1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * SpanDSP - a series of DSP components for telephony
4  *
5  * echo.c - A line echo canceller.  This code is being developed
6  *          against and partially complies with G168.
7  *
8  * Written by Steve Underwood <steveu@coppice.org>
9  *         and David Rowe <david_at_rowetel_dot_com>
10  *
11  * Copyright (C) 2001, 2003 Steve Underwood, 2007 David Rowe
12  *
13  * Based on a bit from here, a bit from there, eye of toad, ear of
14  * bat, 15 years of failed attempts by David and a few fried brain
15  * cells.
16  *
17  * All rights reserved.
18  */
19 
20 /*! \file */
21 
22 /* Implementation Notes
23    David Rowe
24    April 2007
25 
26    This code started life as Steve's NLMS algorithm with a tap
27    rotation algorithm to handle divergence during double talk.  I
28    added a Geigel Double Talk Detector (DTD) [2] and performed some
29    G168 tests.  However I had trouble meeting the G168 requirements,
30    especially for double talk - there were always cases where my DTD
31    failed, for example where near end speech was under the 6dB
32    threshold required for declaring double talk.
33 
34    So I tried a two path algorithm [1], which has so far given better
35    results.  The original tap rotation/Geigel algorithm is available
36    in SVN http://svn.rowetel.com/software/oslec/tags/before_16bit.
37    It's probably possible to make it work if some one wants to put some
38    serious work into it.
39 
40    At present no special treatment is provided for tones, which
41    generally cause NLMS algorithms to diverge.  Initial runs of a
42    subset of the G168 tests for tones (e.g ./echo_test 6) show the
43    current algorithm is passing OK, which is kind of surprising.  The
44    full set of tests needs to be performed to confirm this result.
45 
46    One other interesting change is that I have managed to get the NLMS
47    code to work with 16 bit coefficients, rather than the original 32
48    bit coefficents.  This reduces the MIPs and storage required.
49    I evaulated the 16 bit port using g168_tests.sh and listening tests
50    on 4 real-world samples.
51 
52    I also attempted the implementation of a block based NLMS update
53    [2] but although this passes g168_tests.sh it didn't converge well
54    on the real-world samples.  I have no idea why, perhaps a scaling
55    problem.  The block based code is also available in SVN
56    http://svn.rowetel.com/software/oslec/tags/before_16bit.  If this
57    code can be debugged, it will lead to further reduction in MIPS, as
58    the block update code maps nicely onto DSP instruction sets (it's a
59    dot product) compared to the current sample-by-sample update.
60 
61    Steve also has some nice notes on echo cancellers in echo.h
62 
63    References:
64 
65    [1] Ochiai, Areseki, and Ogihara, "Echo Canceller with Two Echo
66        Path Models", IEEE Transactions on communications, COM-25,
67        No. 6, June
68        1977.
69        https://www.rowetel.com/images/echo/dual_path_paper.pdf
70 
71    [2] The classic, very useful paper that tells you how to
72        actually build a real world echo canceller:
73 	 Messerschmitt, Hedberg, Cole, Haoui, Winship, "Digital Voice
74 	 Echo Canceller with a TMS320020,
75 	 https://www.rowetel.com/images/echo/spra129.pdf
76 
77    [3] I have written a series of blog posts on this work, here is
78        Part 1: http://www.rowetel.com/blog/?p=18
79 
80    [4] The source code http://svn.rowetel.com/software/oslec/
81 
82    [5] A nice reference on LMS filters:
