvailableBytes;
|
cannam@154
|
1438 int start;
|
cannam@154
|
1439 int end;
|
cannam@154
|
1440 int effEnd;
|
cannam@154
|
1441 int codedBands;
|
cannam@154
|
1442 int alloc_trim;
|
cannam@154
|
1443 int pitch_index=COMBFILTER_MINPERIOD;
|
cannam@154
|
1444 opus_val16 gain1 = 0;
|
cannam@154
|
1445 int dual_stereo=0;
|
cannam@154
|
1446 int effectiveBytes;
|
cannam@154
|
1447 int dynalloc_logp;
|
cannam@154
|
1448 opus_int32 vbr_rate;
|
cannam@154
|
1449 opus_int32 total_bits;
|
cannam@154
|
1450 opus_int32 total_boost;
|
cannam@154
|
1451 opus_int32 balance;
|
cannam@154
|
1452 opus_int32 tell;
|
cannam@154
|
1453 opus_int32 tell0_frac;
|
cannam@154
|
1454 int prefilter_tapset=0;
|
cannam@154
|
1455 int pf_on;
|
cannam@154
|
1456 int anti_collapse_rsv;
|
cannam@154
|
1457 int anti_collapse_on=0;
|
cannam@154
|
1458 int silence=0;
|
cannam@154
|
1459 int tf_chan = 0;
|
cannam@154
|
1460 opus_val16 tf_estimate;
|
cannam@154
|
1461 int pitch_change=0;
|
cannam@154
|
1462 opus_int32 tot_boost;
|
cannam@154
|
1463 opus_val32 sample_max;
|
cannam@154
|
1464 opus_val16 maxDepth;
|
cannam@154
|
1465 const OpusCustomMode *mode;
|
cannam@154
|
1466 int nbEBands;
|
cannam@154
|
1467 int overlap;
|
cannam@154
|
1468 const opus_int16 *eBands;
|
cannam@154
|
1469 int secondMdct;
|
cannam@154
|
1470 int signalBandwidth;
|
cannam@154
|
1471 int transient_got_disabled=0;
|
cannam@154
|
1472 opus_val16 surround_masking=0;
|
cannam@154
|
1473 opus_val16 temporal_vbr=0;
|
cannam@154
|
1474 opus_val16 surround_trim = 0;
|
cannam@154
|
1475 opus_int32 equiv_rate;
|
cannam@154
|
1476 int hybrid;
|
cannam@154
|
1477 int weak_transient = 0;
|
cannam@154
|
1478 int enable_tf_analysis;
|
cannam@154
|
1479 VARDECL(opus_val16, surround_dynalloc);
|
cannam@154
|
1480 ALLOC_STACK;
|
cannam@154
|
1481
|
cannam@154
|
1482 mode = st->mode;
|
cannam@154
|
1483 nbEBands = mode->nbEBands;
|
cannam@154
|
1484 overlap = mode->overlap;
|
cannam@154
|
1485 eBands = mode->eBands;
|
cannam@154
|
1486 start = st->start;
|
cannam@154
|
1487 end = st->end;
|
cannam@154
|
1488 hybrid = start != 0;
|
cannam@154
|
1489 tf_estimate = 0;
|
cannam@154
|
1490 if (nbCompressedBytes<2 || pcm==NULL)
|
cannam@154
|
1491 {
|
cannam@154
|
1492 RESTORE_STACK;
|
cannam@154
|
1493 return OPUS_BAD_ARG;
|
cannam@154
|
1494 }
|
cannam@154
|
1495
|
cannam@154
|
1496 frame_size *= st->upsample;
|
cannam@154
|
1497 for (LM=0;LM<=mode->maxLM;LM++)
|
cannam@154
|
1498 if (mode->shortMdctSize<<LM==frame_size)
|
cannam@154
|
1499 break;
|
cannam@154
|
1500 if (LM>mode->maxLM)
|
cannam@154
|
1501 {
|
cannam@154
|
1502 RESTORE_STACK;
|
cannam@154
|
1503 return OPUS_BAD_ARG;
|
cannam@154
|
1504 }
|
cannam@154
|
1505 M=1<<LM;
|
cannam@154
|
1506 N = M*mode->shortMdctSize;
|
cannam@154
|
1507
|
cannam@154
|
1508 prefilter_mem = st->in_mem+CC*(overlap);
|
cannam@154
|
1509 oldBandE = (opus_val16*)(st->in_mem+CC*(overlap+COMBFILTER_MAXPERIOD));
|
cannam@154
|
1510 oldLogE = oldBandE + CC*nbEBands;
|
cannam@154
|
1511 oldLogE2 = oldLogE + CC*nbEBands;
|
cannam@154
|
1512 energyError = oldLogE2 + CC*nbEBands;
|
cannam@154
|
1513
|
cannam@154
|
1514 if (enc==NULL)
|
cannam@154
|
1515 {
|
cannam@154
|
1516 tell0_frac=tell=1;
|
cannam@154
|
1517 nbFilledBytes=0;
|
cannam@154
|
1518 } else {
|
cannam@154
|
1519 tell0_frac=ec_tell_frac(enc);
|
cannam@154
|
1520 tell=ec_tell(enc);
|
cannam@154
|
1521 nbFilledBytes=(tell+4)>>3;
|
cannam@154
|
1522 }
|
cannam@154
|
1523
|
cannam@154
|
1524 #ifdef CUSTOM_MODES
|
cannam@154
|
1525 if (st->signalling && enc==NULL)
|
cannam@154
|
1526 {
|
cannam@154
|
1527 int tmp = (mode->effEBands-end)>>1;
|
cannam@154
|
1528 end = st->end = IMAX(1, mode->effEBands-tmp);
|
cannam@154
|
1529 compressed[0] = tmp<<5;
|
cannam@154
|
1530 compressed[0] |= LM<<3;
|
cannam@154
|
1531 compressed[0] |= (C==2)<<2;
|
cannam@154
|
1532 /* Convert "standard mode" to Opus header */
|
cannam@154
|
1533 if (mode->Fs==48000 && mode->shortMdctSize==120)
|
cannam@154
|
1534 {
|
cannam@154
|
1535 int c0 = toOpus(compressed[0]);
|
cannam@154
|
1536 if (c0<0)
|
cannam@154
|
1537 {
|
cannam@154
|
1538 RESTORE_STACK;
|
cannam@154
|
1539 return OPUS_BAD_ARG;
|
cannam@154
|
1540 }
|
cannam@154
|
1541 compressed[0] = c0;
|
cannam@154
|
1542 }
|
cannam@154
|
1543 compressed++;
|
cannam@154
|
1544 nbCompressedBytes--;
|
cannam@154
|
1545 }
|
cannam@154
|
1546 #else
|
cannam@154
|
1547 celt_assert(st->signalling==0);
|
cannam@154
|
1548 #endif
|
cannam@154
|
1549
|
cannam@154
|
1550 /* Can't produce more than 1275 output bytes */
|
cannam@154
|
1551 nbCompressedBytes = IMIN(nbCompressedBytes,1275);
|
cannam@154
|
1552 nbAvailableBytes = nbCompressedBytes - nbFilledBytes;
|
cannam@154
|
1553
|
cannam@154
|
1554 if (st->vbr && st->bitrate!=OPUS_BITRATE_MAX)
|
cannam@154
|
1555 {
|
cannam@154
|
1556 opus_int32 den=mode->Fs>>BITRES;
|
cannam@154
|
1557 vbr_rate=(st->bitrate*frame_size+(den>>1))/den;
|
cannam@154
|
1558 #ifdef CUSTOM_MODES
|
cannam@154
|
1559 if (st->signalling)
|
cannam@154
|
1560 vbr_rate -= 8<<BITRES;
|
cannam@154
|
1561 #endif
|
cannam@154
|
1562 effectiveBytes = vbr_rate>>(3+BITRES);
|
cannam@154
|
1563 } else {
|
cannam@154
|
1564 opus_int32 tmp;
|
cannam@154
|
1565 vbr_rate = 0;
|
cannam@154
|
1566 tmp = st->bitrate*frame_size;
|
cannam@154
|
1567 if (tell>1)
|
cannam@154
|
1568 tmp += tell;
|
cannam@154
|
1569 if (st->bitrate!=OPUS_BITRATE_MAX)
|
cannam@154
|
1570 nbCompressedBytes = IMAX(2, IMIN(nbCompressedBytes,
|
cannam@154
|
1571 (tmp+4*mode->Fs)/(8*mode->Fs)-!!st->signalling));
|
cannam@154
|
1572 effectiveBytes = nbCompressedBytes - nbFilledBytes;
|
cannam@154
|
1573 }
|
cannam@154
|
1574 equiv_rate = ((opus_int32)nbCompressedBytes*8*50 >> (3-LM)) - (40*C+20)*((400>>LM) - 50);
|
cannam@154
|
1575 if (st->bitrate != OPUS_BITRATE_MAX)
|
cannam@154
|
1576 equiv_rate = IMIN(equiv_rate, st->bitrate - (40*C+20)*((400>>LM) - 50));
|
cannam@154
|
1577
|
cannam@154
|
1578 if (enc==NULL)
|
cannam@154
|
1579 {
|
cannam@154
|
1580 ec_enc_init(&_enc, compressed, nbCompressedBytes);
|
cannam@154
|
1581 enc = &_enc;
|
cannam@154
|
1582 }
|
cannam@154
|
1583
|
cannam@154
|
1584 if (vbr_rate>0)
|
cannam@154
|
1585 {
|
cannam@154
|
1586 /* Computes the max bit-rate allowed in VBR mode to avoid violating the
|
cannam@154
|
1587 target rate and buffering.
|
cannam@154
|
1588 We must do this up front so that bust-prevention logic triggers
|
cannam@154
|
1589 correctly if we don't have enough bits. */
|
cannam@154
|
1590 if (st->constrained_vbr)
|
cannam@154
|
1591 {
|
cannam@154
|
1592 opus_int32 vbr_bound;
|
cannam@154
|
1593 opus_int32 max_allowed;
|
cannam@154
|
1594 /* We could use any multiple of vbr_rate as bound (depending on the
|
cannam@154
|
1595 delay).
