OLD | NEW |
| (Empty) |
1 /* | |
2 * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. | |
3 * | |
4 * Use of this source code is governed by a BSD-style license | |
5 * that can be found in the LICENSE file in the root of the source | |
6 * tree. An additional intellectual property rights grant can be found | |
7 * in the file PATENTS. All contributing project authors may | |
8 * be found in the AUTHORS file in the root of the source tree. | |
9 */ | |
10 | |
11 /* | |
12 * Contains the API functions for the AEC. | |
13 */ | |
14 #include "webrtc/modules/audio_processing/aec/echo_cancellation.h" | |
15 | |
16 #include <math.h> | |
17 #ifdef WEBRTC_AEC_DEBUG_DUMP | |
18 #include <stdio.h> | |
19 #endif | |
20 #include <stdlib.h> | |
21 #include <string.h> | |
22 | |
23 #include "webrtc/common_audio/ring_buffer.h" | |
24 #include "webrtc/common_audio/signal_processing/include/signal_processing_librar
y.h" | |
25 #include "webrtc/modules/audio_processing/aec/aec_core.h" | |
26 #include "webrtc/modules/audio_processing/aec/aec_resampler.h" | |
27 #include "webrtc/modules/audio_processing/aec/echo_cancellation_internal.h" | |
28 #include "webrtc/typedefs.h" | |
29 | |
30 // Measured delays [ms] | |
31 // Device Chrome GTP | |
32 // MacBook Air 10 | |
33 // MacBook Retina 10 100 | |
34 // MacPro 30? | |
35 // | |
36 // Win7 Desktop 70 80? | |
37 // Win7 T430s 110 | |
38 // Win8 T420s 70 | |
39 // | |
40 // Daisy 50 | |
41 // Pixel (w/ preproc?) 240 | |
42 // Pixel (w/o preproc?) 110 110 | |
43 | |
44 // The extended filter mode gives us the flexibility to ignore the system's | |
45 // reported delays. We do this for platforms which we believe provide results | |
46 // which are incompatible with the AEC's expectations. Based on measurements | |
47 // (some provided above) we set a conservative (i.e. lower than measured) | |
48 // fixed delay. | |
49 // | |
50 // WEBRTC_UNTRUSTED_DELAY will only have an impact when |extended_filter_mode| | |
51 // is enabled. See the note along with |DelayCorrection| in | |
52 // echo_cancellation_impl.h for more details on the mode. | |
53 // | |
54 // Justification: | |
55 // Chromium/Mac: Here, the true latency is so low (~10-20 ms), that it plays | |
56 // havoc with the AEC's buffering. To avoid this, we set a fixed delay of 20 ms | |
57 // and then compensate by rewinding by 10 ms (in wideband) through | |
58 // kDelayDiffOffsetSamples. This trick does not seem to work for larger rewind | |
59 // values, but fortunately this is sufficient. | |
60 // | |
61 // Chromium/Linux(ChromeOS): The values we get on this platform don't correspond | |
62 // well to reality. The variance doesn't match the AEC's buffer changes, and the | |
63 // bulk values tend to be too low. However, the range across different hardware | |
64 // appears to be too large to choose a single value. | |
65 // | |
66 // GTP/Linux(ChromeOS): TBD, but for the moment we will trust the values. | |
67 #if defined(WEBRTC_CHROMIUM_BUILD) && defined(WEBRTC_MAC) | |
68 #define WEBRTC_UNTRUSTED_DELAY | |
69 #endif | |
70 | |
71 #if defined(WEBRTC_UNTRUSTED_DELAY) && defined(WEBRTC_MAC) | |
72 static const int kDelayDiffOffsetSamples = -160; | |
73 #else | |
74 // Not enabled for now. | |
75 static const int kDelayDiffOffsetSamples = 0; | |
76 #endif | |
77 | |
78 #if defined(WEBRTC_MAC) | |
79 static const int kFixedDelayMs = 20; | |
80 #else | |
81 static const int kFixedDelayMs = 50; | |
82 #endif | |
83 #if !defined(WEBRTC_UNTRUSTED_DELAY) | |
84 static const int kMinTrustedDelayMs = 20; | |
85 #endif | |
86 static const int kMaxTrustedDelayMs = 500; | |
87 | |
88 // Maximum length of resampled signal. Must be an integer multiple of frames | |
89 // (ceil(1/(1 + MIN_SKEW)*2) + 1)*FRAME_LEN | |
90 // The factor of 2 handles wb, and the + 1 is as a safety margin | |
91 // TODO(bjornv): Replace with kResamplerBufferSize | |
92 #define MAX_RESAMP_LEN (5 * FRAME_LEN) | |