83 	 https://en.wikipedia.org/wiki/Least_mean_squares_filter
84 
85    Credits:
86 
87    Thanks to Steve Underwood, Jean-Marc Valin, and Ramakrishnan
88    Muthukrishnan for their suggestions and email discussions.  Thanks
89    also to those people who collected echo samples for me such as
90    Mark, Pawel, and Pavel.
91 */
92 
93 #include <linux/kernel.h>
94 #include <linux/module.h>
95 #include <linux/slab.h>
96 
97 #include "echo.h"
98 
99 #define MIN_TX_POWER_FOR_ADAPTION	64
100 #define MIN_RX_POWER_FOR_ADAPTION	64
101 #define DTD_HANGOVER			600	/* 600 samples, or 75ms     */
102 #define DC_LOG2BETA			3	/* log2() of DC filter Beta */
103 
104 /* adapting coeffs using the traditional stochastic descent (N)LMS algorithm */
105 
lms_adapt_bg(struct oslec_state * ec,int clean,int shift)106 static inline void lms_adapt_bg(struct oslec_state *ec, int clean, int shift)
107 {
108 	int i;
109 
110 	int offset1;
111 	int offset2;
112 	int factor;
113 	int exp;
114 
115 	if (shift > 0)
116 		factor = clean << shift;
117 	else
118 		factor = clean >> -shift;
119 
120 	/* Update the FIR taps */
121 
122 	offset2 = ec->curr_pos;
123 	offset1 = ec->taps - offset2;
124 
125 	for (i = ec->taps - 1; i >= offset1; i--) {
126 		exp = (ec->fir_state_bg.history[i - offset1] * factor);
127 		ec->fir_taps16[1][i] += (int16_t) ((exp + (1 << 14)) >> 15);
128 	}
129 	for (; i >= 0; i--) {
130 		exp = (ec->fir_state_bg.history[i + offset2] * factor);
131 		ec->fir_taps16[1][i] += (int16_t) ((exp + (1 << 14)) >> 15);
132 	}
133 }
134 
top_bit(unsigned int bits)135 static inline int top_bit(unsigned int bits)
136 {
137 	if (bits == 0)
138 		return -1;
139 	else
140 		return (int)fls((int32_t) bits) - 1;
141 }
142 
oslec_create(int len,int adaption_mode)143 struct oslec_state *oslec_create(int len, int adaption_mode)
144 {
145 	struct oslec_state *ec;
146 	int i;
147 	const int16_t *history;
148 
149 	ec = kzalloc(sizeof(*ec), GFP_KERNEL);
150 	if (!ec)
151 		return NULL;
152 
153 	ec->taps = len;
154 	ec->log2taps = top_bit(len);
155 	ec->curr_pos = ec->taps - 1;
156 
157 	ec->fir_taps16[0] =
158 	    kcalloc(ec->taps, sizeof(int16_t), GFP_KERNEL);
159 	if (!ec->fir_taps16[0])
160 		goto error_oom_0;
161 
162 	ec->fir_taps16[1] =
163 	    kcalloc(ec->taps, sizeof(int16_t), GFP_KERNEL);
164 	if (!ec->fir_taps16[1])
165 		goto error_oom_1;
166 
167 	history = fir16_create(&ec->fir_state, ec->fir_taps16[0], ec->taps);
168 	if (!history)
169 		goto error_state;
170 	history = fir16_create(&ec->fir_state_bg, ec->fir_taps16[1], ec->taps);
171 	if (!history)
172 		goto error_state_bg;
173 
174 	for (i = 0; i < 5; i++)
175 		ec->xvtx[i] = ec->yvtx[i] = ec->xvrx[i] = ec->yvrx[i] = 0;
176 
177 	ec->cng_level = 1000;
178 	oslec_adaption_mode(ec, adaption_mode);
179 
180 	ec->snapshot = kcalloc(ec->taps, sizeof(int16_t), GFP_KERNEL);
181 	if (!ec->snapshot)
182 		goto error_snap;
183 
184 	ec->cond_met = 0;
185 	ec->pstates = 0;
186 	ec->ltxacc = ec->lrxacc = ec->lcleanacc = ec->lclean_bgacc = 0;
187 	ec->ltx = ec->lrx = ec->lclean = ec->lclean_bg = 0;
188 	ec->tx_1 = ec->tx_2 = ec->rx_1 = ec->rx_2 = 0;
189 	ec->lbgn = ec->lbgn_acc = 0;
190 	ec->lbgn_upper = 200;
