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_library.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
WebRtcAec_Create()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
WebRtcAec_Free(void * aecInst)166 void WebRtcAec_Free(void* aecInst) {
167 Aec* aecpc = 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
WebRtcAec_Init(void * aecInst,int32_t sampFreq,int32_t scSampFreq)186 int32_t WebRtcAec_Init(void* aecInst, int32_t sampFreq, int32_t scSampFreq) {
187 Aec* aecpc = aecInst;
188 AecConfig aecConfig;
189
190 if (sampFreq != 8000 &&
191 sampFreq != 16000 &&
192 sampFreq != 32000 &&
193 sampFreq != 48000) {
194 return AEC_BAD_PARAMETER_ERROR;
195 }
196 aecpc->sampFreq = sampFreq;
197
198 if (scSampFreq < 1 || scSampFreq > 96000) {
199 return AEC_BAD_PARAMETER_ERROR;
200 }
201 aecpc->scSampFreq = scSampFreq;
202
203 // Initialize echo canceller core
204 if (WebRtcAec_InitAec(aecpc->aec, aecpc->sampFreq) == -1) {
205 return AEC_UNSPECIFIED_ERROR;
206 }
207
208 if (WebRtcAec_InitResampler(aecpc->resampler, aecpc->scSampFreq) == -1) {
209 return AEC_UNSPECIFIED_ERROR;
210 }
211
212 WebRtc_InitBuffer(aecpc->far_pre_buf);
213 WebRtc_MoveReadPtr(aecpc->far_pre_buf, -PART_LEN); // Start overlap.
214
215 aecpc->initFlag = initCheck; // indicates that initialization has been done
216
217 if (aecpc->sampFreq == 32000 || aecpc->sampFreq == 48000) {
218 aecpc->splitSampFreq = 16000;
219 } else {
220 aecpc->splitSampFreq = sampFreq;
221 }
222
223 aecpc->delayCtr = 0;
224 aecpc->sampFactor = (aecpc->scSampFreq * 1.0f) / aecpc->splitSampFreq;
225 // Sampling frequency multiplier (SWB is processed as 160 frame size).
226 aecpc->rate_factor = aecpc->splitSampFreq / 8000;
227
228 aecpc->sum = 0;
229 aecpc->counter = 0;
230 aecpc->checkBuffSize = 1;
231 aecpc->firstVal = 0;
232
233 // We skip the startup_phase completely (setting to 0) if DA-AEC is enabled,
234 // but not extended_filter mode.
235 aecpc->startup_phase = WebRtcAec_extended_filter_enabled(aecpc->aec) ||
236 !WebRtcAec_delay_agnostic_enabled(aecpc->aec);
237 aecpc->bufSizeStart = 0;
238 aecpc->checkBufSizeCtr = 0;
239 aecpc->msInSndCardBuf = 0;
240 aecpc->filtDelay = -1; // -1 indicates an initialized state.
241 aecpc->timeForDelayChange = 0;
242 aecpc->knownDelay = 0;
243 aecpc->lastDelayDiff = 0;
244
245 aecpc->skewFrCtr = 0;
246 aecpc->resample = kAecFalse;
247 aecpc->highSkewCtr = 0;
248 aecpc->skew = 0;
249
250 aecpc->farend_started = 0;
251
252 // Default settings.
253 aecConfig.nlpMode = kAecNlpModerate;
254 aecConfig.skewMode = kAecFalse;
255 aecConfig.metricsMode = kAecFalse;
256 aecConfig.delay_logging = kAecFalse;
257
258 if (WebRtcAec_set_config(aecpc, aecConfig) == -1) {
259 return AEC_UNSPECIFIED_ERROR;
260 }
261
262 return 0;
263 }
264
265 // Returns any error that is caused when buffering the
266 // far-end signal.