|
cannam@154
|
1596 This is clamped to ensure we use at least two bytes if the encoder
|
cannam@154
|
1597 was entirely empty, but to allow 0 in hybrid mode. */
|
cannam@154
|
1598 vbr_bound = vbr_rate;
|
cannam@154
|
1599 max_allowed = IMIN(IMAX(tell==1?2:0,
|
cannam@154
|
1600 (vbr_rate+vbr_bound-st->vbr_reservoir)>>(BITRES+3)),
|
cannam@154
|
1601 nbAvailableBytes);
|
cannam@154
|
1602 if(max_allowed < nbAvailableBytes)
|
cannam@154
|
1603 {
|
cannam@154
|
1604 nbCompressedBytes = nbFilledBytes+max_allowed;
|
cannam@154
|
1605 nbAvailableBytes = max_allowed;
|
cannam@154
|
1606 ec_enc_shrink(enc, nbCompressedBytes);
|
cannam@154
|
1607 }
|
cannam@154
|
1608 }
|
cannam@154
|
1609 }
|
cannam@154
|
1610 total_bits = nbCompressedBytes*8;
|
cannam@154
|
1611
|
cannam@154
|
1612 effEnd = end;
|
cannam@154
|
1613 if (effEnd > mode->effEBands)
|
cannam@154
|
1614 effEnd = mode->effEBands;
|
cannam@154
|
1615
|
cannam@154
|
1616 ALLOC(in, CC*(N+overlap), celt_sig);
|
cannam@154
|
1617
|
cannam@154
|
1618 sample_max=MAX32(st->overlap_max, celt_maxabs16(pcm, C*(N-overlap)/st->upsample));
|
cannam@154
|
1619 st->overlap_max=celt_maxabs16(pcm+C*(N-overlap)/st->upsample, C*overlap/st->upsample);
|
cannam@154
|
1620 sample_max=MAX32(sample_max, st->overlap_max);
|
cannam@154
|
1621 #ifdef FIXED_POINT
|
cannam@154
|
1622 silence = (sample_max==0);
|
cannam@154
|
1623 #else
|
cannam@154
|
1624 silence = (sample_max <= (opus_val16)1/(1<<st->lsb_depth));
|
cannam@154
|
1625 #endif
|
cannam@154
|
1626 #ifdef FUZZING
|
cannam@154
|
1627 if ((rand()&0x3F)==0)
|
cannam@154
|
1628 silence = 1;
|
cannam@154
|
1629 #endif
|
cannam@154
|
1630 if (tell==1)
|
cannam@154
|
1631 ec_enc_bit_logp(enc, silence, 15);
|
cannam@154
|
1632 else
|
cannam@154
|
1633 silence=0;
|
cannam@154
|
1634 if (silence)
|
cannam@154
|
1635 {
|
cannam@154
|
1636 /*In VBR mode there is no need to send more than the minimum. */
|
cannam@154
|
1637 if (vbr_rate>0)
|
cannam@154
|
1638 {
|
cannam@154
|
1639 effectiveBytes=nbCompressedBytes=IMIN(nbCompressedBytes, nbFilledBytes+2);
|
cannam@154
|
1640 total_bits=nbCompressedBytes*8;
|
cannam@154
|
1641 nbAvailableBytes=2;
|
cannam@154
|
1642 ec_enc_shrink(enc, nbCompressedBytes);
|
cannam@154
|
1643 }
|
cannam@154
|
1644 /* Pretend we've filled all the remaining bits with zeros
|
cannam@154
|
1645 (that's what the initialiser did anyway) */
|
cannam@154
|
1646 tell = nbCompressedBytes*8;
|
cannam@154
|
1647 enc->nbits_total+=tell-ec_tell(enc);
|
cannam@154
|
1648 }
|
cannam@154
|
1649 c=0; do {
|
cannam@154
|
1650 int need_clip=0;
|
cannam@154
|
1651 #ifndef FIXED_POINT
|
cannam@154
|
1652 need_clip = st->clip && sample_max>65536.f;
|
cannam@154
|
1653 #endif
|
cannam@154
|
1654 celt_preemphasis(pcm+c, in+c*(N+overlap)+overlap, N, CC, st->upsample,
|
cannam@154
|
1655 mode->preemph, st->preemph_memE+c, need_clip);
|
cannam@154
|
1656 } while (++c<CC);
|
cannam@154
|
1657
|
cannam@154
|
1658
|
cannam@154
|
1659
|
cannam@154
|
1660 /* Find pitch period and gain */
|
cannam@154
|
1661 {
|
cannam@154
|
1662 int enabled;
|
cannam@154
|
1663 int qg;
|
cannam@154
|
1664 enabled = ((st->lfe&&nbAvailableBytes>3) || nbAvailableBytes>12*C) && !hybrid && !silence && !st->disable_pf
|
cannam@154
|
1665 && st->complexity >= 5;
|
cannam@154
|
1666
|
cannam@154
|
1667 prefilter_tapset = st->tapset_decision;
|
cannam@154
|
1668 pf_on = run_prefilter(st, in, prefilter_mem, CC, N, prefilter_tapset, &pitch_index, &gain1, &qg, enabled, nbAvailableBytes, &st->analysis);
|
cannam@154
|
1669 if ((gain1 > QCONST16(.4f,15) || st->prefilter_gain > QCONST16(.4f,15)) && (!st->analysis.valid || st->analysis.tonality > .3)
|
cannam@154
|
1670 && (pitch_index > 1.26*st->prefilter_period || pitch_index < .79*st->prefilter_period))
|
cannam@154
|
1671 pitch_change = 1;
|
cannam@154
|
1672 if (pf_on==0)
|
cannam@154
|
1673 {
|
cannam@154
|
1674 if(!hybrid && tell+16<=total_bits)
|
cannam@154
|
1675 ec_enc_bit_logp(enc, 0, 1);
|
cannam@154
|
1676 } else {
|
cannam@154
|
1677 /*This block is not gated by a total bits check only because
|
cannam@154
|
1678 of the nbAvailableBytes check above.*/
|
cannam@154
|
1679 int octave;
|
cannam@154
|
1680 ec_enc_bit_logp(enc, 1, 1);
|
cannam@154
|
1681 pitch_index += 1;
|
cannam@154
|
1682 octave = EC_ILOG(pitch_index)-5;
|
cannam@154
|
1683 ec_enc_uint(enc, octave, 6);
|
cannam@154
|
1684 ec_enc_bits(enc, pitch_index-(16<<octave), 4+octave);
|
cannam@154
|
1685 pitch_index -= 1;
|
cannam@154
|
1686 ec_enc_bits(enc, qg, 3);
|
cannam@154
|
1687 ec_enc_icdf(enc, prefilter_tapset, tapset_icdf, 2);
|
cannam@154
|
1688 }
|
cannam@154
|
1689 }
|
cannam@154
|
1690
|
cannam@154
|
1691 isTransient = 0;
|
cannam@154
|
1692 shortBlocks = 0;
|
cannam@154
|
1693 if (st->complexity >= 1 && !st->lfe)
|
cannam@154
|
1694 {
|
cannam@154
|
1695 /* Reduces the likelihood of energy instability on fricatives at low bitrate
|
cannam@154
|
1696 in hybrid mode. It seems like we still want to have real transients on vowels
|
cannam@154
|
1697 though (small SILK quantization offset value). */
|
cannam@154
|
1698 int allow_weak_transients = hybrid && effectiveBytes<15 && st->silk_info.signalType != 2;
|
cannam@154
|
1699 isTransient = transient_analysis(in, N+overlap, CC,
|
cannam@154
|
1700 &tf_estimate, &tf_chan, allow_weak_transients, &weak_transient);
|
cannam@154
|
1701 }
|
cannam@154
|
1702 if (LM>0 && ec_tell(enc)+3<=total_bits)
|
cannam@154
|
1703 {
|
cannam@154
|
1704 if (isTransient)
|
cannam@154
|
1705 shortBlocks = M;
|
cannam@154
|
1706 } else {
|
cannam@154
|
1707 isTransient = 0;
|
cannam@154
|
1708 transient_got_disabled=1;
|
cannam@154
|
1709 }
|
cannam@154
|
1710
|
cannam@154
|
1711 ALLOC(freq, CC*N, celt_sig); /**< Interleaved signal MDCTs */
|
cannam@154
|
1712 ALLOC(bandE,nbEBands*CC, celt_ener);
|
cannam@154
|
1713 ALLOC(bandLogE,nbEBands*CC, opus_val16);
|
cannam@154
|
1714
|
cannam@154
|
1715 secondMdct = shortBlocks && st->complexity>=8;
|
cannam@154
|
1716 ALLOC(bandLogE2, C*nbEBands, opus_val16);
|
cannam@154
|
1717 if (secondMdct)
|
cannam@154
|
1718 {
|
cannam@154
|
1719 compute_mdcts(mode, 0, in, freq, C, CC, LM, st->upsample, st->arch);
|
cannam@154
|
1720 compute_band_energies(mode, freq, bandE, effEnd, C, LM, st->arch);
|
cannam@154
|
1721 amp2Log2(mode, effEnd, end, bandE, bandLogE2, C);
|
cannam@154
|
1722 for (i=0;i<C*nbEBands;i++)
|
cannam@154
|
1723 bandLogE2[i] += HALF16(SHL16(LM, DB_SHIFT));
|
cannam@154
|
1724 }
|
cannam@154
|
1725
|
cannam@154
|
1726 compute_mdcts(mode, shortBlocks, in, freq, C, CC, LM, st->upsample, st->arch);
|
cannam@154
|
1727 /* This should catch any NaN in the CELT input. Since we're not supposed to see any (they're filtered
|
cannam@154
|
1728 at the Opus layer), just abort. */
|
cannam@154
|
1729 celt_assert(!celt_isnan(freq[0]) && (C==1 || !celt_isnan(freq[N])));