191 	ec->lbgn_upper_acc = ec->lbgn_upper << 13;
192 
193 	return ec;
194 
195 error_snap:
196 	fir16_free(&ec->fir_state_bg);
197 error_state_bg:
198 	fir16_free(&ec->fir_state);
199 error_state:
200 	kfree(ec->fir_taps16[1]);
201 error_oom_1:
202 	kfree(ec->fir_taps16[0]);
203 error_oom_0:
204 	kfree(ec);
205 	return NULL;
206 }
207 EXPORT_SYMBOL_GPL(oslec_create);
208 
oslec_free(struct oslec_state * ec)209 void oslec_free(struct oslec_state *ec)
210 {
211 	int i;
212 
213 	fir16_free(&ec->fir_state);
214 	fir16_free(&ec->fir_state_bg);
215 	for (i = 0; i < 2; i++)
216 		kfree(ec->fir_taps16[i]);
217 	kfree(ec->snapshot);
218 	kfree(ec);
219 }
220 EXPORT_SYMBOL_GPL(oslec_free);
221 
oslec_adaption_mode(struct oslec_state * ec,int adaption_mode)222 void oslec_adaption_mode(struct oslec_state *ec, int adaption_mode)
223 {
224 	ec->adaption_mode = adaption_mode;
225 }
226 EXPORT_SYMBOL_GPL(oslec_adaption_mode);
227 
oslec_flush(struct oslec_state * ec)228 void oslec_flush(struct oslec_state *ec)
229 {
230 	int i;
231 
232 	ec->ltxacc = ec->lrxacc = ec->lcleanacc = ec->lclean_bgacc = 0;
233 	ec->ltx = ec->lrx = ec->lclean = ec->lclean_bg = 0;
234 	ec->tx_1 = ec->tx_2 = ec->rx_1 = ec->rx_2 = 0;
235 
236 	ec->lbgn = ec->lbgn_acc = 0;
237 	ec->lbgn_upper = 200;
238 	ec->lbgn_upper_acc = ec->lbgn_upper << 13;
239 
240 	ec->nonupdate_dwell = 0;
241 
242 	fir16_flush(&ec->fir_state);
243 	fir16_flush(&ec->fir_state_bg);
244 	ec->fir_state.curr_pos = ec->taps - 1;
245 	ec->fir_state_bg.curr_pos = ec->taps - 1;
246 	for (i = 0; i < 2; i++)
247 		memset(ec->fir_taps16[i], 0, ec->taps * sizeof(int16_t));
248 
249 	ec->curr_pos = ec->taps - 1;
250 	ec->pstates = 0;
251 }
252 EXPORT_SYMBOL_GPL(oslec_flush);
253 
oslec_snapshot(struct oslec_state * ec)254 void oslec_snapshot(struct oslec_state *ec)
255 {
256 	memcpy(ec->snapshot, ec->fir_taps16[0], ec->taps * sizeof(int16_t));
257 }
258 EXPORT_SYMBOL_GPL(oslec_snapshot);
259 
260 /* Dual Path Echo Canceller */
261 
oslec_update(struct oslec_state * ec,int16_t tx,int16_t rx)262 int16_t oslec_update(struct oslec_state *ec, int16_t tx, int16_t rx)
263 {
264 	int32_t echo_value;
265 	int clean_bg;
266 	int tmp;
267 	int tmp1;
268 
269 	/*
270 	 * Input scaling was found be required to prevent problems when tx
271 	 * starts clipping.  Another possible way to handle this would be the
272 	 * filter coefficent scaling.
273 	 */
274 
275 	ec->tx = tx;
276 	ec->rx = rx;
277 	tx >>= 1;
278 	rx >>= 1;
279 
280 	/*
281 	 * Filter DC, 3dB point is 160Hz (I think), note 32 bit precision
282 	 * required otherwise values do not track down to 0. Zero at DC, Pole
283 	 * at (1-Beta) on real axis.  Some chip sets (like Si labs) don't
284 	 * need this, but something like a $10 X100P card does.  Any DC really
285 	 * slows down convergence.
286 	 *
287 	 * Note: removes some low frequency from the signal, this reduces the
288 	 * speech quality when listening to samples through headphones but may
289 	 * not be obvious through a telephone handset.
290 	 *
291 	 * Note that the 3dB frequency in radians is approx Beta, e.g. for Beta
292 	 * = 2^(-3) = 0.125, 3dB freq is 0.125 rads = 159Hz.