WebRtcAec_GetBufferFarendError(void * aecInst,const float * farend,size_t nrOfSamples)267 int32_t WebRtcAec_GetBufferFarendError(void* aecInst,
268 const float* farend,
269 size_t nrOfSamples) {
270 Aec* aecpc = aecInst;
271
272 if (!farend)
273 return AEC_NULL_POINTER_ERROR;
274
275 if (aecpc->initFlag != initCheck)
276 return AEC_UNINITIALIZED_ERROR;
277
278 // number of samples == 160 for SWB input
279 if (nrOfSamples != 80 && nrOfSamples != 160)
280 return AEC_BAD_PARAMETER_ERROR;
281
282 return 0;
283 }
284
285 // only buffer L band for farend
WebRtcAec_BufferFarend(void * aecInst,const float * farend,size_t nrOfSamples)286 int32_t WebRtcAec_BufferFarend(void* aecInst,
287 const float* farend,
288 size_t nrOfSamples) {
289 Aec* aecpc = aecInst;
290 size_t newNrOfSamples = nrOfSamples;
291 float new_farend[MAX_RESAMP_LEN];
292 const float* farend_ptr = farend;
293
294 // Get any error caused by buffering the farend signal.
295 int32_t error_code = WebRtcAec_GetBufferFarendError(aecInst, farend,
296 nrOfSamples);
297
298 if (error_code != 0)
299 return error_code;
300
301
302 if (aecpc->skewMode == kAecTrue && aecpc->resample == kAecTrue) {
303 // Resample and get a new number of samples
304 WebRtcAec_ResampleLinear(aecpc->resampler,
305 farend,
306 nrOfSamples,
307 aecpc->skew,
308 new_farend,
309 &newNrOfSamples);
310 farend_ptr = new_farend;
311 }
312
313 aecpc->farend_started = 1;
314 WebRtcAec_SetSystemDelay(
315 aecpc->aec, WebRtcAec_system_delay(aecpc->aec) + (int)newNrOfSamples);
316
317 // Write the time-domain data to |far_pre_buf|.
318 WebRtc_WriteBuffer(aecpc->far_pre_buf, farend_ptr, newNrOfSamples);
319
320 // TODO(minyue): reduce to |PART_LEN| samples for each buffering, when
321 // WebRtcAec_BufferFarendPartition() is changed to take |PART_LEN| samples.
322 while (WebRtc_available_read(aecpc->far_pre_buf) >= PART_LEN2) {
323 // We have enough data to pass to the FFT, hence read PART_LEN2 samples.
324 {
325 float* ptmp = NULL;
326 float tmp[PART_LEN2];
327 WebRtc_ReadBuffer(aecpc->far_pre_buf, (void**)&ptmp, tmp, PART_LEN2);
328 WebRtcAec_BufferFarendPartition(aecpc->aec, ptmp);
329 }
330
331 // Rewind |far_pre_buf| PART_LEN samples for overlap before continuing.
332 WebRtc_MoveReadPtr(aecpc->far_pre_buf, -PART_LEN);
333 }
334
335 return 0;
336 }
337
WebRtcAec_Process(void * aecInst,const float * const * nearend,size_t num_bands,float * const * out,size_t nrOfSamples,int16_t msInSndCardBuf,int32_t skew)338 int32_t WebRtcAec_Process(void* aecInst,
339 const float* const* nearend,
340 size_t num_bands,
341 float* const* out,
342 size_t nrOfSamples,
343 int16_t msInSndCardBuf,
344 int32_t skew) {
345 Aec* aecpc = aecInst;
346 int32_t retVal = 0;
347
348 if (out == NULL) {
349 return AEC_NULL_POINTER_ERROR;
350 }
351
352 if (aecpc->initFlag != initCheck) {
353 return AEC_UNINITIALIZED_ERROR;
354 }
355
356 // number of samples == 160 for SWB input
357 if (nrOfSamples != 80 && nrOfSamples != 160) {
358 return AEC_BAD_PARAMETER_ERROR;
359 }
360
361 if (msInSndCardBuf < 0) {
362 msInSndCardBuf = 0;
363 retVal = AEC_BAD_PARAMETER_WARNING;
364 } else if (msInSndCardBuf > kMaxTrustedDelayMs) {
365 // The clamping is now done in ProcessExtended/Normal().
366 retVal = AEC_BAD_PARAMETER_WARNING;
367 }
368
369 // This returns the value of aec->extended_filter_enabled.