93 | |
94 static const int kMaxBufSizeStart = 62; // In partitions | |
95 static const int sampMsNb = 8; // samples per ms in nb | |
96 static const int initCheck = 42; | |
97 | |
98 #ifdef WEBRTC_AEC_DEBUG_DUMP | |
99 int webrtc_aec_instance_count = 0; | |
100 #endif | |
101 | |
102 // Estimates delay to set the position of the far-end buffer read pointer | |
103 // (controlled by knownDelay) | |
104 static void EstBufDelayNormal(Aec* aecInst); | |
105 static void EstBufDelayExtended(Aec* aecInst); | |
106 static int ProcessNormal(Aec* self, | |
107 const float* const* near, | |
108 size_t num_bands, | |
109 float* const* out, | |
110 size_t num_samples, | |
111 int16_t reported_delay_ms, | |
112 int32_t skew); | |
113 static void ProcessExtended(Aec* self, | |
114 const float* const* near, | |
115 size_t num_bands, | |
116 float* const* out, | |
117 size_t num_samples, | |
118 int16_t reported_delay_ms, | |
119 int32_t skew); | |
120 | |
121 void* WebRtcAec_Create() { | |
122 Aec* aecpc = malloc(sizeof(Aec)); | |
123 | |
124 if (!aecpc) { | |
125 return NULL; | |
126 } | |
127 | |
128 aecpc->aec = WebRtcAec_CreateAec(); | |
129 if (!aecpc->aec) { | |
130 WebRtcAec_Free(aecpc); | |
131 return NULL; | |
132 } | |
133 aecpc->resampler = WebRtcAec_CreateResampler(); | |
134 if (!aecpc->resampler) { | |
135 WebRtcAec_Free(aecpc); | |
136 return NULL; | |
137 } | |
138 // Create far-end pre-buffer. The buffer size has to be large enough for | |
139 // largest possible drift compensation (kResamplerBufferSize) + "almost" an | |
140 // FFT buffer (PART_LEN2 - 1). | |
141 aecpc->far_pre_buf = | |
142 WebRtc_CreateBuffer(PART_LEN2 + kResamplerBufferSize, sizeof(float)); | |
143 if (!aecpc->far_pre_buf) { | |
144 WebRtcAec_Free(aecpc); | |
145 return NULL; | |
146 } | |
147 | |
148 aecpc->initFlag = 0; | |
149 | |
150 #ifdef WEBRTC_AEC_DEBUG_DUMP | |
151 { | |
152 char filename[64]; | |
153 sprintf(filename, "aec_buf%d.dat", webrtc_aec_instance_count); | |
154 aecpc->bufFile = fopen(filename, "wb"); | |
155 sprintf(filename, "aec_skew%d.dat", webrtc_aec_instance_count); | |
156 aecpc->skewFile = fopen(filename, "wb"); | |
157 sprintf(filename, "aec_delay%d.dat", webrtc_aec_instance_count); | |
158 aecpc->delayFile = fopen(filename, "wb"); | |
159 webrtc_aec_instance_count++; | |
160 } | |
161 #endif | |
162 | |
163 return aecpc; | |
164 } | |
165 | |
166 void WebRtcAec_Free(void* aecInst) { | |
167 Aec* aecpc = (Aec*)aecInst; | |
168 | |
169 if (aecpc == NULL) { | |
170 return; | |
171 } | |
172 | |
173 WebRtc_FreeBuffer(aecpc->far_pre_buf); | |
174 | |
175 #ifdef WEBRTC_AEC_DEBUG_DUMP | |
176 fclose(aecpc->bufFile); | |
177 fclose(aecpc->skewFile); | |
178 fclose(aecpc->delayFile); | |
179 #endif | |
180 | |
181 WebRtcAec_FreeAec(aecpc->aec); | |
182 WebRtcAec_FreeResampler(aecpc->resampler); | |
183 free(aecpc); | |
184 } | |
185 | |
186 int32_t WebRtcAec_Init(void* aecInst, int32_t sampFreq, int32_t scSampFreq) { | |
187 Aec* aecpc = (Aec*)aecInst; | |
188 AecConfig aecConfig; | |
189 | |
190 if (sampFreq != 8000 && sampFreq != 16000 && sampFreq != 32000 && | |
191 sampFreq != 48000) { | |
192 return AEC_BAD_PARAMETER_ERROR; | |
193 } | |
194 aecpc->sampFreq = sampFreq; | |
195 | |
196 if (scSampFreq < 1 || scSampFreq > 96000) { | |
197 return AEC_BAD_PARAMETER_ERROR; | |
198 } | |
199 aecpc->scSampFreq = scSampFreq; | |
200 | |
201 // Initialize echo canceller core | |
202 if (WebRtcAec_InitAec(aecpc->aec, aecpc->sampFreq) == -1) { | |
203 return AEC_UNSPECIFIED_ERROR; | |
204 } | |
205 | |
206 if (WebRtcAec_InitResampler(aecpc->resampler, aecpc->scSampFreq) == -1) { | |
207 return AEC_UNSPECIFIED_ERROR; | |
208 } | |
209 | |
210 WebRtc_InitBuffer(aecpc->far_pre_buf); | |
211 WebRtc_MoveReadPtr(aecpc->far_pre_buf, -PART_LEN); // Start overlap. | |
212 | |
213 aecpc->initFlag = initCheck; // indicates that initialization has been done | |
214 | |