293 	 */
294 
295 	if (ec->adaption_mode & ECHO_CAN_USE_RX_HPF) {
296 		tmp = rx << 15;
297 
298 		/*
299 		 * Make sure the gain of the HPF is 1.0. This can still
300 		 * saturate a little under impulse conditions, and it might
301 		 * roll to 32768 and need clipping on sustained peak level
302 		 * signals. However, the scale of such clipping is small, and
303 		 * the error due to any saturation should not markedly affect
304 		 * the downstream processing.
305 		 */
306 		tmp -= (tmp >> 4);
307 
308 		ec->rx_1 += -(ec->rx_1 >> DC_LOG2BETA) + tmp - ec->rx_2;
309 
310 		/*
311 		 * hard limit filter to prevent clipping.  Note that at this
312 		 * stage rx should be limited to +/- 16383 due to right shift
313 		 * above
314 		 */
315 		tmp1 = ec->rx_1 >> 15;
316 		if (tmp1 > 16383)
317 			tmp1 = 16383;
318 		if (tmp1 < -16383)
319 			tmp1 = -16383;
320 		rx = tmp1;
321 		ec->rx_2 = tmp;
322 	}
323 
324 	/* Block average of power in the filter states.  Used for
325 	   adaption power calculation. */
326 
327 	{
328 		int new, old;
329 
330 		/* efficient "out with the old and in with the new" algorithm so
331 		   we don't have to recalculate over the whole block of
332 		   samples. */
333 		new = (int)tx * (int)tx;
334 		old = (int)ec->fir_state.history[ec->fir_state.curr_pos] *
335 		    (int)ec->fir_state.history[ec->fir_state.curr_pos];
336 		ec->pstates +=
337 		    ((new - old) + (1 << (ec->log2taps - 1))) >> ec->log2taps;
338 		if (ec->pstates < 0)
339 			ec->pstates = 0;
340 	}
341 
342 	/* Calculate short term average levels using simple single pole IIRs */
343 
344 	ec->ltxacc += abs(tx) - ec->ltx;
345 	ec->ltx = (ec->ltxacc + (1 << 4)) >> 5;
346 	ec->lrxacc += abs(rx) - ec->lrx;
347 	ec->lrx = (ec->lrxacc + (1 << 4)) >> 5;
348 
349 	/* Foreground filter */
350 
351 	ec->fir_state.coeffs = ec->fir_taps16[0];
352 	echo_value = fir16(&ec->fir_state, tx);
353 	ec->clean = rx - echo_value;
354 	ec->lcleanacc += abs(ec->clean) - ec->lclean;
355 	ec->lclean = (ec->lcleanacc + (1 << 4)) >> 5;
356 
357 	/* Background filter */
358 
359 	echo_value = fir16(&ec->fir_state_bg, tx);
360 	clean_bg = rx - echo_value;
361 	ec->lclean_bgacc += abs(clean_bg) - ec->lclean_bg;
362 	ec->lclean_bg = (ec->lclean_bgacc + (1 << 4)) >> 5;
363 
364 	/* Background Filter adaption */
365 
366 	/* Almost always adap bg filter, just simple DT and energy
367 	   detection to minimise adaption in cases of strong double talk.
368 	   However this is not critical for the dual path algorithm.
369 	 */
370 	ec->factor = 0;
371 	ec->shift = 0;
372 	if (!ec->nonupdate_dwell) {
373 		int p, logp, shift;
374 
375 		/* Determine:
376 
377 		   f = Beta * clean_bg_rx/P ------ (1)
378 
379 		   where P is the total power in the filter states.
380 
381 		   The Boffins have shown that if we obey (1) we converge
382 		   quickly and avoid instability.
383 
384 		   The correct factor f must be in Q30, as this is the fixed
385 		   point format required by the lms_adapt_bg() function,
386 		   therefore the scaled version of (1) is:
387 
388 		   (2^30) * f  = (2^30) * Beta * clean_bg_rx/P
389 		   factor      = (2^30) * Beta * clean_bg_rx/P     ----- (2)
390 
391 		   We have chosen Beta = 0.25 by experiment, so:
392 
393 		   factor      = (2^30) * (2^-2) * clean_bg_rx/P
394 
395 		   (30 - 2 - log2(P))
396 		   factor      = clean_bg_rx 2                     ----- (3)
397 
398 		   To avoid a divide we approximate log2(P) as top_bit(P),
399 		   which returns the position of the highest non-zero bit in
400 		   P.  This approximation introduces an error as large as a
401 		   factor of 2, but the algorithm seems to handle it OK.