370 if (WebRtcAec_extended_filter_enabled(aecpc->aec)) {
371 ProcessExtended(aecpc,
372 nearend,
373 num_bands,
374 out,
375 nrOfSamples,
376 msInSndCardBuf,
377 skew);
378 } else {
379 retVal = ProcessNormal(aecpc,
380 nearend,
381 num_bands,
382 out,
383 nrOfSamples,
384 msInSndCardBuf,
385 skew);
386 }
387
388 #ifdef WEBRTC_AEC_DEBUG_DUMP
389 {
390 int16_t far_buf_size_ms = (int16_t)(WebRtcAec_system_delay(aecpc->aec) /
391 (sampMsNb * aecpc->rate_factor));
392 (void)fwrite(&far_buf_size_ms, 2, 1, aecpc->bufFile);
393 (void)fwrite(
394 &aecpc->knownDelay, sizeof(aecpc->knownDelay), 1, aecpc->delayFile);
395 }
396 #endif
397
398 return retVal;
399 }
400
WebRtcAec_set_config(void * handle,AecConfig config)401 int WebRtcAec_set_config(void* handle, AecConfig config) {
402 Aec* self = (Aec*)handle;
403 if (self->initFlag != initCheck) {
404 return AEC_UNINITIALIZED_ERROR;
405 }
406
407 if (config.skewMode != kAecFalse && config.skewMode != kAecTrue) {
408 return AEC_BAD_PARAMETER_ERROR;
409 }
410 self->skewMode = config.skewMode;
411
412 if (config.nlpMode != kAecNlpConservative &&
413 config.nlpMode != kAecNlpModerate &&
414 config.nlpMode != kAecNlpAggressive) {
415 return AEC_BAD_PARAMETER_ERROR;
416 }
417
418 if (config.metricsMode != kAecFalse && config.metricsMode != kAecTrue) {
419 return AEC_BAD_PARAMETER_ERROR;
420 }
421
422 if (config.delay_logging != kAecFalse && config.delay_logging != kAecTrue) {
423 return AEC_BAD_PARAMETER_ERROR;
424 }
425
426 WebRtcAec_SetConfigCore(
427 self->aec, config.nlpMode, config.metricsMode, config.delay_logging);
428 return 0;
429 }
430
WebRtcAec_get_echo_status(void * handle,int * status)431 int WebRtcAec_get_echo_status(void* handle, int* status) {
432 Aec* self = (Aec*)handle;
433 if (status == NULL) {
434 return AEC_NULL_POINTER_ERROR;
435 }
436 if (self->initFlag != initCheck) {
437 return AEC_UNINITIALIZED_ERROR;
438 }
439
440 *status = WebRtcAec_echo_state(self->aec);
441
442 return 0;
443 }
444
WebRtcAec_GetMetrics(void * handle,AecMetrics * metrics)445 int WebRtcAec_GetMetrics(void* handle, AecMetrics* metrics) {
446 const float kUpWeight = 0.7f;
447 float dtmp;
448 int stmp;
449 Aec* self = (Aec*)handle;
450 Stats erl;
451 Stats erle;
452 Stats a_nlp;
453
454 if (handle == NULL) {
455 return -1;
456 }
457 if (metrics == NULL) {
458 return AEC_NULL_POINTER_ERROR;
459 }
460 if (self->initFlag != initCheck) {
461 return AEC_UNINITIALIZED_ERROR;
462 }
463
464 WebRtcAec_GetEchoStats(self->aec, &erl, &erle, &a_nlp);
465
466 // ERL
467 metrics->erl.instant = (int)erl.instant;
468
469 if ((erl.himean > kOffsetLevel) && (erl.average > kOffsetLevel)) {
470 // Use a mix between regular average and upper part average.
471 dtmp = kUpWeight * erl.himean + (1 - kUpWeight) * erl.average;
472 metrics->erl.average = (int)dtmp;
473 } else {
474 metrics->erl.average = kOffsetLevel;
475 }
476
477 metrics->erl.max = (int)erl.max;
478
479 if (erl.min < (kOffsetLevel * (-1))) {
480 metrics->erl.min = (int)erl.min;
481 } else {
482 metrics->erl.min = kOffsetLevel;
483 }
484
485 // ERLE
486 metrics->erle.instant = (int)erle.instant;
487
488 if ((erle.himean > kOffsetLevel) && (erle.average > kOffsetLevel)) {
489 // Use a mix between regular average and upper part average.