215 if (aecpc->sampFreq == 32000 || aecpc->sampFreq == 48000) { | |
216 aecpc->splitSampFreq = 16000; | |
217 } else { | |
218 aecpc->splitSampFreq = sampFreq; | |
219 } | |
220 | |
221 aecpc->delayCtr = 0; | |
222 aecpc->sampFactor = (aecpc->scSampFreq * 1.0f) / aecpc->splitSampFreq; | |
223 // Sampling frequency multiplier (SWB is processed as 160 frame size). | |
224 aecpc->rate_factor = aecpc->splitSampFreq / 8000; | |
225 | |
226 aecpc->sum = 0; | |
227 aecpc->counter = 0; | |
228 aecpc->checkBuffSize = 1; | |
229 aecpc->firstVal = 0; | |
230 | |
231 // We skip the startup_phase completely (setting to 0) if DA-AEC is enabled, | |
232 // but not extended_filter mode. | |
233 aecpc->startup_phase = WebRtcAec_extended_filter_enabled(aecpc->aec) || | |
234 !WebRtcAec_delay_agnostic_enabled(aecpc->aec); | |
235 aecpc->bufSizeStart = 0; | |
236 aecpc->checkBufSizeCtr = 0; | |
237 aecpc->msInSndCardBuf = 0; | |
238 aecpc->filtDelay = -1; // -1 indicates an initialized state. | |
239 aecpc->timeForDelayChange = 0; | |
240 aecpc->knownDelay = 0; | |
241 aecpc->lastDelayDiff = 0; | |
242 | |
243 aecpc->skewFrCtr = 0; | |
244 aecpc->resample = kAecFalse; | |
245 aecpc->highSkewCtr = 0; | |
246 aecpc->skew = 0; | |
247 | |
248 aecpc->farend_started = 0; | |
249 | |
250 // Default settings. | |
251 aecConfig.nlpMode = kAecNlpModerate; | |
252 aecConfig.skewMode = kAecFalse; | |
253 aecConfig.metricsMode = kAecFalse; | |
254 aecConfig.delay_logging = kAecFalse; | |
255 | |
256 if (WebRtcAec_set_config(aecpc, aecConfig) == -1) { | |
257 return AEC_UNSPECIFIED_ERROR; | |
258 } | |
259 | |
260 return 0; | |
261 } | |
262 | |
263 // Returns any error that is caused when buffering the | |
264 // far-end signal. | |
265 int32_t WebRtcAec_GetBufferFarendError(void* aecInst, | |
266 const float* farend, | |
267 size_t nrOfSamples) { | |
268 Aec* aecpc = (Aec*)aecInst; | |
269 | |
270 if (!farend) | |
271 return AEC_NULL_POINTER_ERROR; | |
272 | |
273 if (aecpc->initFlag != initCheck) | |
274 return AEC_UNINITIALIZED_ERROR; | |
275 | |
276 // number of samples == 160 for SWB input | |
277 if (nrOfSamples != 80 && nrOfSamples != 160) | |
278 return AEC_BAD_PARAMETER_ERROR; | |
279 | |
280 return 0; | |
281 } | |
282 | |
283 // only buffer L band for farend | |
284 int32_t WebRtcAec_BufferFarend(void* aecInst, | |
285 const float* farend, | |
286 size_t nrOfSamples) { | |
287 Aec* aecpc = (Aec*)aecInst; | |
288 size_t newNrOfSamples = nrOfSamples; | |
289 float new_farend[MAX_RESAMP_LEN]; | |
290 const float* farend_ptr = farend; | |
291 | |
292 // Get any error caused by buffering the farend signal. | |
293 int32_t error_code = | |
294 WebRtcAec_GetBufferFarendError(aecInst, farend, nrOfSamples); | |
295 | |
296 if (error_code != 0) | |
297 return error_code; | |
298 | |
299 if (aecpc->skewMode == kAecTrue && aecpc->resample == kAecTrue) { | |
300 // Resample and get a new number of samples | |
301 WebRtcAec_ResampleLinear(aecpc->resampler, farend, nrOfSamples, aecpc->skew, | |
302 new_farend, &newNrOfSamples); | |
303 farend_ptr = new_farend; | |
304 } | |
305 | |
306 aecpc->farend_started = 1; | |
307 WebRtcAec_SetSystemDelay( | |
308 aecpc->aec, WebRtcAec_system_delay(aecpc->aec) + (int)newNrOfSamples); | |
309 | |
310 // Write the time-domain data to |far_pre_buf|. | |
311 WebRtc_WriteBuffer(aecpc->far_pre_buf, farend_ptr, newNrOfSamples); | |
312 | |
313 // TODO(minyue): reduce to |PART_LEN| samples for each buffering, when | |
314 // WebRtcAec_BufferFarendPartition() is changed to take |PART_LEN| samples. | |
315 while (WebRtc_available_read(aecpc->far_pre_buf) >= PART_LEN2) { | |
316 // We have enough data to pass to the FFT, hence read PART_LEN2 samples. | |
317 { | |
318 float* ptmp = NULL; | |
319 float tmp[PART_LEN2]; | |
320 WebRtc_ReadBuffer(aecpc->far_pre_buf, (void**)&ptmp, tmp, PART_LEN2); | |
321 WebRtcAec_BufferFarendPartition(aecpc->aec, ptmp); | |
322 } | |
323 | |