402 
403 		   Come to think of it a divide may not be a big deal on a
404 		   modern DSP, so its probably worth checking out the cycles
405 		   for a divide versus a top_bit() implementation.
406 		 */
407 
408 		p = MIN_TX_POWER_FOR_ADAPTION + ec->pstates;
409 		logp = top_bit(p) + ec->log2taps;
410 		shift = 30 - 2 - logp;
411 		ec->shift = shift;
412 
413 		lms_adapt_bg(ec, clean_bg, shift);
414 	}
415 
416 	/* very simple DTD to make sure we dont try and adapt with strong
417 	   near end speech */
418 
419 	ec->adapt = 0;
420 	if ((ec->lrx > MIN_RX_POWER_FOR_ADAPTION) && (ec->lrx > ec->ltx))
421 		ec->nonupdate_dwell = DTD_HANGOVER;
422 	if (ec->nonupdate_dwell)
423 		ec->nonupdate_dwell--;
424 
425 	/* Transfer logic */
426 
427 	/* These conditions are from the dual path paper [1], I messed with
428 	   them a bit to improve performance. */
429 
430 	if ((ec->adaption_mode & ECHO_CAN_USE_ADAPTION) &&
431 	    (ec->nonupdate_dwell == 0) &&
432 	    /* (ec->Lclean_bg < 0.875*ec->Lclean) */
433 	    (8 * ec->lclean_bg < 7 * ec->lclean) &&
434 	    /* (ec->Lclean_bg < 0.125*ec->Ltx) */
435 	    (8 * ec->lclean_bg < ec->ltx)) {
436 		if (ec->cond_met == 6) {
437 			/*
438 			 * BG filter has had better results for 6 consecutive
439 			 * samples
440 			 */
441 			ec->adapt = 1;
442 			memcpy(ec->fir_taps16[0], ec->fir_taps16[1],
443 			       ec->taps * sizeof(int16_t));
444 		} else
445 			ec->cond_met++;
446 	} else
447 		ec->cond_met = 0;
448 
449 	/* Non-Linear Processing */
450 
451 	ec->clean_nlp = ec->clean;
452 	if (ec->adaption_mode & ECHO_CAN_USE_NLP) {
453 		/*
454 		 * Non-linear processor - a fancy way to say "zap small
455 		 * signals, to avoid residual echo due to (uLaw/ALaw)
456 		 * non-linearity in the channel.".
457 		 */
458 
459 		if ((16 * ec->lclean < ec->ltx)) {
460 			/*
461 			 * Our e/c has improved echo by at least 24 dB (each
462 			 * factor of 2 is 6dB, so 2*2*2*2=16 is the same as
463 			 * 6+6+6+6=24dB)
464 			 */
465 			if (ec->adaption_mode & ECHO_CAN_USE_CNG) {
466 				ec->cng_level = ec->lbgn;
467 
468 				/*
469 				 * Very elementary comfort noise generation.
470 				 * Just random numbers rolled off very vaguely
471 				 * Hoth-like.  DR: This noise doesn't sound
472 				 * quite right to me - I suspect there are some
473 				 * overflow issues in the filtering as it's too
474 				 * "crackly".
475 				 * TODO: debug this, maybe just play noise at
476 				 * high level or look at spectrum.