490 dtmp = kUpWeight * erle.himean + (1 - kUpWeight) * erle.average;
491 metrics->erle.average = (int)dtmp;
492 } else {
493 metrics->erle.average = kOffsetLevel;
494 }
495
496 metrics->erle.max = (int)erle.max;
497
498 if (erle.min < (kOffsetLevel * (-1))) {
499 metrics->erle.min = (int)erle.min;
500 } else {
501 metrics->erle.min = kOffsetLevel;
502 }
503
504 // RERL
505 if ((metrics->erl.average > kOffsetLevel) &&
506 (metrics->erle.average > kOffsetLevel)) {
507 stmp = metrics->erl.average + metrics->erle.average;
508 } else {
509 stmp = kOffsetLevel;
510 }
511 metrics->rerl.average = stmp;
512
513 // No other statistics needed, but returned for completeness.
514 metrics->rerl.instant = stmp;
515 metrics->rerl.max = stmp;
516 metrics->rerl.min = stmp;
517
518 // A_NLP
519 metrics->aNlp.instant = (int)a_nlp.instant;
520
521 if ((a_nlp.himean > kOffsetLevel) && (a_nlp.average > kOffsetLevel)) {
522 // Use a mix between regular average and upper part average.
523 dtmp = kUpWeight * a_nlp.himean + (1 - kUpWeight) * a_nlp.average;
524 metrics->aNlp.average = (int)dtmp;
525 } else {
526 metrics->aNlp.average = kOffsetLevel;
527 }
528
529 metrics->aNlp.max = (int)a_nlp.max;
530
531 if (a_nlp.min < (kOffsetLevel * (-1))) {
532 metrics->aNlp.min = (int)a_nlp.min;
533 } else {
534 metrics->aNlp.min = kOffsetLevel;
535 }
536
537 return 0;
538 }
539
WebRtcAec_GetDelayMetrics(void * handle,int * median,int * std,float * fraction_poor_delays)540 int WebRtcAec_GetDelayMetrics(void* handle,
541 int* median,
542 int* std,
543 float* fraction_poor_delays) {
544 Aec* self = handle;
545 if (median == NULL) {
546 return AEC_NULL_POINTER_ERROR;
547 }
548 if (std == NULL) {
549 return AEC_NULL_POINTER_ERROR;
550 }
551 if (self->initFlag != initCheck) {
552 return AEC_UNINITIALIZED_ERROR;
553 }
554 if (WebRtcAec_GetDelayMetricsCore(self->aec, median, std,
555 fraction_poor_delays) ==
556 -1) {
557 // Logging disabled.
558 return AEC_UNSUPPORTED_FUNCTION_ERROR;
559 }
560
561 return 0;
562 }
563
564
WebRtcAec_aec_core(void * handle)565 AecCore* WebRtcAec_aec_core(void* handle) {
566 if (!handle) {
567 return NULL;
568 }
569 return ((Aec*)handle)->aec;
570 }
571
ProcessNormal(Aec * aecpc,const float * const * nearend,size_t num_bands,float * const * out,size_t nrOfSamples,int16_t msInSndCardBuf,int32_t skew)572 static int ProcessNormal(Aec* aecpc,
573 const float* const* nearend,
574 size_t num_bands,
575 float* const* out,
576 size_t nrOfSamples,
577 int16_t msInSndCardBuf,
578 int32_t skew) {
579 int retVal = 0;
580 size_t i;
581 size_t nBlocks10ms;
582 // Limit resampling to doubling/halving of signal
583 const float minSkewEst = -0.5f;
584 const float maxSkewEst = 1.0f;
585
586 msInSndCardBuf =
587 msInSndCardBuf > kMaxTrustedDelayMs ? kMaxTrustedDelayMs : msInSndCardBuf;
588 // TODO(andrew): we need to investigate if this +10 is really wanted.