|
cannam@154
|
1730 if (CC==2&&C==1)
|
cannam@154
|
1731 tf_chan = 0;
|
cannam@154
|
1732 compute_band_energies(mode, freq, bandE, effEnd, C, LM, st->arch);
|
cannam@154
|
1733
|
cannam@154
|
1734 if (st->lfe)
|
cannam@154
|
1735 {
|
cannam@154
|
1736 for (i=2;i<end;i++)
|
cannam@154
|
1737 {
|
cannam@154
|
1738 bandE[i] = IMIN(bandE[i], MULT16_32_Q15(QCONST16(1e-4f,15),bandE[0]));
|
cannam@154
|
1739 bandE[i] = MAX32(bandE[i], EPSILON);
|
cannam@154
|
1740 }
|
cannam@154
|
1741 }
|
cannam@154
|
1742 amp2Log2(mode, effEnd, end, bandE, bandLogE, C);
|
cannam@154
|
1743
|
cannam@154
|
1744 ALLOC(surround_dynalloc, C*nbEBands, opus_val16);
|
cannam@154
|
1745 OPUS_CLEAR(surround_dynalloc, end);
|
cannam@154
|
1746 /* This computes how much masking takes place between surround channels */
|
cannam@154
|
1747 if (!hybrid&&st->energy_mask&&!st->lfe)
|
cannam@154
|
1748 {
|
cannam@154
|
1749 int mask_end;
|
cannam@154
|
1750 int midband;
|
cannam@154
|
1751 int count_dynalloc;
|
cannam@154
|
1752 opus_val32 mask_avg=0;
|
cannam@154
|
1753 opus_val32 diff=0;
|
cannam@154
|
1754 int count=0;
|
cannam@154
|
1755 mask_end = IMAX(2,st->lastCodedBands);
|
cannam@154
|
1756 for (c=0;c<C;c++)
|
cannam@154
|
1757 {
|
cannam@154
|
1758 for(i=0;i<mask_end;i++)
|
cannam@154
|
1759 {
|
cannam@154
|
1760 opus_val16 mask;
|
cannam@154
|
1761 mask = MAX16(MIN16(st->energy_mask[nbEBands*c+i],
|
cannam@154
|
1762 QCONST16(.25f, DB_SHIFT)), -QCONST16(2.0f, DB_SHIFT));
|
cannam@154
|
1763 if (mask > 0)
|
cannam@154
|
1764 mask = HALF16(mask);
|
cannam@154
|
1765 mask_avg += MULT16_16(mask, eBands[i+1]-eBands[i]);
|
cannam@154
|
1766 count += eBands[i+1]-eBands[i];
|
cannam@154
|
1767 diff += MULT16_16(mask, 1+2*i-mask_end);
|
cannam@154
|
1768 }
|
cannam@154
|
1769 }
|
cannam@154
|
1770 celt_assert(count>0);
|
cannam@154
|
1771 mask_avg = DIV32_16(mask_avg,count);
|
cannam@154
|
1772 mask_avg += QCONST16(.2f, DB_SHIFT);
|
cannam@154
|
1773 diff = diff*6/(C*(mask_end-1)*(mask_end+1)*mask_end);
|
cannam@154
|
1774 /* Again, being conservative */
|
cannam@154
|
1775 diff = HALF32(diff);
|
cannam@154
|
1776 diff = MAX32(MIN32(diff, QCONST32(.031f, DB_SHIFT)), -QCONST32(.031f, DB_SHIFT));
|
cannam@154
|
1777 /* Find the band that's in the middle of the coded spectrum */
|
cannam@154
|
1778 for (midband=0;eBands[midband+1] < eBands[mask_end]/2;midband++);
|
cannam@154
|
1779 count_dynalloc=0;
|
cannam@154
|
1780 for(i=0;i<mask_end;i++)
|
cannam@154
|
1781 {
|
cannam@154
|
1782 opus_val32 lin;
|
cannam@154
|
1783 opus_val16 unmask;
|
cannam@154
|
1784 lin = mask_avg + diff*(i-midband);
|
cannam@154
|
1785 if (C==2)
|
cannam@154
|
1786 unmask = MAX16(st->energy_mask[i], st->energy_mask[nbEBands+i]);
|
cannam@154
|
1787 else
|
cannam@154
|
1788 unmask = st->energy_mask[i];
|
cannam@154
|
1789 unmask = MIN16(unmask, QCONST16(.0f, DB_SHIFT));
|
cannam@154
|
1790 unmask -= lin;
|
cannam@154
|
1791 if (unmask > QCONST16(.25f, DB_SHIFT))
|
cannam@154
|
1792 {
|
cannam@154
|
1793 surround_dynalloc[i] = unmask - QCONST16(.25f, DB_SHIFT);
|
cannam@154
|
1794 count_dynalloc++;
|
cannam@154
|
1795 }
|
cannam@154
|
1796 }
|
cannam@154
|
1797 if (count_dynalloc>=3)
|
cannam@154
|
1798 {
|
cannam@154
|
1799 /* If we need dynalloc in many bands, it's probably because our
|
cannam@154
|
1800 initial masking rate was too low. */
|
cannam@154
|
1801 mask_avg += QCONST16(.25f, DB_SHIFT);
|
cannam@154
|
1802 if (mask_avg>0)
|
cannam@154
|
1803 {
|
cannam@154
|
1804 /* Something went really wrong in the original calculations,
|
cannam@154
|
1805 disabling masking. */
|
cannam@154
|
1806 mask_avg = 0;
|
cannam@154
|
1807 diff = 0;
|
cannam@154
|
1808 OPUS_CLEAR(surround_dynalloc, mask_end);
|
cannam@154
|
1809 } else {
|
cannam@154
|
1810 for(i=0;i<mask_end;i++)
|
cannam@154
|
1811 surround_dynalloc[i] = MAX16(0, surround_dynalloc[i]-QCONST16(.25f, DB_SHIFT));
|
cannam@154
|
1812 }
|
cannam@154
|
1813 }
|
cannam@154
|
1814 mask_avg += QCONST16(.2f, DB_SHIFT);
|
cannam@154
|
1815 /* Convert to 1/64th units used for the trim */
|
cannam@154
|
1816 surround_trim = 64*diff;
|
cannam@154
|
1817 /*printf("%d %d ", mask_avg, surround_trim);*/
|
cannam@154
|
1818 surround_masking = mask_avg;
|
cannam@154
|
1819 }
|
cannam@154
|
1820 /* Temporal VBR (but not for LFE) */
|
cannam@154
|
1821 if (!st->lfe)
|
cannam@154
|
1822 {
|
cannam@154
|
1823 opus_val16 follow=-QCONST16(10.0f,DB_SHIFT);
|
cannam@154
|
1824 opus_val32 frame_avg=0;
|
cannam@154
|
1825 opus_val16 offset = shortBlocks?HALF16(SHL16(LM, DB_SHIFT)):0;
|
cannam@154
|
1826 for(i=start;i<end;i++)
|
cannam@154
|
1827 {
|
cannam@154
|
1828 follow = MAX16(follow-QCONST16(1.f, DB_SHIFT), bandLogE[i]-offset);
|
cannam@154
|
1829 if (C==2)
|
cannam@154
|
1830 follow = MAX16(follow, bandLogE[i+nbEBands]-offset);
|
cannam@154
|
1831 frame_avg += follow;
|
cannam@154
|
1832 }
|
cannam@154
|
1833 frame_avg /= (end-start);
|
cannam@154
|
1834 temporal_vbr = SUB16(frame_avg,st->spec_avg);
|
cannam@154
|
1835 temporal_vbr = MIN16(QCONST16(3.f, DB_SHIFT), MAX16(-QCONST16(1.5f, DB_SHIFT), temporal_vbr));
|
cannam@154
|
1836 st->spec_avg += MULT16_16_Q15(QCONST16(.02f, 15), temporal_vbr);
|
cannam@154
|
1837 }
|
cannam@154
|
1838 /*for (i=0;i<21;i++)
|
cannam@154
|
1839 printf("%f ", bandLogE[i]);
|
cannam@154
|
1840 printf("\n");*/
|
cannam@154
|
1841
|
cannam@154
|
1842 if (!secondMdct)
|
cannam@154
|
1843 {
|
cannam@154
|
1844 OPUS_COPY(bandLogE2, bandLogE, C*nbEBands);
|
cannam@154
|
1845 }
|
cannam@154
|
1846
|
cannam@154
|
1847 /* Last chance to catch any transient we might have missed in the
|
cannam@154
|
1848 time-domain analysis */
|
cannam@154
|
1849 if (LM>0 && ec_tell(enc)+3<=total_bits && !isTransient && st->complexity>=5 && !st->lfe && !hybrid)
|
cannam@154
|
1850 {
|
cannam@154
|
1851 if (patch_transient_decision(bandLogE, oldBandE, nbEBands, start, end, C))
|
cannam@154
|
1852 {
|
cannam@154
|
1853 isTransient = 1;
|
cannam@154
|
1854 shortBlocks = M;
|
cannam@154
|
1855 compute_mdcts(mode, shortBlocks, in, freq, C, CC, LM, st->upsample, st->arch);
|
cannam@154
|
1856 compute_band_energies(mode, freq, bandE, effEnd, C, LM, st->arch);
|
cannam@154
|
1857 amp2Log2(mode, effEnd, end, bandE, bandLogE, C);
|
cannam@154
|
1858 /* Compensate for the scaling of short vs long mdcts */
|
cannam@154
|
1859 for (i=0;i<C*nbEBands;i++)
|
cannam@154
|
1860 bandLogE2[i] += HALF16(SHL16(LM, DB_SHIFT));
|
cannam@154
|
1861 tf_estimate = QCONST16(.2f,14);
|
cannam@154
|
1862 }
|
cannam@154
|
1863 }
|
cannam@154
|
1864
|
cannam@154
|
1865 if (LM>0 && ec_tell(enc)+3<=total_bits)
|
cannam@154
|
1866 ec_enc_bit_logp(enc, isTransient, 3);
|
cannam@154
|
1867
|
cannam@154
|
1868 ALLOC(X, C*N, celt_norm); /**< Interleaved normalised MDCTs */
|
cannam@154
|
1869
|
cannam@154