477 				 */
478 
479 				ec->cng_rndnum =
480 				    1664525U * ec->cng_rndnum + 1013904223U;
481 				ec->cng_filter =
482 				    ((ec->cng_rndnum & 0xFFFF) - 32768 +
483 				     5 * ec->cng_filter) >> 3;
484 				ec->clean_nlp =
485 				    (ec->cng_filter * ec->cng_level * 8) >> 14;
486 
487 			} else if (ec->adaption_mode & ECHO_CAN_USE_CLIP) {
488 				/* This sounds much better than CNG */
489 				if (ec->clean_nlp > ec->lbgn)
490 					ec->clean_nlp = ec->lbgn;
491 				if (ec->clean_nlp < -ec->lbgn)
492 					ec->clean_nlp = -ec->lbgn;
493 			} else {
494 				/*
495 				 * just mute the residual, doesn't sound very
496 				 * good, used mainly in G168 tests
497 				 */
498 				ec->clean_nlp = 0;
499 			}
500 		} else {
501 			/*
502 			 * Background noise estimator.  I tried a few
503 			 * algorithms here without much luck.  This very simple
504 			 * one seems to work best, we just average the level
505 			 * using a slow (1 sec time const) filter if the
506 			 * current level is less than a (experimentally
507 			 * derived) constant.  This means we dont include high
508 			 * level signals like near end speech.  When combined
509 			 * with CNG or especially CLIP seems to work OK.
510 			 */
511 			if (ec->lclean < 40) {
512 				ec->lbgn_acc += abs(ec->clean) - ec->lbgn;
513 				ec->lbgn = (ec->lbgn_acc + (1 << 11)) >> 12;
514 			}
515 		}
516 	}
517 
518 	/* Roll around the taps buffer */
519 	if (ec->curr_pos <= 0)
520 		ec->curr_pos = ec->taps;
521 	ec->curr_pos--;
522 
523 	if (ec->adaption_mode & ECHO_CAN_DISABLE)
524 		ec->clean_nlp = rx;
525 
526 	/* Output scaled back up again to match input scaling */
527 
528 	return (int16_t) ec->clean_nlp << 1;
529 }
530 EXPORT_SYMBOL_GPL(oslec_update);
531 
532 /* This function is separated from the echo canceller is it is usually called
533    as part of the tx process.  See rx HP (DC blocking) filter above, it's
534    the same design.
535 
536    Some soft phones send speech signals with a lot of low frequency
537    energy, e.g. down to 20Hz.  This can make the hybrid non-linear
538    which causes the echo canceller to fall over.  This filter can help
539    by removing any low frequency before it gets to the tx port of the
540    hybrid.
541 
542    It can also help by removing and DC in the tx signal.  DC is bad
543    for LMS algorithms.
544 
545    This is one of the classic DC removal filters, adjusted to provide
546    sufficient bass rolloff to meet the above requirement to protect hybrids
547    from things that upset them. The difference between successive samples
548    produces a lousy HPF, and then a suitably placed pole flattens things out.
549    The final result is a nicely rolled off bass end. The filtering is
550    implemented with extended fractional precision, which noise shapes things,
551    giving very clean DC removal.
552 */
553 
oslec_hpf_tx(struct oslec_state * ec,int16_t tx)554 int16_t oslec_hpf_tx(struct oslec_state *ec, int16_t tx)
555 {
556 	int tmp;
557 	int tmp1;
558 
559 	if (ec->adaption_mode & ECHO_CAN_USE_TX_HPF) {
560 		tmp = tx << 15;
561 
562 		/*
563 		 * Make sure the gain of the HPF is 1.0. The first can still
564 		 * saturate a little under impulse conditions, and it might
565 		 * roll to 32768 and need clipping on sustained peak level
566 		 * signals. However, the scale of such clipping is small, and
567 		 * the error due to any saturation should not markedly affect
568 		 * the downstream processing.
569 		 */
570 		tmp -= (tmp >> 4);
571 
572 		ec->tx_1 += -(ec->tx_1 >> DC_LOG2BETA) + tmp - ec->tx_2;
573 		tmp1 = ec->tx_1 >> 15;
574 		if (tmp1 > 32767)
575 			tmp1 = 32767;
576 		if (tmp1 < -32767)
577 			tmp1 = -32767;
578 		tx = tmp1;
579 		ec->tx_2 = tmp;
580 	}
581 
582 	return tx;
583 }
584 EXPORT_SYMBOL_GPL(oslec_hpf_tx);
585 
586 MODULE_LICENSE("GPL");
587 MODULE_AUTHOR("David Rowe");
588 MODULE_DESCRIPTION("Open Source Line Echo Canceller");
589 MODULE_VERSION("0.3.0");
590