589 msInSndCardBuf += 10;
590 aecpc->msInSndCardBuf = msInSndCardBuf;
591
592 if (aecpc->skewMode == kAecTrue) {
593 if (aecpc->skewFrCtr < 25) {
594 aecpc->skewFrCtr++;
595 } else {
596 retVal = WebRtcAec_GetSkew(aecpc->resampler, skew, &aecpc->skew);
597 if (retVal == -1) {
598 aecpc->skew = 0;
599 retVal = AEC_BAD_PARAMETER_WARNING;
600 }
601
602 aecpc->skew /= aecpc->sampFactor * nrOfSamples;
603
604 if (aecpc->skew < 1.0e-3 && aecpc->skew > -1.0e-3) {
605 aecpc->resample = kAecFalse;
606 } else {
607 aecpc->resample = kAecTrue;
608 }
609
610 if (aecpc->skew < minSkewEst) {
611 aecpc->skew = minSkewEst;
612 } else if (aecpc->skew > maxSkewEst) {
613 aecpc->skew = maxSkewEst;
614 }
615
616 #ifdef WEBRTC_AEC_DEBUG_DUMP
617 (void)fwrite(&aecpc->skew, sizeof(aecpc->skew), 1, aecpc->skewFile);
618 #endif
619 }
620 }
621
622 nBlocks10ms = nrOfSamples / (FRAME_LEN * aecpc->rate_factor);
623
624 if (aecpc->startup_phase) {
625 for (i = 0; i < num_bands; ++i) {
626 // Only needed if they don't already point to the same place.
627 if (nearend[i] != out[i]) {
628 memcpy(out[i], nearend[i], sizeof(nearend[i][0]) * nrOfSamples);
629 }
630 }
631
632 // The AEC is in the start up mode
633 // AEC is disabled until the system delay is OK
634
635 // Mechanism to ensure that the system delay is reasonably stable.
636 if (aecpc->checkBuffSize) {
637 aecpc->checkBufSizeCtr++;
638 // Before we fill up the far-end buffer we require the system delay
639 // to be stable (+/-8 ms) compared to the first value. This
640 // comparison is made during the following 6 consecutive 10 ms
641 // blocks. If it seems to be stable then we start to fill up the
642 // far-end buffer.
643 if (aecpc->counter == 0) {
644 aecpc->firstVal = aecpc->msInSndCardBuf;
645 aecpc->sum = 0;
646 }
647
648 if (abs(aecpc->firstVal - aecpc->msInSndCardBuf) <
649 WEBRTC_SPL_MAX(0.2 * aecpc->msInSndCardBuf, sampMsNb)) {
650 aecpc->sum += aecpc->msInSndCardBuf;
651 aecpc->counter++;
652 } else {
653 aecpc->counter = 0;
654 }
655
656 if (aecpc->counter * nBlocks10ms >= 6) {
657 // The far-end buffer size is determined in partitions of
658 // PART_LEN samples. Use 75% of the average value of the system
659 // delay as buffer size to start with.
660 aecpc->bufSizeStart =
661 WEBRTC_SPL_MIN((3 * aecpc->sum * aecpc->rate_factor * 8) /
662 (4 * aecpc->counter * PART_LEN),
663 kMaxBufSizeStart);
664 // Buffer size has now been determined.
665 aecpc->checkBuffSize = 0;
666 }
667
668 if (aecpc->checkBufSizeCtr * nBlocks10ms > 50) {
669 // For really bad systems, don't disable the echo canceller for
670 // more than 0.5 sec.
671 aecpc->bufSizeStart = WEBRTC_SPL_MIN(
672 (aecpc->msInSndCardBuf * aecpc->rate_factor * 3) / 40,
673 kMaxBufSizeStart);
674 aecpc->checkBuffSize = 0;
675 }
676 }
677
678 // If |checkBuffSize| changed in the if-statement above.
679 if (!aecpc->checkBuffSize) {
680 // The system delay is now reasonably stable (or has been unstable
681 // for too long). When the far-end buffer is filled with
682 // approximately the same amount of data as reported by the system
683 // we end the startup phase.
684 int overhead_elements =
685 WebRtcAec_system_delay(aecpc->aec) / PART_LEN - aecpc->bufSizeStart;
686 if (overhead_elements == 0) {
687 // Enable the AEC
688 aecpc->startup_phase = 0;
689 } else if (overhead_elements > 0) {
690 // TODO(bjornv): Do we need a check on how much we actually
691 // moved the read pointer? It should always be possible to move
692 // the pointer |overhead_elements| since we have only added data
693 // to the buffer and no delay compensation nor AEC processing
694 // has been done.
695 WebRtcAec_MoveFarReadPtr(aecpc->aec, overhead_elements);
696
697 // Enable the AEC
698 aecpc->startup_phase = 0;
699 }
700 }
701 } else {
702 // AEC is enabled.