324 // Rewind |far_pre_buf| PART_LEN samples for overlap before continuing. | |
325 WebRtc_MoveReadPtr(aecpc->far_pre_buf, -PART_LEN); | |
326 } | |
327 | |
328 return 0; | |
329 } | |
330 | |
331 int32_t WebRtcAec_Process(void* aecInst, | |
332 const float* const* nearend, | |
333 size_t num_bands, | |
334 float* const* out, | |
335 size_t nrOfSamples, | |
336 int16_t msInSndCardBuf, | |
337 int32_t skew) { | |
338 Aec* aecpc = (Aec*)aecInst; | |
339 int32_t retVal = 0; | |
340 | |
341 if (out == NULL) { | |
342 return AEC_NULL_POINTER_ERROR; | |
343 } | |
344 | |
345 if (aecpc->initFlag != initCheck) { | |
346 return AEC_UNINITIALIZED_ERROR; | |
347 } | |
348 | |
349 // number of samples == 160 for SWB input | |
350 if (nrOfSamples != 80 && nrOfSamples != 160) { | |
351 return AEC_BAD_PARAMETER_ERROR; | |
352 } | |
353 | |
354 if (msInSndCardBuf < 0) { | |
355 msInSndCardBuf = 0; | |
356 retVal = AEC_BAD_PARAMETER_WARNING; | |
357 } else if (msInSndCardBuf > kMaxTrustedDelayMs) { | |
358 // The clamping is now done in ProcessExtended/Normal(). | |
359 retVal = AEC_BAD_PARAMETER_WARNING; | |
360 } | |
361 | |
362 // This returns the value of aec->extended_filter_enabled. | |
363 if (WebRtcAec_extended_filter_enabled(aecpc->aec)) { | |
364 ProcessExtended(aecpc, nearend, num_bands, out, nrOfSamples, msInSndCardBuf, | |
365 skew); | |
366 } else { | |
367 retVal = ProcessNormal(aecpc, nearend, num_bands, out, nrOfSamples, | |
368 msInSndCardBuf, skew); | |
369 } | |
370 | |
371 #ifdef WEBRTC_AEC_DEBUG_DUMP | |
372 { | |
373 int16_t far_buf_size_ms = (int16_t)(WebRtcAec_system_delay(aecpc->aec) / | |
374 (sampMsNb * aecpc->rate_factor)); | |
375 (void)fwrite(&far_buf_size_ms, 2, 1, aecpc->bufFile); | |
376 (void)fwrite(&aecpc->knownDelay, sizeof(aecpc->knownDelay), 1, | |
377 aecpc->delayFile); | |
378 } | |
379 #endif | |
380 | |
381 return retVal; | |
382 } | |
383 | |
384 int WebRtcAec_set_config(void* handle, AecConfig config) { | |
385 Aec* self = (Aec*)handle; | |
386 if (self->initFlag != initCheck) { | |
387 return AEC_UNINITIALIZED_ERROR; | |
388 } | |
389 | |
390 if (config.skewMode != kAecFalse && config.skewMode != kAecTrue) { | |
391 return AEC_BAD_PARAMETER_ERROR; | |
392 } | |
393 self->skewMode = config.skewMode; | |
394 | |
395 if (config.nlpMode != kAecNlpConservative && | |
396 config.nlpMode != kAecNlpModerate && | |
397 config.nlpMode != kAecNlpAggressive) { | |
398 return AEC_BAD_PARAMETER_ERROR; | |
399 } | |
400 | |
401 if (config.metricsMode != kAecFalse && config.metricsMode != kAecTrue) { | |
402 return AEC_BAD_PARAMETER_ERROR; | |
403 } | |
404 | |
405 if (config.delay_logging != kAecFalse && config.delay_logging != kAecTrue) { | |
406 return AEC_BAD_PARAMETER_ERROR; | |
407 } | |
408 | |
409 WebRtcAec_SetConfigCore(self->aec, config.nlpMode, config.metricsMode, | |
410 config.delay_logging); | |
411 return 0; | |
412 } | |
413 | |
414 int WebRtcAec_get_echo_status(void* handle, int* status) { | |
415 Aec* self = (Aec*)handle; | |
416 if (status == NULL) { | |
417 return AEC_NULL_POINTER_ERROR; | |
418 } | |
419 if (self->initFlag != initCheck) { | |
420 return AEC_UNINITIALIZED_ERROR; | |
421 } | |
422 | |
423 *status = WebRtcAec_echo_state(self->aec); | |
424 | |
425 return 0; | |
426 } | |
427 | |
428 int WebRtcAec_GetMetrics(void* handle, AecMetrics* metrics) { | |
429 const float kUpWeight = 0.7f; | |
430 float dtmp; | |
431 int stmp; | |
432 Aec* self = (Aec*)handle; | |
433 Stats erl; | |
434 Stats erle; | |
435 Stats a_nlp; | |
436 | |
437 if (handle == NULL) { | |
438 return -1; | |
439 } | |
440 if (metrics == NULL) { | |
441 return AEC_NULL_POINTER_ERROR; | |
442 } | |
443 if (self->initFlag != initCheck) { | |
444 return AEC_UNINITIALIZED_ERROR; | |
445 } | |
446 | |
447 WebRtcAec_GetEchoStats(self->aec, &erl, &erle, &a_nlp); | |
448 | |
449 // ERL | |
450 metrics->erl.instant = (int)erl.instant; | |
451 | |