|
1870 /* Band normalisation */
|
cannam@154
|
1871 normalise_bands(mode, freq, X, bandE, effEnd, C, M);
|
cannam@154
|
1872
|
cannam@154
|
1873 enable_tf_analysis = effectiveBytes>=15*C && !hybrid && st->complexity>=2 && !st->lfe;
|
cannam@154
|
1874
|
cannam@154
|
1875 ALLOC(offsets, nbEBands, int);
|
cannam@154
|
1876 ALLOC(importance, nbEBands, int);
|
cannam@154
|
1877 ALLOC(spread_weight, nbEBands, int);
|
cannam@154
|
1878
|
cannam@154
|
1879 maxDepth = dynalloc_analysis(bandLogE, bandLogE2, nbEBands, start, end, C, offsets,
|
cannam@154
|
1880 st->lsb_depth, mode->logN, isTransient, st->vbr, st->constrained_vbr,
|
cannam@154
|
1881 eBands, LM, effectiveBytes, &tot_boost, st->lfe, surround_dynalloc, &st->analysis, importance, spread_weight);
|
cannam@154
|
1882
|
cannam@154
|
1883 ALLOC(tf_res, nbEBands, int);
|
cannam@154
|
1884 /* Disable variable tf resolution for hybrid and at very low bitrate */
|
cannam@154
|
1885 if (enable_tf_analysis)
|
cannam@154
|
1886 {
|
cannam@154
|
1887 int lambda;
|
cannam@154
|
1888 lambda = IMAX(80, 20480/effectiveBytes + 2);
|
cannam@154
|
1889 tf_select = tf_analysis(mode, effEnd, isTransient, tf_res, lambda, X, N, LM, tf_estimate, tf_chan, importance);
|
cannam@154
|
1890 for (i=effEnd;i<end;i++)
|
cannam@154
|
1891 tf_res[i] = tf_res[effEnd-1];
|
cannam@154
|
1892 } else if (hybrid && weak_transient)
|
cannam@154
|
1893 {
|
cannam@154
|
1894 /* For weak transients, we rely on the fact that improving time resolution using
|
cannam@154
|
1895 TF on a long window is imperfect and will not result in an energy collapse at
|
cannam@154
|
1896 low bitrate. */
|
cannam@154
|
1897 for (i=0;i<end;i++)
|
cannam@154
|
1898 tf_res[i] = 1;
|
cannam@154
|
1899 tf_select=0;
|
cannam@154
|
1900 } else if (hybrid && effectiveBytes<15 && st->silk_info.signalType != 2)
|
cannam@154
|
1901 {
|
cannam@154
|
1902 /* For low bitrate hybrid, we force temporal resolution to 5 ms rather than 2.5 ms. */
|
cannam@154
|
1903 for (i=0;i<end;i++)
|
cannam@154
|
1904 tf_res[i] = 0;
|
cannam@154
|
1905 tf_select=isTransient;
|
cannam@154
|
1906 } else {
|
cannam@154
|
1907 for (i=0;i<end;i++)
|
cannam@154
|
1908 tf_res[i] = isTransient;
|
cannam@154
|
1909 tf_select=0;
|
cannam@154
|
1910 }
|
cannam@154
|
1911
|
cannam@154
|
1912 ALLOC(error, C*nbEBands, opus_val16);
|
cannam@154
|
1913 c=0;
|
cannam@154
|
1914 do {
|
cannam@154
|
1915 for (i=start;i<end;i++)
|
cannam@154
|
1916 {
|
cannam@154
|
1917 /* When the energy is stable, slightly bias energy quantization towards
|
cannam@154
|
1918 the previous error to make the gain more stable (a constant offset is
|
cannam@154
|
1919 better than fluctuations). */
|
cannam@154
|
1920 if (ABS32(SUB32(bandLogE[i+c*nbEBands], oldBandE[i+c*nbEBands])) < QCONST16(2.f, DB_SHIFT))
|
cannam@154
|
1921 {
|
cannam@154
|
1922 bandLogE[i+c*nbEBands] -= MULT16_16_Q15(energyError[i+c*nbEBands], QCONST16(0.25f, 15));
|
cannam@154
|
1923 }
|
cannam@154
|
1924 }
|
cannam@154
|
1925 } while (++c < C);
|
cannam@154
|
1926 quant_coarse_energy(mode, start, end, effEnd, bandLogE,
|
cannam@154
|
1927 oldBandE, total_bits, error, enc,
|
cannam@154
|
1928 C, LM, nbAvailableBytes, st->force_intra,
|
cannam@154
|
1929 &st->delayedIntra, st->complexity >= 4, st->loss_rate, st->lfe);
|
cannam@154
|
1930
|
cannam@154
|
1931 tf_encode(start, end, isTransient, tf_res, LM, tf_select, enc);
|
cannam@154
|
1932
|
cannam@154
|
1933 if (ec_tell(enc)+4<=total_bits)
|
cannam@154
|
1934 {
|
cannam@154
|
1935 if (st->lfe)
|
cannam@154
|
1936 {
|
cannam@154
|
1937 st->tapset_decision = 0;
|
cannam@154
|
1938 st->spread_decision = SPREAD_NORMAL;
|
cannam@154
|
1939 } else if (hybrid)
|
cannam@154
|
1940 {
|
cannam@154
|
1941 if (st->complexity == 0)
|
cannam@154
|
1942 st->spread_decision = SPREAD_NONE;
|
cannam@154
|
1943 else if (isTransient)
|
cannam@154
|
1944 st->spread_decision = SPREAD_NORMAL;
|
cannam@154
|
1945 else
|
cannam@154
|
1946 st->spread_decision = SPREAD_AGGRESSIVE;
|
cannam@154
|
1947 } else if (shortBlocks || st->complexity < 3 || nbAvailableBytes < 10*C)
|
cannam@154
|
1948 {
|
cannam@154
|
1949 if (st->complexity == 0)
|
cannam@154
|
1950 st->spread_decision = SPREAD_NONE;
|
cannam@154
|
1951 else
|
cannam@154
|
1952 st->spread_decision = SPREAD_NORMAL;
|
cannam@154
|
1953 } else {
|
cannam@154
|
1954 /* Disable new spreading+tapset estimator until we can show it works
|
cannam@154
|
1955 better than the old one. So far it seems like spreading_decision()
|
cannam@154
|
1956 works best. */
|
cannam@154
|
1957 #if 0
|
cannam@154
|
1958 if (st->analysis.valid)
|
cannam@154
|
1959 {
|
cannam@154
|
1960 static const opus_val16 spread_thresholds[3] = {-QCONST16(.6f, 15), -QCONST16(.2f, 15), -QCONST16(.07f, 15)};
|
cannam@154
|
1961 static const opus_val16 spread_histeresis[3] = {QCONST16(.15f, 15), QCONST16(.07f, 15), QCONST16(.02f, 15)};
|
cannam@154
|
1962 static const opus_val16 tapset_thresholds[2] = {QCONST16(.0f, 15), QCONST16(.15f, 15)};
|
cannam@154
|
1963 static const opus_val16 tapset_histeresis[2] = {QCONST16(.1f, 15), QCONST16(.05f, 15)};
|
cannam@154
|
1964 st->spread_decision = hysteresis_decision(-st->analysis.tonality, spread_thresholds, spread_histeresis, 3, st->spread_decision);
|
cannam@154
|
1965 st->tapset_decision = hysteresis_decision(st->analysis.tonality_slope, tapset_thresholds, tapset_histeresis, 2, st->tapset_decision);
|
cannam@154
|
1966 } else
|
cannam@154
|
1967 #endif
|
cannam@154
|
1968 {
|
cannam@154
|
1969 st->spread_decision = spreading_decision(mode, X,
|
cannam@154
|
1970 &st->tonal_average, st->spread_decision, &st->hf_average,
|
cannam@154
|
1971 &st->tapset_decision, pf_on&&!shortBlocks, effEnd, C, M, spread_weight);
|
cannam@154
|
1972 }
|
cannam@154
|
1973 /*printf("%d %d\n", st->tapset_decision, st->spread_decision);*/
|
cannam@154
|
1974 /*printf("%f %d %f %d\n\n", st->analysis.tonality, st->spread_decision, st->analysis.tonality_slope, st->tapset_decision);*/
|
cannam@154
|
1975 }
|
cannam@154
|
1976 ec_enc_icdf(enc, st->spread_decision, spread_icdf, 5);
|
cannam@154
|
1977 }
|
cannam@154
|
1978
|
cannam@154
|
1979 /* For LFE, everything interesting is in the first band */
|
cannam@154
|
1980 if (st->lfe)
|
cannam@154
|
1981 offsets[0] = IMIN(8, effectiveBytes/3);
|
cannam@154
|
1982 ALLOC(cap, nbEBands, int);
|
cannam@154
|
1983 init_caps(mode,cap,LM,C);
|
cannam@154
|
1984
|
cannam@154
|
1985 dynalloc_logp = 6;
|
cannam@154
|
1986 total_bits<<=BITRES;
|
cannam@154
|
1987 total_boost = 0;
|
cannam@154
|
1988 tell = ec_tell_frac(enc);
|
cannam@154
|
1989 for (i=start;i<end;i++)
|
cannam@154
|
1990 {
|
cannam@154
|
1991 int width, quanta;
|
cannam@154
|
1992 int dynalloc_loop_logp;
|
cannam@154
|
1993 int boost;
|
cannam@154
|
1994 int j;
|
cannam@154
|
1995 width = C*(eBands[i+1]-eBands[i])<<LM;
|
