diff options
Diffstat (limited to 'drivers/misc/echo')
| -rw-r--r-- | drivers/misc/echo/Kconfig | 9 | ||||
| -rw-r--r-- | drivers/misc/echo/Makefile | 1 | ||||
| -rw-r--r-- | drivers/misc/echo/echo.c | 674 | ||||
| -rw-r--r-- | drivers/misc/echo/echo.h | 187 | ||||
| -rw-r--r-- | drivers/misc/echo/fir.h | 216 | ||||
| -rw-r--r-- | drivers/misc/echo/oslec.h | 94 |
6 files changed, 1181 insertions, 0 deletions
diff --git a/drivers/misc/echo/Kconfig b/drivers/misc/echo/Kconfig new file mode 100644 index 000000000000..f1d41ea9cd48 --- /dev/null +++ b/drivers/misc/echo/Kconfig @@ -0,0 +1,9 @@ +config ECHO + tristate "Line Echo Canceller support" + default n + ---help--- + This driver provides line echo cancelling support for mISDN and + Zaptel drivers. + + To compile this driver as a module, choose M here. The module + will be called echo. diff --git a/drivers/misc/echo/Makefile b/drivers/misc/echo/Makefile new file mode 100644 index 000000000000..7d4caac12a8d --- /dev/null +++ b/drivers/misc/echo/Makefile @@ -0,0 +1 @@ +obj-$(CONFIG_ECHO) += echo.o diff --git a/drivers/misc/echo/echo.c b/drivers/misc/echo/echo.c new file mode 100644 index 000000000000..9597e9523cac --- /dev/null +++ b/drivers/misc/echo/echo.c @@ -0,0 +1,674 @@ +/* + * SpanDSP - a series of DSP components for telephony + * + * echo.c - A line echo canceller. This code is being developed + * against and partially complies with G168. + * + * Written by Steve Underwood <steveu@coppice.org> + * and David Rowe <david_at_rowetel_dot_com> + * + * Copyright (C) 2001, 2003 Steve Underwood, 2007 David Rowe + * + * Based on a bit from here, a bit from there, eye of toad, ear of + * bat, 15 years of failed attempts by David and a few fried brain + * cells. + * + * All rights reserved. + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2, as + * published by the Free Software Foundation. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + */ + +/*! \file */ + +/* Implementation Notes + David Rowe + April 2007 + + This code started life as Steve's NLMS algorithm with a tap + rotation algorithm to handle divergence during double talk. I + added a Geigel Double Talk Detector (DTD) [2] and performed some + G168 tests. However I had trouble meeting the G168 requirements, + especially for double talk - there were always cases where my DTD + failed, for example where near end speech was under the 6dB + threshold required for declaring double talk. + + So I tried a two path algorithm [1], which has so far given better + results. The original tap rotation/Geigel algorithm is available + in SVN http://svn.rowetel.com/software/oslec/tags/before_16bit. + It's probably possible to make it work if some one wants to put some + serious work into it. + + At present no special treatment is provided for tones, which + generally cause NLMS algorithms to diverge. Initial runs of a + subset of the G168 tests for tones (e.g ./echo_test 6) show the + current algorithm is passing OK, which is kind of surprising. The + full set of tests needs to be performed to confirm this result. + + One other interesting change is that I have managed to get the NLMS + code to work with 16 bit coefficients, rather than the original 32 + bit coefficents. This reduces the MIPs and storage required. + I evaulated the 16 bit port using g168_tests.sh and listening tests + on 4 real-world samples. + + I also attempted the implementation of a block based NLMS update + [2] but although this passes g168_tests.sh it didn't converge well + on the real-world samples. I have no idea why, perhaps a scaling + problem. The block based code is also available in SVN + http://svn.rowetel.com/software/oslec/tags/before_16bit. If this + code can be debugged, it will lead to further reduction in MIPS, as + the block update code maps nicely onto DSP instruction sets (it's a + dot product) compared to the current sample-by-sample update. + + Steve also has some nice notes on echo cancellers in echo.h + + References: + + [1] Ochiai, Areseki, and Ogihara, "Echo Canceller with Two Echo + Path Models", IEEE Transactions on communications, COM-25, + No. 6, June + 1977. + http://www.rowetel.com/images/echo/dual_path_paper.pdf + + [2] The classic, very useful paper that tells you how to + actually build a real world echo canceller: + Messerschmitt, Hedberg, Cole, Haoui, Winship, "Digital Voice + Echo Canceller with a TMS320020, + http://www.rowetel.com/images/echo/spra129.pdf + + [3] I have written a series of blog posts on this work, here is + Part 1: http://www.rowetel.com/blog/?p=18 + + [4] The source code http://svn.rowetel.com/software/oslec/ + + [5] A nice reference on LMS filters: + http://en.wikipedia.org/wiki/Least_mean_squares_filter + + Credits: + + Thanks to Steve Underwood, Jean-Marc Valin, and Ramakrishnan + Muthukrishnan for their suggestions and email discussions. Thanks + also to those people who collected echo samples for me such as + Mark, Pawel, and Pavel. +*/ + +#include <linux/kernel.h> +#include <linux/module.h> +#include <linux/slab.h> + +#include "echo.h" + +#define MIN_TX_POWER_FOR_ADAPTION 64 +#define MIN_RX_POWER_FOR_ADAPTION 64 +#define DTD_HANGOVER 600 /* 600 samples, or 75ms */ +#define DC_LOG2BETA 3 /* log2() of DC filter Beta */ + +/* adapting coeffs using the traditional stochastic descent (N)LMS algorithm */ + +#ifdef __bfin__ +static inline void lms_adapt_bg(struct oslec_state *ec, int clean, int shift) +{ + int i; + int offset1; + int offset2; + int factor; + int exp; + int16_t *phist; + int n; + + if (shift > 0) + factor = clean << shift; + else + factor = clean >> -shift; + + /* Update the FIR taps */ + + offset2 = ec->curr_pos; + offset1 = ec->taps - offset2; + phist = &ec->fir_state_bg.history[offset2]; + + /* st: and en: help us locate the assembler in echo.s */ + + /* asm("st:"); */ + n = ec->taps; + for (i = 0; i < n; i++) { + exp = *phist++ * factor; + ec->fir_taps16[1][i] += (int16_t) ((exp + (1 << 14)) >> 15); + } + /* asm("en:"); */ + + /* Note the asm for the inner loop above generated by Blackfin gcc + 4.1.1 is pretty good (note even parallel instructions used): + + R0 = W [P0++] (X); + R0 *= R2; + R0 = R0 + R3 (NS) || + R1 = W [P1] (X) || + nop; + R0 >>>= 15; + R0 = R0 + R1; + W [P1++] = R0; + + A block based update algorithm would be much faster but the + above can't be improved on much. Every instruction saved in + the loop above is 2 MIPs/ch! The for loop above is where the + Blackfin spends most of it's time - about 17 MIPs/ch measured + with speedtest.c with 256 taps (32ms). Write-back and + Write-through cache gave about the same performance. + */ +} + +/* + IDEAS for further optimisation of lms_adapt_bg(): + + 1/ The rounding is quite costly. Could we keep as 32 bit coeffs + then make filter pluck the MS 16-bits of the coeffs when filtering? + However this would lower potential optimisation of filter, as I + think the dual-MAC architecture requires packed 16 bit coeffs. + + 2/ Block based update would be more efficient, as per comments above, + could use dual MAC architecture. + + 3/ Look for same sample Blackfin LMS code, see if we can get dual-MAC + packing. + + 4/ Execute the whole e/c in a block of say 20ms rather than sample + by sample. Processing a few samples every ms is inefficient. +*/ + +#else +static inline void lms_adapt_bg(struct oslec_state *ec, int clean, int shift) +{ + int i; + + int offset1; + int offset2; + int factor; + int exp; + + if (shift > 0) + factor = clean << shift; + else + factor = clean >> -shift; + + /* Update the