452 if ((erl.himean > kOffsetLevel) && (erl.average > kOffsetLevel)) { | |
453 // Use a mix between regular average and upper part average. | |
454 dtmp = kUpWeight * erl.himean + (1 - kUpWeight) * erl.average; | |
455 metrics->erl.average = (int)dtmp; | |
456 } else { | |
457 metrics->erl.average = kOffsetLevel; | |
458 } | |
459 | |
460 metrics->erl.max = (int)erl.max; | |
461 | |
462 if (erl.min < (kOffsetLevel * (-1))) { | |
463 metrics->erl.min = (int)erl.min; | |
464 } else { | |
465 metrics->erl.min = kOffsetLevel; | |
466 } | |
467 | |
468 // ERLE | |
469 metrics->erle.instant = (int)erle.instant; | |
470 | |
471 if ((erle.himean > kOffsetLevel) && (erle.average > kOffsetLevel)) { | |
472 // Use a mix between regular average and upper part average. | |
473 dtmp = kUpWeight * erle.himean + (1 - kUpWeight) * erle.average; | |
474 metrics->erle.average = (int)dtmp; | |
475 } else { | |
476 metrics->erle.average = kOffsetLevel; | |
477 } | |
478 | |
479 metrics->erle.max = (int)erle.max; | |
480 | |
481 if (erle.min < (kOffsetLevel * (-1))) { | |
482 metrics->erle.min = (int)erle.min; | |
483 } else { | |
484 metrics->erle.min = kOffsetLevel; | |
485 } | |
486 | |
487 // RERL | |
488 if ((metrics->erl.average > kOffsetLevel) && | |
489 (metrics->erle.average > kOffsetLevel)) { | |
490 stmp = metrics->erl.average + metrics->erle.average; | |
491 } else { | |
492 stmp = kOffsetLevel; | |
493 } | |
494 metrics->rerl.average = stmp; | |
495 | |
496 // No other statistics needed, but returned for completeness. | |
497 metrics->rerl.instant = stmp; | |
498 metrics->rerl.max = stmp; | |
499 metrics->rerl.min = stmp; | |
500 | |
501 // A_NLP | |
502 metrics->aNlp.instant = (int)a_nlp.instant; | |
503 | |
504 if ((a_nlp.himean > kOffsetLevel) && (a_nlp.average > kOffsetLevel)) { | |
505 // Use a mix between regular average and upper part average. | |
506 dtmp = kUpWeight * a_nlp.himean + (1 - kUpWeight) * a_nlp.average; | |
507 metrics->aNlp.average = (int)dtmp; | |
508 } else { | |
509 metrics->aNlp.average = kOffsetLevel; | |
510 } | |
511 | |
512 metrics->aNlp.max = (int)a_nlp.max; | |
513 | |
514 if (a_nlp.min < (kOffsetLevel * (-1))) { | |
515 metrics->aNlp.min = (int)a_nlp.min; | |
516 } else { | |
517 metrics->aNlp.min = kOffsetLevel; | |
518 } | |
519 | |
520 return 0; | |
521 } | |
522 | |
523 int WebRtcAec_GetDelayMetrics(void* handle, | |
524 int* median, | |
525 int* std, | |
526 float* fraction_poor_delays) { | |
527 Aec* self = (Aec*)handle; | |
528 if (median == NULL) { | |
529 return AEC_NULL_POINTER_ERROR; | |
530 } | |
531 if (std == NULL) { | |
532 return AEC_NULL_POINTER_ERROR; | |
533 } | |
534 if (self->initFlag != initCheck) { | |
535 return AEC_UNINITIALIZED_ERROR; | |
536 } | |
537 if (WebRtcAec_GetDelayMetricsCore(self->aec, median, std, | |
538 fraction_poor_delays) == -1) { | |
539 // Logging disabled. | |
540 return AEC_UNSUPPORTED_FUNCTION_ERROR; | |
541 } | |
542 | |
543 return 0; | |
544 } | |
545 | |
546 AecCore* WebRtcAec_aec_core(void* handle) { | |
547 if (!handle) { | |
548 return NULL; | |
549 } | |
550 return ((Aec*)handle)->aec; | |
551 } | |
552 | |
553 static int ProcessNormal(Aec* aecpc, | |
554 const float* const* nearend, | |
555 size_t num_bands, | |
556 float* const* out, | |
557 size_t nrOfSamples, | |
558 int16_t msInSndCardBuf, | |
559 int32_t skew) { | |
560 int retVal = 0; | |
561 size_t i; | |
562 size_t nBlocks10ms; | |
563 // Limit resampling to doubling/halving of signal | |
564 const float minSkewEst = -0.5f; | |
565 const float maxSkewEst = 1.0f; | |
566 | |
567 msInSndCardBuf = | |
568 msInSndCardBuf > kMaxTrustedDelayMs ? kMaxTrustedDelayMs : msInSndCardBuf; | |
569 // TODO(andrew): we need to investigate if this +10 is really wanted. | |
570 msInSndCardBuf += 10; | |
571 aecpc->msInSndCardBuf = msInSndCardBuf; | |
572 | |
573 if (aecpc->skewMode == kAecTrue) { | |
574 if (aecpc->skewFrCtr < 25) { | |
575 aecpc->skewFrCtr++; | |