cannam@154
|
1996 /* quanta is 6 bits, but no more than 1 bit/sample
|
cannam@154
|
1997 and no less than 1/8 bit/sample */
|
cannam@154
|
1998 quanta = IMIN(width<<BITRES, IMAX(6<<BITRES, width));
|
cannam@154
|
1999 dynalloc_loop_logp = dynalloc_logp;
|
cannam@154
|
2000 boost = 0;
|
cannam@154
|
2001 for (j = 0; tell+(dynalloc_loop_logp<<BITRES) < total_bits-total_boost
|
cannam@154
|
2002 && boost < cap[i]; j++)
|
cannam@154
|
2003 {
|
cannam@154
|
2004 int flag;
|
cannam@154
|
2005 flag = j<offsets[i];
|
cannam@154
|
2006 ec_enc_bit_logp(enc, flag, dynalloc_loop_logp);
|
cannam@154
|
2007 tell = ec_tell_frac(enc);
|
cannam@154
|
2008 if (!flag)
|
cannam@154
|
2009 break;
|
cannam@154
|
2010 boost += quanta;
|
cannam@154
|
2011 total_boost += quanta;
|
cannam@154
|
2012 dynalloc_loop_logp = 1;
|
cannam@154
|
2013 }
|
cannam@154
|
2014 /* Making dynalloc more likely */
|
cannam@154
|
2015 if (j)
|
cannam@154
|
2016 dynalloc_logp = IMAX(2, dynalloc_logp-1);
|
cannam@154
|
2017 offsets[i] = boost;
|
cannam@154
|
2018 }
|
cannam@154
|
2019
|
cannam@154
|
2020 if (C==2)
|
cannam@154
|
2021 {
|
cannam@154
|
2022 static const opus_val16 intensity_thresholds[21]=
|
cannam@154
|
2023 /* 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 off*/
|
cannam@154
|
2024 { 1, 2, 3, 4, 5, 6, 7, 8,16,24,36,44,50,56,62,67,72,79,88,106,134};
|
cannam@154
|
2025 static const opus_val16 intensity_histeresis[21]=
|
cannam@154
|
2026 { 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 3, 3, 4, 5, 6, 8, 8};
|
cannam@154
|
2027
|
cannam@154
|
2028 /* Always use MS for 2.5 ms frames until we can do a better analysis */
|
cannam@154
|
2029 if (LM!=0)
|
cannam@154
|
2030 dual_stereo = stereo_analysis(mode, X, LM, N);
|
cannam@154
|
2031
|
cannam@154
|
2032 st->intensity = hysteresis_decision((opus_val16)(equiv_rate/1000),
|
cannam@154
|
2033 intensity_thresholds, intensity_histeresis, 21, st->intensity);
|
cannam@154
|
2034 st->intensity = IMIN(end,IMAX(start, st->intensity));
|
cannam@154
|
2035 }
|
cannam@154
|
2036
|
cannam@154
|
2037 alloc_trim = 5;
|
cannam@154
|
2038 if (tell+(6<<BITRES) <= total_bits - total_boost)
|
cannam@154
|
2039 {
|
cannam@154
|
2040 if (start > 0 || st->lfe)
|
cannam@154
|
2041 {
|
cannam@154
|
2042 st->stereo_saving = 0;
|
cannam@154
|
2043 alloc_trim = 5;
|
cannam@154
|
2044 } else {
|
cannam@154
|
2045 alloc_trim = alloc_trim_analysis(mode, X, bandLogE,
|
cannam@154
|
2046 end, LM, C, N, &st->analysis, &st->stereo_saving, tf_estimate,
|
cannam@154
|
2047 st->intensity, surround_trim, equiv_rate, st->arch);
|
cannam@154
|
2048 }
|
cannam@154
|
2049 ec_enc_icdf(enc, alloc_trim, trim_icdf, 7);
|
cannam@154
|
2050 tell = ec_tell_frac(enc);
|
cannam@154
|
2051 }
|
cannam@154
|
2052
|
cannam@154
|
2053 /* Variable bitrate */
|
cannam@154
|
2054 if (vbr_rate>0)
|
cannam@154
|
2055 {
|
cannam@154
|
2056 opus_val16 alpha;
|
cannam@154
|
2057 opus_int32 delta;
|
cannam@154
|
2058 /* The target rate in 8th bits per frame */
|
cannam@154
|
2059 opus_int32 target, base_target;
|
cannam@154
|
2060 opus_int32 min_allowed;
|
cannam@154
|
2061 int lm_diff = mode->maxLM - LM;
|
cannam@154
|
2062
|
cannam@154
|
2063 /* Don't attempt to use more than 510 kb/s, even for frames smaller than 20 ms.
|
cannam@154
|
2064 The CELT allocator will just not be able to use more than that anyway. */
|
cannam@154
|
2065 nbCompressedBytes = IMIN(nbCompressedBytes,1275>>(3-LM));
|
cannam@154
|
2066 if (!hybrid)
|
cannam@154
|
2067 {
|
cannam@154
|
2068 base_target = vbr_rate - ((40*C+20)<<BITRES);
|
cannam@154
|
2069 } else {
|
cannam@154
|
2070 base_target = IMAX(0, vbr_rate - ((9*C+4)<<BITRES));
|
cannam@154
|
2071 }
|
cannam@154
|
2072
|
cannam@154
|
2073 if (st->constrained_vbr)
|
cannam@154
|
2074 base_target += (st->vbr_offset>>lm_diff);
|
cannam@154
|
2075
|
cannam@154
|
2076 if (!hybrid)
|
cannam@154
|
2077 {
|
cannam@154
|
2078 target = compute_vbr(mode, &st->analysis, base_target, LM, equiv_rate,
|
cannam@154
|
2079 st->lastCodedBands, C, st->intensity, st->constrained_vbr,
|
cannam@154
|
2080 st->stereo_saving, tot_boost, tf_estimate, pitch_change, maxDepth,
|
cannam@154
|
2081 st->lfe, st->energy_mask!=NULL, surround_masking,
|
cannam@154
|
2082 temporal_vbr);
|
cannam@154
|
2083 } else {
|
cannam@154
|
2084 target = base_target;
|
cannam@154
|
2085 /* Tonal frames (offset<100) need more bits than noisy (offset>100) ones. */
|
cannam@154
|
2086 if (st->silk_info.offset < 100) target += 12 << BITRES >> (3-LM);
|
cannam@154
|
2087 if (st->silk_info.offset > 100) target -= 18 << BITRES >> (3-LM);
|
cannam@154
|
2088 /* Boosting bitrate on transients and vowels with significant temporal
|
cannam@154
|
2089 spikes. */
|
cannam@154
|
2090 target += (opus_int32)MULT16_16_Q14(tf_estimate-QCONST16(.25f,14), (50<<BITRES));
|
cannam@154
|
2091 /* If we have a strong transient, let's make sure it has enough bits to code
|
cannam@154
|
2092 the first two bands, so that it can use folding rather than noise. */
|
cannam@154
|
2093 if (tf_estimate > QCONST16(.7f,14))
|
cannam@154
|
2094 target = IMAX(target, 50<<BITRES);
|
cannam@154
|
2095 }
|
cannam@154
|
2096 /* The current offset is removed from the target and the space used
|
cannam@154
|
2097 so far is added*/
|
cannam@154
|
2098 target=target+tell;
|
cannam@154
|
2099 /* In VBR mode the frame size must not be reduced so much that it would
|
cannam@154
|
2100 result in the encoder running out of bits.
|
cannam@154
|
2101 The margin of 2 bytes ensures that none of the bust-prevention logic
|
cannam@154
|
2102 in the decoder will have triggered so far. */
|
cannam@154
|
2103 min_allowed = ((tell+total_boost+(1<<(BITRES+3))-1)>>(BITRES+3)) + 2;
|
cannam@154
|
2104 /* Take into account the 37 bits we need to have left in the packet to
|
cannam@154
|
2105 signal a redundant frame in hybrid mode. Creating a shorter packet would
|
cannam@154
|
2106 create an entropy coder desync. */
|
cannam@154
|
2107 if (hybrid)
|
cannam@154
|
2108 min_allowed = IMAX(min_allowed, (tell0_frac+(37<<BITRES)+total_boost+(1<<(BITRES+3))-1)>>(BITRES+3));
|
cannam@154
|
2109
|
cannam@154
|
2110 nbAvailableBytes = (target+(1<<(BITRES+2)))>>(BITRES+3);
|
cannam@154
|
2111 nbAvailableBytes = IMAX(min_allowed,nbAvailableBytes);
|
cannam@154
|
2112 nbAvailableBytes = IMIN(nbCompressedBytes,nbAvailableBytes);
|
cannam@154
|
2113
|
cannam@154
|
2114 /* By how much did we "miss" the target on that frame */
|
cannam@154
|
2115 delta = target - vbr_rate;
|
cannam@154
|
2116
|
cannam@154
|
2117 target=nbAvailableBytes<<(BITRES+3);
|
cannam@154
|
2118
|
cannam@154
|
2119 /*If the frame is silent we don't adjust our drift, otherwise
|
cannam@154
|
2120 the encoder will shoot to very high rates after hitting a
|
cannam@154
|
2121 span of silence, but we do allow the bitres to refill.