703 EstBufDelayNormal(aecpc);
704
705 // Call the AEC.
706 // TODO(bjornv): Re-structure such that we don't have to pass
707 // |aecpc->knownDelay| as input. Change name to something like
708 // |system_buffer_diff|.
709 WebRtcAec_ProcessFrames(aecpc->aec,
710 nearend,
711 num_bands,
712 nrOfSamples,
713 aecpc->knownDelay,
714 out);
715 }
716
717 return retVal;
718 }
719
ProcessExtended(Aec * self,const float * const * near,size_t num_bands,float * const * out,size_t num_samples,int16_t reported_delay_ms,int32_t skew)720 static void ProcessExtended(Aec* self,
721 const float* const* near,
722 size_t num_bands,
723 float* const* out,
724 size_t num_samples,
725 int16_t reported_delay_ms,
726 int32_t skew) {
727 size_t i;
728 const int delay_diff_offset = kDelayDiffOffsetSamples;
729 #if defined(WEBRTC_UNTRUSTED_DELAY)
730 reported_delay_ms = kFixedDelayMs;
731 #else
732 // This is the usual mode where we trust the reported system delay values.
733 // Due to the longer filter, we no longer add 10 ms to the reported delay
734 // to reduce chance of non-causality. Instead we apply a minimum here to avoid
735 // issues with the read pointer jumping around needlessly.
736 reported_delay_ms = reported_delay_ms < kMinTrustedDelayMs
737 ? kMinTrustedDelayMs
738 : reported_delay_ms;
739 // If the reported delay appears to be bogus, we attempt to recover by using
740 // the measured fixed delay values. We use >= here because higher layers
741 // may already clamp to this maximum value, and we would otherwise not
742 // detect it here.
743 reported_delay_ms = reported_delay_ms >= kMaxTrustedDelayMs
744 ? kFixedDelayMs
745 : reported_delay_ms;
746 #endif
747 self->msInSndCardBuf = reported_delay_ms;
748
749 if (!self->farend_started) {
750 for (i = 0; i < num_bands; ++i) {
751 // Only needed if they don't already point to the same place.
752 if (near[i] != out[i]) {
753 memcpy(out[i], near[i], sizeof(near[i][0]) * num_samples);
754 }
755 }
756 return;
757 }
758 if (self->startup_phase) {
759 // In the extended mode, there isn't a startup "phase", just a special
760 // action on the first frame. In the trusted delay case, we'll take the
761 // current reported delay, unless it's less then our conservative
762 // measurement.
763 int startup_size_ms =
764 reported_delay_ms < kFixedDelayMs ? kFixedDelayMs : reported_delay_ms;
765 #if defined(WEBRTC_ANDROID)
766 int target_delay = startup_size_ms * self->rate_factor * 8;
767 #else
768 // To avoid putting the AEC in a non-causal state we're being slightly
769 // conservative and scale by 2. On Android we use a fixed delay and
770 // therefore there is no need to scale the target_delay.
771 int target_delay = startup_size_ms * self->rate_factor * 8 / 2;
772 #endif
773 int overhead_elements =
774 (WebRtcAec_system_delay(self->aec) - target_delay) / PART_LEN;
775 WebRtcAec_MoveFarReadPtr(self->aec, overhead_elements);
776 self->startup_phase = 0;
777 }
778
779 EstBufDelayExtended(self);
780
781 {
782 // |delay_diff_offset| gives us the option to manually rewind the delay on
783 // very low delay platforms which can't be expressed purely through
784 // |reported_delay_ms|.
785 const int adjusted_known_delay =
786 WEBRTC_SPL_MAX(0, self->knownDelay + delay_diff_offset);
787
788 WebRtcAec_ProcessFrames(self->aec,
789 near,
790 num_bands,
791 num_samples,
792 adjusted_known_delay,
793 out);
794 }
795 }
796
EstBufDelayNormal(Aec * aecpc)797 static void EstBufDelayNormal(Aec* aecpc) {
798 int nSampSndCard = aecpc->msInSndCardBuf * sampMsNb * aecpc->rate_factor;
799 int current_delay = nSampSndCard - WebRtcAec_system_delay(aecpc->aec);
800 int delay_difference = 0;
801
802 // Before we proceed with the delay estimate filtering we:
803 // 1) Compensate for the frame that will be read.