576 } else { | |
577 retVal = WebRtcAec_GetSkew(aecpc->resampler, skew, &aecpc->skew); | |
578 if (retVal == -1) { | |
579 aecpc->skew = 0; | |
580 retVal = AEC_BAD_PARAMETER_WARNING; | |
581 } | |
582 | |
583 aecpc->skew /= aecpc->sampFactor * nrOfSamples; | |
584 | |
585 if (aecpc->skew < 1.0e-3 && aecpc->skew > -1.0e-3) { | |
586 aecpc->resample = kAecFalse; | |
587 } else { | |
588 aecpc->resample = kAecTrue; | |
589 } | |
590 | |
591 if (aecpc->skew < minSkewEst) { | |
592 aecpc->skew = minSkewEst; | |
593 } else if (aecpc->skew > maxSkewEst) { | |
594 aecpc->skew = maxSkewEst; | |
595 } | |
596 | |
597 #ifdef WEBRTC_AEC_DEBUG_DUMP | |
598 (void)fwrite(&aecpc->skew, sizeof(aecpc->skew), 1, aecpc->skewFile); | |
599 #endif | |
600 } | |
601 } | |
602 | |
603 nBlocks10ms = nrOfSamples / (FRAME_LEN * aecpc->rate_factor); | |
604 | |
605 if (aecpc->startup_phase) { | |
606 for (i = 0; i < num_bands; ++i) { | |
607 // Only needed if they don't already point to the same place. | |
608 if (nearend[i] != out[i]) { | |
609 memcpy(out[i], nearend[i], sizeof(nearend[i][0]) * nrOfSamples); | |
610 } | |
611 } | |
612 | |
613 // The AEC is in the start up mode | |
614 // AEC is disabled until the system delay is OK | |
615 | |
616 // Mechanism to ensure that the system delay is reasonably stable. | |
617 if (aecpc->checkBuffSize) { | |
618 aecpc->checkBufSizeCtr++; | |
619 // Before we fill up the far-end buffer we require the system delay | |
620 // to be stable (+/-8 ms) compared to the first value. This | |
621 // comparison is made during the following 6 consecutive 10 ms | |
622 // blocks. If it seems to be stable then we start to fill up the | |
623 // far-end buffer. | |
624 if (aecpc->counter == 0) { | |
625 aecpc->firstVal = aecpc->msInSndCardBuf; | |
626 aecpc->sum = 0; | |
627 } | |
628 | |
629 if (abs(aecpc->firstVal - aecpc->msInSndCardBuf) < | |
630 WEBRTC_SPL_MAX(0.2 * aecpc->msInSndCardBuf, sampMsNb)) { | |
631 aecpc->sum += aecpc->msInSndCardBuf; | |
632 aecpc->counter++; | |
633 } else { | |
634 aecpc->counter = 0; | |
635 } | |
636 | |
637 if (aecpc->counter * nBlocks10ms >= 6) { | |
638 // The far-end buffer size is determined in partitions of | |
639 // PART_LEN samples. Use 75% of the average value of the system | |
640 // delay as buffer size to start with. | |
641 aecpc->bufSizeStart = | |
642 WEBRTC_SPL_MIN((3 * aecpc->sum * aecpc->rate_factor * 8) / | |
643 (4 * aecpc->counter * PART_LEN), | |
644 kMaxBufSizeStart); | |
645 // Buffer size has now been determined. | |
646 aecpc->checkBuffSize = 0; | |
647 } | |
648 | |
649 if (aecpc->checkBufSizeCtr * nBlocks10ms > 50) { | |
650 // For really bad systems, don't disable the echo canceller for | |
651 // more than 0.5 sec. | |
652 aecpc->bufSizeStart = WEBRTC_SPL_MIN( | |
653 (aecpc->msInSndCardBuf * aecpc->rate_factor * 3) / 40, | |
654 kMaxBufSizeStart); | |
655 aecpc->checkBuffSize = 0; | |
656 } | |
657 } | |
658 | |
659 // If |checkBuffSize| changed in the if-statement above. | |
660 if (!aecpc->checkBuffSize) { | |
661 // The system delay is now reasonably stable (or has been unstable | |
662 // for too long). When the far-end buffer is filled with | |
663 // approximately the same amount of data as reported by the system | |
664 // we end the startup phase. | |
665 int overhead_elements = | |
666 WebRtcAec_system_delay(aecpc->aec) / PART_LEN - aecpc->bufSizeStart; | |
667 if (overhead_elements == 0) { | |
668 // Enable the AEC | |
669 aecpc->startup_phase = 0; | |
670 } else if (overhead_elements > 0) { | |
671 // TODO(bjornv): Do we need a check on how much we actually | |
672 // moved the read pointer? It should always be possible to move | |
673 // the pointer |overhead_elements| since we have only added data | |
674 // to the buffer and no delay compensation nor AEC processing | |
675 // has been done. | |
676 WebRtcAec_MoveFarReadPtr(aecpc->aec, overhead_elements); | |