|
cannam@154
|
2122 This means that we'll undershoot our target in CVBR/VBR modes
|
cannam@154
|
2123 on files with lots of silence. */
|
cannam@154
|
2124 if(silence)
|
cannam@154
|
2125 {
|
cannam@154
|
2126 nbAvailableBytes = 2;
|
cannam@154
|
2127 target = 2*8<<BITRES;
|
cannam@154
|
2128 delta = 0;
|
cannam@154
|
2129 }
|
cannam@154
|
2130
|
cannam@154
|
2131 if (st->vbr_count < 970)
|
cannam@154
|
2132 {
|
cannam@154
|
2133 st->vbr_count++;
|
cannam@154
|
2134 alpha = celt_rcp(SHL32(EXTEND32(st->vbr_count+20),16));
|
cannam@154
|
2135 } else
|
cannam@154
|
2136 alpha = QCONST16(.001f,15);
|
cannam@154
|
2137 /* How many bits have we used in excess of what we're allowed */
|
cannam@154
|
2138 if (st->constrained_vbr)
|
cannam@154
|
2139 st->vbr_reservoir += target - vbr_rate;
|
cannam@154
|
2140 /*printf ("%d\n", st->vbr_reservoir);*/
|
cannam@154
|
2141
|
cannam@154
|
2142 /* Compute the offset we need to apply in order to reach the target */
|
cannam@154
|
2143 if (st->constrained_vbr)
|
cannam@154
|
2144 {
|
cannam@154
|
2145 st->vbr_drift += (opus_int32)MULT16_32_Q15(alpha,(delta*(1<<lm_diff))-st->vbr_offset-st->vbr_drift);
|
cannam@154
|
2146 st->vbr_offset = -st->vbr_drift;
|
cannam@154
|
2147 }
|
cannam@154
|
2148 /*printf ("%d\n", st->vbr_drift);*/
|
cannam@154
|
2149
|
cannam@154
|
2150 if (st->constrained_vbr && st->vbr_reservoir < 0)
|
cannam@154
|
2151 {
|
cannam@154
|
2152 /* We're under the min value -- increase rate */
|
cannam@154
|
2153 int adjust = (-st->vbr_reservoir)/(8<<BITRES);
|
cannam@154
|
2154 /* Unless we're just coding silence */
|
cannam@154
|
2155 nbAvailableBytes += silence?0:adjust;
|
cannam@154
|
2156 st->vbr_reservoir = 0;
|
cannam@154
|
2157 /*printf ("+%d\n", adjust);*/
|
cannam@154
|
2158 }
|
cannam@154
|
2159 nbCompressedBytes = IMIN(nbCompressedBytes,nbAvailableBytes);
|
cannam@154
|
2160 /*printf("%d\n", nbCompressedBytes*50*8);*/
|
cannam@154
|
2161 /* This moves the raw bits to take into account the new compressed size */
|
cannam@154
|
2162 ec_enc_shrink(enc, nbCompressedBytes);
|
cannam@154
|
2163 }
|
cannam@154
|
2164
|
cannam@154
|
2165 /* Bit allocation */
|
cannam@154
|
2166 ALLOC(fine_quant, nbEBands, int);
|
cannam@154
|
2167 ALLOC(pulses, nbEBands, int);
|
cannam@154
|
2168 ALLOC(fine_priority, nbEBands, int);
|
cannam@154
|
2169
|
cannam@154
|
2170 /* bits = packet size - where we are - safety*/
|
cannam@154
|
2171 bits = (((opus_int32)nbCompressedBytes*8)<<BITRES) - ec_tell_frac(enc) - 1;
|
cannam@154
|
2172 anti_collapse_rsv = isTransient&&LM>=2&&bits>=((LM+2)<<BITRES) ? (1<<BITRES) : 0;
|
cannam@154
|
2173 bits -= anti_collapse_rsv;
|
cannam@154
|
2174 signalBandwidth = end-1;
|
cannam@154
|
2175 #ifndef DISABLE_FLOAT_API
|
cannam@154
|
2176 if (st->analysis.valid)
|
cannam@154
|
2177 {
|
cannam@154
|
2178 int min_bandwidth;
|
cannam@154
|
2179 if (equiv_rate < (opus_int32)32000*C)
|
cannam@154
|
2180 min_bandwidth = 13;
|
cannam@154
|
2181 else if (equiv_rate < (opus_int32)48000*C)
|
cannam@154
|
2182 min_bandwidth = 16;
|
cannam@154
|
2183 else if (equiv_rate < (opus_int32)60000*C)
|
cannam@154
|
2184 min_bandwidth = 18;
|
cannam@154
|
2185 else if (equiv_rate < (opus_int32)80000*C)
|
cannam@154
|
2186 min_bandwidth = 19;
|
cannam@154
|
2187 else
|
cannam@154
|
2188 min_bandwidth = 20;
|
cannam@154
|
2189 signalBandwidth = IMAX(st->analysis.bandwidth, min_bandwidth);
|
cannam@154
|
2190 }
|
cannam@154
|
2191 #endif
|
cannam@154
|
2192 if (st->lfe)
|
cannam@154
|
2193 signalBandwidth = 1;
|
cannam@154
|
2194 codedBands = clt_compute_allocation(mode, start, end, offsets, cap,
|
cannam@154
|
2195 alloc_trim, &st->intensity, &dual_stereo, bits, &balance, pulses,
|
cannam@154
|
2196 fine_quant, fine_priority, C, LM, enc, 1, st->lastCodedBands, signalBandwidth);
|
cannam@154
|
2197 if (st->lastCodedBands)
|
cannam@154
|
2198 st->lastCodedBands = IMIN(st->lastCodedBands+1,IMAX(st->lastCodedBands-1,codedBands));
|
cannam@154
|
2199 else
|
cannam@154
|
2200 st->lastCodedBands = codedBands;
|
cannam@154
|
2201
|
cannam@154
|
2202 quant_fine_energy(mode, start, end, oldBandE, error, fine_quant, enc, C);
|
cannam@154
|
2203
|
cannam@154
|
2204 /* Residual quantisation */
|
cannam@154
|
2205 ALLOC(collapse_masks, C*nbEBands, unsigned char);
|
cannam@154
|
2206 quant_all_bands(1, mode, start, end, X, C==2 ? X+N : NULL, collapse_masks,
|
cannam@154
|
2207 bandE, pulses, shortBlocks, st->spread_decision,
|
cannam@154
|
2208 dual_stereo, st->intensity, tf_res, nbCompressedBytes*(8<<BITRES)-anti_collapse_rsv,
|
cannam@154
|
2209 balance, enc, LM, codedBands, &st->rng, st->complexity, st->arch, st->disable_inv);
|
cannam@154
|
2210
|
cannam@154
|
2211 if (anti_collapse_rsv > 0)
|
cannam@154
|
2212 {
|
cannam@154
|
2213 anti_collapse_on = st->consec_transient<2;
|
cannam@154
|
2214 #ifdef FUZZING
|
cannam@154
|
2215 anti_collapse_on = rand()&0x1;
|
cannam@154
|
2216 #endif
|
cannam@154
|
2217 ec_enc_bits(enc, anti_collapse_on, 1);
|
cannam@154
|
2218 }
|
cannam@154
|
2219 quant_energy_finalise(mode, start, end, oldBandE, error, fine_quant, fine_priority, nbCompressedBytes*8-ec_tell(enc), enc, C);
|
cannam@154
|
2220 OPUS_CLEAR(energyError, nbEBands*CC);
|
cannam@154
|
2221 c=0;
|
cannam@154
|
2222 do {
|
cannam@154
|
2223 for (i=start;i<end;i++)
|
cannam@154
|
2224 {
|
cannam@154
|
2225 energyError[i+c*nbEBands] = MAX16(-QCONST16(0.5f, 15), MIN16(QCONST16(0.5f, 15), error[i+c*nbEBands]));
|
cannam@154
|
2226 }
|
cannam@154
|
2227 } while (++c < C);
|
cannam@154
|
2228
|
cannam@154
|
2229 if (silence)
|
cannam@154
|
2230 {
|
cannam@154
|
2231 for (i=0;i<C*nbEBands;i++)
|
cannam@154
|
2232 oldBandE[i] = -QCONST16(28.f,DB_SHIFT);
|
cannam@154
|
2233 }
|
cannam@154
|
2234
|
cannam@154
|
2235 #ifdef RESYNTH
|
cannam@154
|
2236 /* Re-synthesis of the coded audio if required */
|
cannam@154
|
2237 {
|
cannam@154
|
2238 celt_sig *out_mem[2];
|
cannam@154
|
2239
|
cannam@154
|
2240 if (anti_collapse_on)
|
cannam@154
|
2241 {
|
cannam@154
|
2242 anti_collapse(mode, X, collapse_masks, LM, C, N,
|
cannam@154
|
2243 start, end, oldBandE, oldLogE, oldLogE2, pulses, st->rng);
|
cannam@154
|
2244 }
|
cannam@154
|
2245
|
cannam@154
|
2246 c=0; do {
|
cannam@154
|
2247 OPUS_MOVE(st->syn_mem[c], st->syn_mem[c]+N, 2*MAX_PERIOD-N+overlap/2);
|
cannam@154
|
2248 } while (++c<CC);
|
cannam@154
|
2249
|
cannam@154
|
2250 c=0; do {
|
cannam@154
|
2251 out_mem[c] = st->syn_mem[c]+2*MAX_PERIOD-N;
|
cannam@154
|
2252 } while (++c<CC);
|
cannam@154
|
2253
|
cannam@154
|
2254 celt_synthesis(mode, X, out_mem, oldBandE, start, effEnd,
|
cannam@154
|
2255 C, CC, isTransient, LM, st->upsample, silence, st->arch);
|
cannam@154
|
2256
|
cannam@154
|
2257 c=0; do {
|
cannam@154
|
2258 st->prefilter_period=IMAX(st->prefilter_period, COMBFILTER_MINPERIOD);
|
cannam@154
|