FIR taps */ + + offset2 = ec->curr_pos; + offset1 = ec->taps - offset2; + + for (i = ec->taps - 1; i >= offset1; i--) { + exp = (ec->fir_state_bg.history[i - offset1] * factor); + ec->fir_taps16[1][i] += (int16_t) ((exp + (1 << 14)) >> 15); + } + for (; i >= 0; i--) { + exp = (ec->fir_state_bg.history[i + offset2] * factor); + ec->fir_taps16[1][i] += (int16_t) ((exp + (1 << 14)) >> 15); + } +} +#endif + +static inline int top_bit(unsigned int bits) +{ + if (bits == 0) + return -1; + else + return (int)fls((int32_t) bits) - 1; +} + +struct oslec_state *oslec_create(int len, int adaption_mode) +{ + struct oslec_state *ec; + int i; + const int16_t *history; + + ec = kzalloc(sizeof(*ec), GFP_KERNEL); + if (!ec) + return NULL; + + ec->taps = len; + ec->log2taps = top_bit(len); + ec->curr_pos = ec->taps - 1; + + ec->fir_taps16[0] = + kcalloc(ec->taps, sizeof(int16_t), GFP_KERNEL); + if (!ec->fir_taps16[0]) + goto error_oom_0; + + ec->fir_taps16[1] = + kcalloc(ec->taps, sizeof(int16_t), GFP_KERNEL); + if (!ec->fir_taps16[1]) + goto error_oom_1; + + history = fir16_create(&ec->fir_state, ec->fir_taps16[0], ec->taps); + if (!history) + goto error_state; + history = fir16_create(&ec->fir_state_bg, ec->fir_taps16[1], ec->taps); + if (!history) + goto error_state_bg; + + for (i = 0; i < 5; i++) + ec->xvtx[i] = ec->yvtx[i] = ec->xvrx[i] = ec->yvrx[i] = 0; + + ec->cng_level = 1000; + oslec_adaption_mode(ec, adaption_mode); + + ec->snapshot = kcalloc(ec->taps, sizeof(int16_t), GFP_KERNEL); + if (!ec->snapshot) + goto error_snap; + + ec->cond_met = 0; + ec->pstates = 0; + ec->ltxacc = ec->lrxacc = ec->lcleanacc = ec->lclean_bgacc = 0; + ec->ltx = ec->lrx = ec->lclean = ec->lclean_bg = 0; + ec->tx_1 = ec->tx_2 = ec->rx_1 = ec->rx_2 = 0; + ec->lbgn = ec->lbgn_acc = 0; + ec->lbgn_upper = 200; + ec->lbgn_upper_acc = ec->lbgn_upper << 13; + + return ec; + +error_snap: + fir16_free(&ec->fir_state_bg); +error_state_bg: + fir16_free(&ec->fir_state); +error_state: + kfree(ec->fir_taps16[1]); +error_oom_1: + kfree(ec->fir_taps16[0]); +error_oom_0: + kfree(ec); + return NULL; +} +EXPORT_SYMBOL_GPL(oslec_create); + +void oslec_free(struct oslec_state *ec) +{ + int i; + + fir16_free(&ec->fir_state); + fir16_free(&ec->fir_state_bg); + for (i = 0; i < 2; i++) + kfree(ec->fir_taps16[i]); + kfree(ec->snapshot); + kfree(ec); +} +EXPORT_SYMBOL_GPL(oslec_free); + +void oslec_adaption_mode(struct oslec_state *ec, int adaption_mode) +{ + ec->adaption_mode = adaption_mode; +} +EXPORT_SYMBOL_GPL(oslec_adaption_mode); + +void oslec_flush(struct oslec_state *ec) +{ + int i; + + ec->ltxacc = ec->lrxacc = ec->lcleanacc = ec->lclean_bgacc = 0; + ec->ltx = ec->lrx = ec->lclean = ec->lclean_bg = 0; + ec->tx_1 = ec->tx_2 = ec->rx_1 = ec->rx_2 = 0; + + ec->lbgn = ec->lbgn_acc = 0; + ec->lbgn_upper = 200; + ec->lbgn_upper_acc = ec->lbgn_upper << 13; + + ec->nonupdate_dwell = 0; + + fir16_flush(&ec->fir_state); + fir16_flush(&ec->fir_state_bg); + ec->fir_state.curr_pos = ec->taps - 1; + ec->fir_state_bg.curr_pos = ec->taps - 1; + for (i = 0; i < 2; i++) + memset(ec->fir_taps16[i], 0, ec->taps * sizeof(int16_t)); + + ec->curr_pos = ec->taps - 1; + ec->pstates = 0; +} +EXPORT_SYMBOL_GPL(oslec_flush); + +void oslec_snapshot(struct oslec_state *ec) +{ + memcpy(ec->snapshot, ec->fir_taps16[0], ec->taps * sizeof(int16_t)); +} +EXPORT_SYMBOL_GPL(oslec_snapshot); + +/* Dual Path Echo Canceller */ + +int16_t oslec_update(struct oslec_state *ec, int16_t tx, int16_t rx) +{ + int32_t echo_value; + int clean_bg; + int tmp; + int tmp1; + + /* + * Input scaling was found be required to prevent problems when tx + * starts clipping. Another possible way to handle this would be the + * filter coefficent scaling. + */ + + ec->tx = tx; + ec->rx = rx; + tx >>= 1; + rx >>= 1; + + /* + * Filter DC, 3dB point is 160Hz (I think), note 32 bit precision + * required otherwise values do not track down to 0. Zero at DC, Pole + * at (1-Beta) on real axis. Some chip sets (like Si labs) don't + * need this, but something like a $10 X100P card does. Any DC really + * slows down convergence. + * + * Note: removes some low frequency from the signal, this reduces the + * speech quality when listening to samples through headphones but may + * not be obvious through a telephone handset. + * + * Note that the 3dB frequency in radians is approx Beta, e.g. for Beta + * = 2^(-3) = 0.125, 3dB freq is 0.125 rads = 159Hz. + */ + + if (ec->adaption_mode & ECHO_CAN_USE_RX_HPF) { + tmp = rx << 15; + + /* + * Make sure the gain of the HPF is 1.0. This can still + * saturate a little under impulse conditions, and it might + * roll to 32768 and need clipping on sustained peak level + * signals. However, the scale of such clipping is small, and + * the error due to any saturation should not markedly affect + * the downstream processing. + */ + tmp -= (tmp >> 4); + + ec->rx_1 += -(ec->rx_1 >> DC_LOG2BETA) + tmp - ec->rx_2; + + /* + * hard limit filter to prevent clipping. Note that at this + * stage rx should be limited to +/- 16383 due to right shift + * above + */ + tmp1 = ec->rx_1 >> 15; + if (tmp1 > 16383) + tmp1 = 16383; + if (tmp1 < -16383) + tmp1 = -16383; + rx = tmp1; + ec->rx_2 = tmp; + } + + /* Block average of power in the filter states. Used for + adaption power calculation. */ + + { + int new, old; + + /* efficient "out with the old and in with the new" algorithm so + we don't have to recalculate over the whole block of + samples. */ + new = (int)tx * (int)tx; + old = (int)ec->fir_state.history[ec->fir_state.curr_pos] * + (int)ec->fir_state.history[ec->fir_state.curr_pos]; + ec->pstates += + ((new - old) + (1 << (ec->log2taps - 1))) >> ec->log2taps; + if (ec->pstates < 0) + ec->pstates = 0; + } + + /* Calculate short term average levels using simple single pole IIRs */ + + ec->ltxacc += abs(tx) - ec->ltx; + ec->ltx = (ec->ltxacc + (1 << 4)) >> 5; + ec->lrxacc += abs(rx) - ec->lrx; + ec->lrx = (ec->lrxacc + (1 << 4)) >> 5; + + /* Foreground filter */ + + ec->fir_state.coeffs = ec->fir_taps16[0]; + echo_value = fir16(&ec->fir_state, tx); + ec->clean = rx - echo_value; + ec->lcleanacc += abs(ec->clean) - ec->lclean; + ec->lclean = (ec->lcleanacc + (1 << 4)) >> 5; + + /* Background filter */ + + echo_value = fir16(&ec->fir_state_bg, tx); + clean_bg = rx - echo_value; + ec->lclean_bgacc += abs(clean_bg) - ec->lclean_bg; + ec->lclean_bg = (ec->lclean_bgacc + (1 << 4)) >> 5; + + /* Background Filter adaption */ + + /* Almost always adap bg filter, just simple DT and energy + detection to minimise adaption in cases of strong double talk. + However this is not critical for the dual path algorithm. + */ + ec->factor = 0; + ec->shift = 0; + if ((ec->nonupdate_dwell == 0)) { + int p, logp, shift; + + /* Determine: + + f = Beta * clean_bg_rx/P ------ (1) + + where P is the total power in the filter states. + + The Boffins have shown that if we obey (1) we converge + quickly and avoid instability. + + The correct factor f must be in Q30, as this is the fixed + point format required by the lms_adapt_bg() function, + therefore the scaled version of (1) is: + + (2^30) * f = (2^30) * Beta * clean_bg_rx/P + factor = (2^30) * Beta * clean_bg_rx/P ----- (2) + + We have chosen Beta = 0.25 by experiment, so: + + factor = (2^30) * (2^-2) * clean_bg_rx/P + + (30 - 2 - log2(P)) + factor = clean_bg_rx 2 ----- (3) + + To avoid a divide we approximate log2(P) as top_bit(P), + which returns the position of the highest non-zero bit in + P. This approximation introduces an error as large as a + factor of 2, but