677 | |
678 // Enable the AEC | |
679 aecpc->startup_phase = 0; | |
680 } | |
681 } | |
682 } else { | |
683 // AEC is enabled. | |
684 EstBufDelayNormal(aecpc); | |
685 | |
686 // Call the AEC. | |
687 // TODO(bjornv): Re-structure such that we don't have to pass | |
688 // |aecpc->knownDelay| as input. Change name to something like | |
689 // |system_buffer_diff|. | |
690 WebRtcAec_ProcessFrames(aecpc->aec, nearend, num_bands, nrOfSamples, | |
691 aecpc->knownDelay, out); | |
692 } | |
693 | |
694 return retVal; | |
695 } | |
696 | |
697 static void ProcessExtended(Aec* self, | |
698 const float* const* near, | |
699 size_t num_bands, | |
700 float* const* out, | |
701 size_t num_samples, | |
702 int16_t reported_delay_ms, | |
703 int32_t skew) { | |
704 size_t i; | |
705 const int delay_diff_offset = kDelayDiffOffsetSamples; | |
706 #if defined(WEBRTC_UNTRUSTED_DELAY) | |
707 reported_delay_ms = kFixedDelayMs; | |
708 #else | |
709 // This is the usual mode where we trust the reported system delay values. | |
710 // Due to the longer filter, we no longer add 10 ms to the reported delay | |
711 // to reduce chance of non-causality. Instead we apply a minimum here to avoid | |
712 // issues with the read pointer jumping around needlessly. | |
713 reported_delay_ms = reported_delay_ms < kMinTrustedDelayMs | |
714 ? kMinTrustedDelayMs | |
715 : reported_delay_ms; | |
716 // If the reported delay appears to be bogus, we attempt to recover by using | |
717 // the measured fixed delay values. We use >= here because higher layers | |
718 // may already clamp to this maximum value, and we would otherwise not | |
719 // detect it here. | |
720 reported_delay_ms = reported_delay_ms >= kMaxTrustedDelayMs | |
721 ? kFixedDelayMs | |
722 : reported_delay_ms; | |
723 #endif | |
724 self->msInSndCardBuf = reported_delay_ms; | |
725 | |
726 if (!self->farend_started) { | |
727 for (i = 0; i < num_bands; ++i) { | |
728 // Only needed if they don't already point to the same place. | |
729 if (near[i] != out[i]) { | |
730 memcpy(out[i], near[i], sizeof(near[i][0]) * num_samples); | |
731 } | |
732 } | |
733 return; | |
734 } | |
735 if (self->startup_phase) { | |
736 // In the extended mode, there isn't a startup "phase", just a special | |
737 // action on the first frame. In the trusted delay case, we'll take the | |
738 // current reported delay, unless it's less then our conservative | |
739 // measurement. | |
740 int startup_size_ms = | |
741 reported_delay_ms < kFixedDelayMs ? kFixedDelayMs : reported_delay_ms; | |
742 #if defined(WEBRTC_ANDROID) | |
743 int target_delay = startup_size_ms * self->rate_factor * 8; | |
744 #else | |
745 // To avoid putting the AEC in a non-causal state we're being slightly | |
746 // conservative and scale by 2. On Android we use a fixed delay and | |
747 // therefore there is no need to scale the target_delay. | |
748 int target_delay = startup_size_ms * self->rate_factor * 8 / 2; | |
749 #endif | |
750 int overhead_elements = | |
751 (WebRtcAec_system_delay(self->aec) - target_delay) / PART_LEN; | |
752 WebRtcAec_MoveFarReadPtr(self->aec, overhead_elements); | |
753 self->startup_phase = 0; | |
754 } | |
755 | |
756 EstBufDelayExtended(self); | |
757 | |
758 { | |
759 // |delay_diff_offset| gives us the option to manually rewind the delay on | |
760 // very low delay platforms which can't be expressed purely through | |
761 // |reported_delay_ms|. | |
762 const int adjusted_known_delay = | |
763 WEBRTC_SPL_MAX(0, self->knownDelay + delay_diff_offset); | |
764 | |
765 WebRtcAec_ProcessFrames(self->aec, near, num_bands, num_samples, | |
766 adjusted_known_delay, out); | |
767 } | |
768 } | |
769 | |
770 static void EstBufDelayNormal(Aec* aecpc) { | |
771 int nSampSndCard = aecpc->msInSndCardBuf * sampMsNb * aecpc->rate_factor; | |
772 int current_delay = nSampSndCard - WebRtcAec_system_delay(aecpc->aec); | |
773 int delay_difference = 0; | |