2259 st->prefilter_period_old=IMAX(st->prefilter_period_old, COMBFILTER_MINPERIOD);
|
cannam@154
|
2260 comb_filter(out_mem[c], out_mem[c], st->prefilter_period_old, st->prefilter_period, mode->shortMdctSize,
|
cannam@154
|
2261 st->prefilter_gain_old, st->prefilter_gain, st->prefilter_tapset_old, st->prefilter_tapset,
|
cannam@154
|
2262 mode->window, overlap);
|
cannam@154
|
2263 if (LM!=0)
|
cannam@154
|
2264 comb_filter(out_mem[c]+mode->shortMdctSize, out_mem[c]+mode->shortMdctSize, st->prefilter_period, pitch_index, N-mode->shortMdctSize,
|
cannam@154
|
2265 st->prefilter_gain, gain1, st->prefilter_tapset, prefilter_tapset,
|
cannam@154
|
2266 mode->window, overlap);
|
cannam@154
|
2267 } while (++c<CC);
|
cannam@154
|
2268
|
cannam@154
|
2269 /* We reuse freq[] as scratch space for the de-emphasis */
|
cannam@154
|
2270 deemphasis(out_mem, (opus_val16*)pcm, N, CC, st->upsample, mode->preemph, st->preemph_memD);
|
cannam@154
|
2271 st->prefilter_period_old = st->prefilter_period;
|
cannam@154
|
2272 st->prefilter_gain_old = st->prefilter_gain;
|
cannam@154
|
2273 st->prefilter_tapset_old = st->prefilter_tapset;
|
cannam@154
|
2274 }
|
cannam@154
|
2275 #endif
|
cannam@154
|
2276
|
cannam@154
|
2277 st->prefilter_period = pitch_index;
|
cannam@154
|
2278 st->prefilter_gain = gain1;
|
cannam@154
|
2279 st->prefilter_tapset = prefilter_tapset;
|
cannam@154
|
2280 #ifdef RESYNTH
|
cannam@154
|
2281 if (LM!=0)
|
cannam@154
|
2282 {
|
cannam@154
|
2283 st->prefilter_period_old = st->prefilter_period;
|
cannam@154
|
2284 st->prefilter_gain_old = st->prefilter_gain;
|
cannam@154
|
2285 st->prefilter_tapset_old = st->prefilter_tapset;
|
cannam@154
|
2286 }
|
cannam@154
|
2287 #endif
|
cannam@154
|
2288
|
cannam@154
|
2289 if (CC==2&&C==1) {
|
cannam@154
|
2290 OPUS_COPY(&oldBandE[nbEBands], oldBandE, nbEBands);
|
cannam@154
|
2291 }
|
cannam@154
|
2292
|
cannam@154
|
2293 if (!isTransient)
|
cannam@154
|
2294 {
|
cannam@154
|
2295 OPUS_COPY(oldLogE2, oldLogE, CC*nbEBands);
|
cannam@154
|
2296 OPUS_COPY(oldLogE, oldBandE, CC*nbEBands);
|
cannam@154
|
2297 } else {
|
cannam@154
|
2298 for (i=0;i<CC*nbEBands;i++)
|
cannam@154
|
2299 oldLogE[i] = MIN16(oldLogE[i], oldBandE[i]);
|
cannam@154
|
2300 }
|
cannam@154
|
2301 /* In case start or end were to change */
|
cannam@154
|
2302 c=0; do
|
cannam@154
|
2303 {
|
cannam@154
|
2304 for (i=0;i<start;i++)
|
cannam@154
|
2305 {
|
cannam@154
|
2306 oldBandE[c*nbEBands+i]=0;
|
cannam@154
|
2307 oldLogE[c*nbEBands+i]=oldLogE2[c*nbEBands+i]=-QCONST16(28.f,DB_SHIFT);
|
cannam@154
|
2308 }
|
cannam@154
|
2309 for (i=end;i<nbEBands;i++)
|
cannam@154
|
2310 {
|
cannam@154
|
2311 oldBandE[c*nbEBands+i]=0;
|
cannam@154
|
2312 oldLogE[c*nbEBands+i]=oldLogE2[c*nbEBands+i]=-QCONST16(28.f,DB_SHIFT);
|
cannam@154
|
2313 }
|
cannam@154
|
2314 } while (++c<CC);
|
cannam@154
|
2315
|
cannam@154
|
2316 if (isTransient || transient_got_disabled)
|
cannam@154
|
2317 st->consec_transient++;
|
cannam@154
|
2318 else
|
cannam@154
|
2319 st->consec_transient=0;
|
cannam@154
|
2320 st->rng = enc->rng;
|
cannam@154
|
2321
|
cannam@154
|
2322 /* If there's any room left (can only happen for very high rates),
|
cannam@154
|
2323 it's already filled with zeros */
|
cannam@154
|
2324 ec_enc_done(enc);
|
cannam@154
|
2325
|
cannam@154
|
2326 #ifdef CUSTOM_MODES
|
cannam@154
|
2327 if (st->signalling)
|
cannam@154
|
2328 nbCompressedBytes++;
|
cannam@154
|
2329 #endif
|
cannam@154
|
2330
|
cannam@154
|
2331 RESTORE_STACK;
|
cannam@154
|
2332 if (ec_get_error(enc))
|
cannam@154
|
2333 return OPUS_INTERNAL_ERROR;
|
cannam@154
|
2334 else
|
cannam@154
|
2335 return nbCompressedBytes;
|
cannam@154
|
2336 }
|
cannam@154
|
2337
|
cannam@154
|
2338
|
cannam@154
|
2339 #ifdef CUSTOM_MODES
|
cannam@154
|
2340
|
cannam@154
|
2341 #ifdef FIXED_POINT
|
cannam@154
|
2342 int opus_custom_encode(CELTEncoder * OPUS_RESTRICT st, const opus_int16 * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes)
|
cannam@154
|
2343 {
|
cannam@154
|
2344 return celt_encode_with_ec(st, pcm, frame_size, compressed, nbCompressedBytes, NULL);
|
cannam@154
|
2345 }
|
cannam@154
|
2346
|
cannam@154
|
2347 #ifndef DISABLE_FLOAT_API
|
cannam@154
|
2348 int opus_custom_encode_float(CELTEncoder * OPUS_RESTRICT st, const float * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes)
|
cannam@154
|
2349 {
|
cannam@154
|
2350 int j, ret, C, N;
|
cannam@154
|
2351 VARDECL(opus_int16, in);
|
cannam@154
|
2352 ALLOC_STACK;
|
cannam@154
|
2353
|
cannam@154
|
2354 if (pcm==NULL)
|
cannam@154
|
2355 return OPUS_BAD_ARG;
|
cannam@154
|
2356
|
cannam@154
|
2357 C = st->channels;
|
cannam@154
|
2358 N = frame_size;
|
cannam@154
|
2359 ALLOC(in, C*N, opus_int16);
|
cannam@154
|
2360
|
cannam@154
|
2361 for (j=0;j<C*N;j++)
|
cannam@154
|
2362 in[j] = FLOAT2INT16(pcm[j]);
|
cannam@154
|
2363
|
cannam@154
|
2364 ret=celt_encode_with_ec(st,in,frame_size,compressed,nbCompressedBytes, NULL);
|
cannam@154
|
2365 #ifdef RESYNTH
|
cannam@154
|
2366 for (j=0;j<C*N;j++)
|
cannam@154
|
2367 ((float*)pcm)[j]=in[j]*(1.f/32768.f);
|
cannam@154
|
2368 #endif
|
cannam@154
|
2369 RESTORE_STACK;
|
cannam@154
|
2370 return ret;
|
cannam@154
|
2371 }
|
cannam@154
|
2372 #endif /* DISABLE_FLOAT_API */
|
cannam@154
|
2373 #else
|
cannam@154
|
2374
|
cannam@154
|
2375 int opus_custom_encode(CELTEncoder * OPUS_RESTRICT st, const opus_int16 * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes)
|
cannam@154
|
2376 {
|
cannam@154
|
2377 int j, ret, C, N;
|
cannam@154
|
2378 VARDECL(celt_sig, in);
|
cannam@154
|
2379 ALLOC_STACK;
|
cannam@154
|
2380
|
cannam@154
|
2381 if (pcm==NULL)
|
cannam@154
|
2382 return OPUS_BAD_ARG;
|
cannam@154
|
2383
|
cannam@154
|
2384 C=st->channels;
|
cannam@154
|
2385 N=frame_size;
|
cannam@154
|
2386 ALLOC(in, C*N, celt_sig);
|
cannam@154
|
2387 for (j=0;j<C*N;j++) {
|
cannam@154
|
2388 in[j] = SCALEOUT(pcm[j]);
|
cannam@154
|
2389 }
|
cannam@154
|
2390
|
cannam@154
|
2391 ret = celt_encode_with_ec(st,in,frame_size,compressed,nbCompressedBytes, NULL);
|
cannam@154
|
2392 #ifdef RESYNTH
|
cannam@154
|
2393 for (j=0;j<C*N;j++)
|
cannam@154
|
2394 ((opus_int16*)pcm)[j] = FLOAT2INT16(in[j]);
|
cannam@154
|
2395 #endif
|
cannam@154
|
2396 RESTORE_STACK;
|
cannam@154
|
2397 return ret;
|
cannam@154
|
2398 }
|
cannam@154
|
2399
|
cannam@154
|
2400 int opus_custom_encode_float(CELTEncoder * OPUS_RESTRICT st, const float * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes)
|
cannam@154
|
2401 {
|
cannam@154
|
2402 return celt_encode_with_ec(st, pcm, frame_size, compressed, nbCompressedBytes, NULL);
|
cannam@154
|
2403 }
|
cannam@154
|
2404
|
cannam@154
|
2405 #endif
|
cannam@154
|
2406
|
cannam@154
|
2407 #endif /* CUSTOM_MODES */
|
cannam@154
|
2408
|
cannam@154
|
2409 int opus_custom_encoder_ctl(CELTEncoder * OPUS_RESTRICT st, int request, ...)