804 // 2) Compensate for drift resampling.
805 // 3) Compensate for non-causality if needed, since the estimated delay can't
806 // be negative.
807
808 // 1) Compensating for the frame(s) that will be read/processed.
809 current_delay += FRAME_LEN * aecpc->rate_factor;
810
811 // 2) Account for resampling frame delay.
812 if (aecpc->skewMode == kAecTrue && aecpc->resample == kAecTrue) {
813 current_delay -= kResamplingDelay;
814 }
815
816 // 3) Compensate for non-causality, if needed, by flushing one block.
817 if (current_delay < PART_LEN) {
818 current_delay += WebRtcAec_MoveFarReadPtr(aecpc->aec, 1) * PART_LEN;
819 }
820
821 // We use -1 to signal an initialized state in the "extended" implementation;
822 // compensate for that.
823 aecpc->filtDelay = aecpc->filtDelay < 0 ? 0 : aecpc->filtDelay;
824 aecpc->filtDelay =
825 WEBRTC_SPL_MAX(0, (short)(0.8 * aecpc->filtDelay + 0.2 * current_delay));
826
827 delay_difference = aecpc->filtDelay - aecpc->knownDelay;
828 if (delay_difference > 224) {
829 if (aecpc->lastDelayDiff < 96) {
830 aecpc->timeForDelayChange = 0;
831 } else {
832 aecpc->timeForDelayChange++;
833 }
834 } else if (delay_difference < 96 && aecpc->knownDelay > 0) {
835 if (aecpc->lastDelayDiff > 224) {
836 aecpc->timeForDelayChange = 0;
837 } else {
838 aecpc->timeForDelayChange++;
839 }
840 } else {
841 aecpc->timeForDelayChange = 0;
842 }
843 aecpc->lastDelayDiff = delay_difference;
844
845 if (aecpc->timeForDelayChange > 25) {
846 aecpc->knownDelay = WEBRTC_SPL_MAX((int)aecpc->filtDelay - 160, 0);
847 }
848 }
849
EstBufDelayExtended(Aec * self)850 static void EstBufDelayExtended(Aec* self) {
851 int reported_delay = self->msInSndCardBuf * sampMsNb * self->rate_factor;
852 int current_delay = reported_delay - WebRtcAec_system_delay(self->aec);
853 int delay_difference = 0;
854
855 // Before we proceed with the delay estimate filtering we:
856 // 1) Compensate for the frame that will be read.
857 // 2) Compensate for drift resampling.
858 // 3) Compensate for non-causality if needed, since the estimated delay can't
859 // be negative.
860
861 // 1) Compensating for the frame(s) that will be read/processed.
862 current_delay += FRAME_LEN * self->rate_factor;
863
864 // 2) Account for resampling frame delay.
865 if (self->skewMode == kAecTrue && self->resample == kAecTrue) {
866 current_delay -= kResamplingDelay;
867 }
868
869 // 3) Compensate for non-causality, if needed, by flushing two blocks.
870 if (current_delay < PART_LEN) {
871 current_delay += WebRtcAec_MoveFarReadPtr(self->aec, 2) * PART_LEN;
872 }
873
874 if (self->filtDelay == -1) {
875 self->filtDelay = WEBRTC_SPL_MAX(0, 0.5 * current_delay);
876 } else {
877 self->filtDelay = WEBRTC_SPL_MAX(
878 0, (short)(0.95 * self->filtDelay + 0.05 * current_delay));
879 }
880
881 delay_difference = self->filtDelay - self->knownDelay;
882 if (delay_difference > 384) {
883 if (self->lastDelayDiff < 128) {
884 self->timeForDelayChange = 0;
885 } else {
886 self->timeForDelayChange++;
887 }
888 } else if (delay_difference < 128 && self->knownDelay > 0) {
889 if (self->lastDelayDiff > 384) {
890 self->timeForDelayChange = 0;
891 } else {
892 self->timeForDelayChange++;
893 }
894 } else {
895 self->timeForDelayChange = 0;
896 }
897 self->lastDelayDiff = delay_difference;
898
899 if (self->timeForDelayChange > 25) {
900 self->knownDelay = WEBRTC_SPL_MAX((int)self->filtDelay - 256, 0);
901 }
902 }
903