774 | |
775 // Before we proceed with the delay estimate filtering we: | |
776 // 1) Compensate for the frame that will be read. | |
777 // 2) Compensate for drift resampling. | |
778 // 3) Compensate for non-causality if needed, since the estimated delay can't | |
779 // be negative. | |
780 | |
781 // 1) Compensating for the frame(s) that will be read/processed. | |
782 current_delay += FRAME_LEN * aecpc->rate_factor; | |
783 | |
784 // 2) Account for resampling frame delay. | |
785 if (aecpc->skewMode == kAecTrue && aecpc->resample == kAecTrue) { | |
786 current_delay -= kResamplingDelay; | |
787 } | |
788 | |
789 // 3) Compensate for non-causality, if needed, by flushing one block. | |
790 if (current_delay < PART_LEN) { | |
791 current_delay += WebRtcAec_MoveFarReadPtr(aecpc->aec, 1) * PART_LEN; | |
792 } | |
793 | |
794 // We use -1 to signal an initialized state in the "extended" implementation; | |
795 // compensate for that. | |
796 aecpc->filtDelay = aecpc->filtDelay < 0 ? 0 : aecpc->filtDelay; | |
797 aecpc->filtDelay = | |
798 WEBRTC_SPL_MAX(0, (short)(0.8 * aecpc->filtDelay + 0.2 * current_delay)); | |
799 | |
800 delay_difference = aecpc->filtDelay - aecpc->knownDelay; | |
801 if (delay_difference > 224) { | |
802 if (aecpc->lastDelayDiff < 96) { | |
803 aecpc->timeForDelayChange = 0; | |
804 } else { | |
805 aecpc->timeForDelayChange++; | |
806 } | |
807 } else if (delay_difference < 96 && aecpc->knownDelay > 0) { | |
808 if (aecpc->lastDelayDiff > 224) { | |
809 aecpc->timeForDelayChange = 0; | |
810 } else { | |
811 aecpc->timeForDelayChange++; | |
812 } | |
813 } else { | |
814 aecpc->timeForDelayChange = 0; | |
815 } | |
816 aecpc->lastDelayDiff = delay_difference; | |
817 | |
818 if (aecpc->timeForDelayChange > 25) { | |
819 aecpc->knownDelay = WEBRTC_SPL_MAX((int)aecpc->filtDelay - 160, 0); | |
820 } | |
821 } | |
822 | |
823 static void EstBufDelayExtended(Aec* self) { | |
824 int reported_delay = self->msInSndCardBuf * sampMsNb * self->rate_factor; | |
825 int current_delay = reported_delay - WebRtcAec_system_delay(self->aec); | |
826 int delay_difference = 0; | |
827 | |
828 // Before we proceed with the delay estimate filtering we: | |
829 // 1) Compensate for the frame that will be read. | |
830 // 2) Compensate for drift resampling. | |
831 // 3) Compensate for non-causality if needed, since the estimated delay can't | |
832 // be negative. | |
833 | |
834 // 1) Compensating for the frame(s) that will be read/processed. | |
835 current_delay += FRAME_LEN * self->rate_factor; | |
836 | |
837 // 2) Account for resampling frame delay. | |
838 if (self->skewMode == kAecTrue && self->resample == kAecTrue) { | |
839 current_delay -= kResamplingDelay; | |
840 } | |
841 | |
842 // 3) Compensate for non-causality, if needed, by flushing two blocks. | |
843 if (current_delay < PART_LEN) { | |
844 current_delay += WebRtcAec_MoveFarReadPtr(self->aec, 2) * PART_LEN; | |
845 } | |
846 | |
847 if (self->filtDelay == -1) { | |
848 self->filtDelay = WEBRTC_SPL_MAX(0, 0.5 * current_delay); | |
849 } else { | |
850 self->filtDelay = WEBRTC_SPL_MAX( | |
851 0, (short)(0.95 * self->filtDelay + 0.05 * current_delay)); | |
852 } | |
853 | |
854 delay_difference = self->filtDelay - self->knownDelay; | |
855 if (delay_difference > 384) { | |
856 if (self->lastDelayDiff < 128) { | |
857 self->timeForDelayChange = 0; | |
858 } else { | |
859 self->timeForDelayChange++; | |
860 } | |
861 } else if (delay_difference < 128 && self->knownDelay > 0) { | |
862 if (self->lastDelayDiff > 384) { | |
863 self->timeForDelayChange = 0; | |
864 } else { | |
865 self->timeForDelayChange++; | |
866 } | |
867 } else { | |
868 self->timeForDelayChange = 0; | |
869 } | |
870 self->lastDelayDiff = delay_difference; | |
871 | |
872 if (self->timeForDelayChange > 25) { | |
873 self->knownDelay = WEBRTC_SPL_MAX((int)self->filtDelay - 256, 0); | |
874 } | |
875 } | |
OLD | NEW |