|
cannam@154
|
2410 {
|
cannam@154
|
2411 va_list ap;
|
cannam@154
|
2412
|
cannam@154
|
2413 va_start(ap, request);
|
cannam@154
|
2414 switch (request)
|
cannam@154
|
2415 {
|
cannam@154
|
2416 case OPUS_SET_COMPLEXITY_REQUEST:
|
cannam@154
|
2417 {
|
cannam@154
|
2418 int value = va_arg(ap, opus_int32);
|
cannam@154
|
2419 if (value<0 || value>10)
|
cannam@154
|
2420 goto bad_arg;
|
cannam@154
|
2421 st->complexity = value;
|
cannam@154
|
2422 }
|
cannam@154
|
2423 break;
|
cannam@154
|
2424 case CELT_SET_START_BAND_REQUEST:
|
cannam@154
|
2425 {
|
cannam@154
|
2426 opus_int32 value = va_arg(ap, opus_int32);
|
cannam@154
|
2427 if (value<0 || value>=st->mode->nbEBands)
|
cannam@154
|
2428 goto bad_arg;
|
cannam@154
|
2429 st->start = value;
|
cannam@154
|
2430 }
|
cannam@154
|
2431 break;
|
cannam@154
|
2432 case CELT_SET_END_BAND_REQUEST:
|
cannam@154
|
2433 {
|
cannam@154
|
2434 opus_int32 value = va_arg(ap, opus_int32);
|
cannam@154
|
2435 if (value<1 || value>st->mode->nbEBands)
|
cannam@154
|
2436 goto bad_arg;
|
cannam@154
|
2437 st->end = value;
|
cannam@154
|
2438 }
|
cannam@154
|
2439 break;
|
cannam@154
|
2440 case CELT_SET_PREDICTION_REQUEST:
|
cannam@154
|
2441 {
|
cannam@154
|
2442 int value = va_arg(ap, opus_int32);
|
cannam@154
|
2443 if (value<0 || value>2)
|
cannam@154
|
2444 goto bad_arg;
|
cannam@154
|
2445 st->disable_pf = value<=1;
|
cannam@154
|
2446 st->force_intra = value==0;
|
cannam@154
|
2447 }
|
cannam@154
|
2448 break;
|
cannam@154
|
2449 case OPUS_SET_PACKET_LOSS_PERC_REQUEST:
|
cannam@154
|
2450 {
|
cannam@154
|
2451 int value = va_arg(ap, opus_int32);
|
cannam@154
|
2452 if (value<0 || value>100)
|
cannam@154
|
2453 goto bad_arg;
|
cannam@154
|
2454 st->loss_rate = value;
|
cannam@154
|
2455 }
|
cannam@154
|
2456 break;
|
cannam@154
|
2457 case OPUS_SET_VBR_CONSTRAINT_REQUEST:
|
cannam@154
|
2458 {
|
cannam@154
|
2459 opus_int32 value = va_arg(ap, opus_int32);
|
cannam@154
|
2460 st->constrained_vbr = value;
|
cannam@154
|
2461 }
|
cannam@154
|
2462 break;
|
cannam@154
|
2463 case OPUS_SET_VBR_REQUEST:
|
cannam@154
|
2464 {
|
cannam@154
|
2465 opus_int32 value = va_arg(ap, opus_int32);
|
cannam@154
|
2466 st->vbr = value;
|
cannam@154
|
2467 }
|
cannam@154
|
2468 break;
|
cannam@154
|
2469 case OPUS_SET_BITRATE_REQUEST:
|
cannam@154
|
2470 {
|
cannam@154
|
2471 opus_int32 value = va_arg(ap, opus_int32);
|
cannam@154
|
2472 if (value<=500 && value!=OPUS_BITRATE_MAX)
|
cannam@154
|
2473 goto bad_arg;
|
cannam@154
|
2474 value = IMIN(value, 260000*st->channels);
|
cannam@154
|
2475 st->bitrate = value;
|
cannam@154
|
2476 }
|
cannam@154
|
2477 break;
|
cannam@154
|
2478 case CELT_SET_CHANNELS_REQUEST:
|
cannam@154
|
2479 {
|
cannam@154
|
2480 opus_int32 value = va_arg(ap, opus_int32);
|
cannam@154
|
2481 if (value<1 || value>2)
|
cannam@154
|
2482 goto bad_arg;
|
cannam@154
|
2483 st->stream_channels = value;
|
cannam@154
|
2484 }
|
cannam@154
|
2485 break;
|
cannam@154
|
2486 case OPUS_SET_LSB_DEPTH_REQUEST:
|
cannam@154
|
2487 {
|
cannam@154
|
2488 opus_int32 value = va_arg(ap, opus_int32);
|
cannam@154
|
2489 if (value<8 || value>24)
|
cannam@154
|
2490 goto bad_arg;
|
cannam@154
|
2491 st->lsb_depth=value;
|
cannam@154
|
2492 }
|
cannam@154
|
2493 break;
|
cannam@154
|
2494 case OPUS_GET_LSB_DEPTH_REQUEST:
|
cannam@154
|
2495 {
|
cannam@154
|
2496 opus_int32 *value = va_arg(ap, opus_int32*);
|
cannam@154
|
2497 *value=st->lsb_depth;
|
cannam@154
|
2498 }
|
cannam@154
|
2499 break;
|
cannam@154
|
2500 case OPUS_SET_PHASE_INVERSION_DISABLED_REQUEST:
|
cannam@154
|
2501 {
|
cannam@154
|
2502 opus_int32 value = va_arg(ap, opus_int32);
|
cannam@154
|
2503 if(value<0 || value>1)
|
cannam@154
|
2504 {
|
cannam@154
|
2505 goto bad_arg;
|
cannam@154
|
2506 }
|
cannam@154
|
2507 st->disable_inv = value;
|
cannam@154
|
2508 }
|
cannam@154
|
2509 break;
|
cannam@154
|
2510 case OPUS_GET_PHASE_INVERSION_DISABLED_REQUEST:
|
cannam@154
|
2511 {
|
cannam@154
|
2512 opus_int32 *value = va_arg(ap, opus_int32*);
|
cannam@154
|
2513 if (!value)
|
cannam@154
|
2514 {
|
cannam@154
|
2515 goto bad_arg;
|
cannam@154
|
2516 }
|
cannam@154
|
2517 *value = st->disable_inv;
|
cannam@154
|
2518 }
|
cannam@154
|
2519 break;
|
cannam@154
|
2520 case OPUS_RESET_STATE:
|
cannam@154
|
2521 {
|
cannam@154
|
2522 int i;
|
cannam@154
|
2523 opus_val16 *oldBandE, *oldLogE, *oldLogE2;
|
cannam@154
|
2524 oldBandE = (opus_val16*)(st->in_mem+st->channels*(st->mode->overlap+COMBFILTER_MAXPERIOD));
|
cannam@154
|
2525 oldLogE = oldBandE + st->channels*st->mode->nbEBands;
|
cannam@154
|
2526 oldLogE2 = oldLogE + st->channels*st->mode->nbEBands;
|
cannam@154
|
2527 OPUS_CLEAR((char*)&st->ENCODER_RESET_START,
|
cannam@154
|
2528 opus_custom_encoder_get_size(st->mode, st->channels)-
|
cannam@154
|
2529 ((char*)&st->ENCODER_RESET_START - (char*)st));
|
cannam@154
|
2530 for (i=0;i<st->channels*st->mode->nbEBands;i++)
|
cannam@154
|
2531 oldLogE[i]=oldLogE2[i]=-QCONST16(28.f,DB_SHIFT);
|
cannam@154
|
2532 st->vbr_offset = 0;
|
cannam@154
|
2533 st->delayedIntra = 1;
|
cannam@154
|
2534 st->spread_decision = SPREAD_NORMAL;
|
cannam@154
|
2535 st->tonal_average = 256;
|
cannam@154
|
2536 st->hf_average = 0;
|
cannam@154
|
2537 st->tapset_decision = 0;
|
cannam@154
|
2538 }
|
cannam@154
|
2539 break;
|
cannam@154
|
2540 #ifdef CUSTOM_MODES
|
cannam@154
|
2541 case CELT_SET_INPUT_CLIPPING_REQUEST:
|
cannam@154
|
2542 {
|
cannam@154
|
2543 opus_int32 value = va_arg(ap, opus_int32);
|
cannam@154
|
2544 st->clip = value;
|
cannam@154
|
2545 }
|
cannam@154
|
2546 break;
|
cannam@154
|
2547 #endif
|
cannam@154
|
2548 case CELT_SET_SIGNALLING_REQUEST:
|
cannam@154
|
2549 {
|
cannam@154
|
2550 opus_int32 value = va_arg(ap, opus_int32);
|
cannam@154
|
2551 st->signalling = value;
|
cannam@154
|
2552 }
|
cannam@154
|
2553 break;
|
cannam@154
|
2554 case CELT_SET_ANALYSIS_REQUEST:
|
cannam@154
|
2555 {
|
cannam@154
|
2556 AnalysisInfo *info = va_arg(ap, AnalysisInfo *);
|
cannam@154
|
2557 if (info)
|
cannam@154
|
2558 OPUS_COPY(&st->analysis, info, 1);
|
cannam@154
|
2559 }
|
cannam@154
|
2560 break;
|
cannam@154
|
2561 case CELT_SET_SILK_INFO_REQUEST:
|
cannam@154
|
2562 {
|
cannam@154
|
2563 SILKInfo *info = va_arg(ap, SILKInfo *);
|
cannam@154
|
2564 if (info)
|
cannam@154
|
2565 OPUS_COPY(&st->silk_info, info, 1);
|
cannam@154
|
2566 }
|
cannam@154
|
2567 break;
|
cannam@154
|
2568 case CELT_GET_MODE_REQUEST:
|
cannam@154
|
2569 {
|
cannam@154
|
2570 const CELTMode ** value = va_arg(ap, const CELTMode**);
|
cannam@154
|
2571 if (value==0)
|
cannam@154
|
2572 goto bad_arg;
|
cannam@154
|
2573 *value=st->mode;
|
cannam@154
|
2574 }
|
cannam@154
|
2575 break;
|
cannam@154
|
2576 case OPUS_GET_FINAL_RANGE_REQUEST:
|
cannam@154
|
2577 {
|
cannam@154
|
2578 opus_uint32 * value = va_arg(ap, opus_uint32 *);
|
cannam@154
|
2579 if (value==0)
|
cannam@154
|
2580 goto bad_arg;
|
cannam@154
|
2581 *value=st->rng;
|
cannam@154
|
2582 }
|
cannam@154
|
2583 break;
|
cannam@154
|
2584 case OPUS_SET_LFE_REQUEST:
|
cannam@154
|
2585 {
|
cannam@154
|
2586 opus_int32 value = va_arg(ap, opus_int32);
|
cannam@154
|
2587 st->lfe = value;
|
cannam@154
|
2588 }
|
cannam@154
|
2589 break;
|
cannam@154
|
2590 case OPUS_SET_ENERGY_MASK_REQUEST:
|
cannam@154
|
2591 {
|
cannam@154
|
2592 opus_val16 *value = va_arg(ap, opus_val16*);
|
cannam@154
|
2593 st->energy_mask = value;
|
cannam@154
|
2594 }
|
cannam@154
|
2595 break;
|
cannam@154
|
2596 default:
|
cannam@154
|
2597 goto bad_request;
|
cannam@154
|
2598 }
|
cannam@154
|
2599 va_end(ap);
|
cannam@154
|
2600 return OPUS_OK;
|
cannam@154
|
2601 bad_arg:
|
cannam@154
|
2602 va_end(ap);
|
cannam@154
|
2603 return OPUS_BAD_ARG;
|
cannam@154
|
2604 bad_request:
|
cannam@154
|
2605 va_end(ap);
|
cannam@154
|
2606 return OPUS_UNIMPLEMENTED;
|
cannam@154
|
2607 }
|