the algorithm seems to handle it OK. + + Come to think of it a divide may not be a big deal on a + modern DSP, so its probably worth checking out the cycles + for a divide versus a top_bit() implementation. + */ + + p = MIN_TX_POWER_FOR_ADAPTION + ec->pstates; + logp = top_bit(p) + ec->log2taps; + shift = 30 - 2 - logp; + ec->shift = shift; + + lms_adapt_bg(ec, clean_bg, shift); + } + + /* very simple DTD to make sure we dont try and adapt with strong + near end speech */ + + ec->adapt = 0; + if ((ec->lrx > MIN_RX_POWER_FOR_ADAPTION) && (ec->lrx > ec->ltx)) + ec->nonupdate_dwell = DTD_HANGOVER; + if (ec->nonupdate_dwell) + ec->nonupdate_dwell--; + + /* Transfer logic */ + + /* These conditions are from the dual path paper [1], I messed with + them a bit to improve performance. */ + + if ((ec->adaption_mode & ECHO_CAN_USE_ADAPTION) && + (ec->nonupdate_dwell == 0) && + /* (ec->Lclean_bg < 0.875*ec->Lclean) */ + (8 * ec->lclean_bg < 7 * ec->lclean) && + /* (ec->Lclean_bg < 0.125*ec->Ltx) */ + (8 * ec->lclean_bg < ec->ltx)) { + if (ec->cond_met == 6) { + /* + * BG filter has had better results for 6 consecutive + * samples + */ + ec->adapt = 1; + memcpy(ec->fir_taps16[0], ec->fir_taps16[1], + ec->taps * sizeof(int16_t)); + } else + ec->cond_met++; + } else + ec->cond_met = 0; + + /* Non-Linear Processing */ + + ec->clean_nlp = ec->clean; + if (ec->adaption_mode & ECHO_CAN_USE_NLP) { + /* + * Non-linear processor - a fancy way to say "zap small + * signals, to avoid residual echo due to (uLaw/ALaw) + * non-linearity in the channel.". + */ + + if ((16 * ec->lclean < ec->ltx)) { + /* + * Our e/c has improved echo by at least 24 dB (each + * factor of 2 is 6dB, so 2*2*2*2=16 is the same as + * 6+6+6+6=24dB) + */ + if (ec->adaption_mode & ECHO_CAN_USE_CNG) { + ec->cng_level = ec->lbgn; + + /* + * Very elementary comfort noise generation. + * Just random numbers rolled off very vaguely + * Hoth-like. DR: This noise doesn't sound + * quite right to me - I suspect there are some + * overflow issues in the filtering as it's too + * "crackly". + * TODO: debug this, maybe just play noise at + * high level or look at spectrum. + */ + + ec->cng_rndnum = + 1664525U * ec->cng_rndnum + 1013904223U; + ec->cng_filter = + ((ec->cng_rndnum & 0xFFFF) - 32768 + + 5 * ec->cng_filter) >> 3; + ec->clean_nlp = + (ec->cng_filter * ec->cng_level * 8) >> 14; + + } else if (ec->adaption_mode & ECHO_CAN_USE_CLIP) { + /* This sounds much better than CNG */ + if (ec->clean_nlp > ec->lbgn) + ec->clean_nlp = ec->lbgn; + if (ec->clean_nlp < -ec->lbgn) + ec->clean_nlp = -ec->lbgn; + } else { + /* + * just mute the residual, doesn't sound very + * good, used mainly in G168 tests + */ + ec->clean_nlp = 0; + } + } else { + /* + * Background noise estimator. I tried a few + * algorithms here without much luck. This very simple + * one seems to work best, we just average the level + * using a slow (1 sec time const) filter if the + * current level is less than a (experimentally + * derived) constant. This means we dont include high + * level signals like near end speech. When combined + * with CNG or especially CLIP seems to work OK. + */ + if (ec->lclean < 40) { + ec->lbgn_acc += abs(ec->clean) - ec->lbgn; + ec->lbgn = (ec->lbgn_acc + (1 << 11)) >> 12; + } + } + } + + /* Roll around the taps buffer */ + if (ec->curr_pos <= 0) + ec->curr_pos = ec->taps; + ec->curr_pos--; + + if (ec->adaption_mode & ECHO_CAN_DISABLE) + ec->clean_nlp = rx; + + /* Output scaled back up again to match input scaling */ + + return (int16_t) ec->clean_nlp << 1; +} +EXPORT_SYMBOL_GPL(oslec_update); + +/* This function is separated from the echo canceller is it is usually called + as part of the tx process. See rx HP (DC blocking) filter above, it's + the same design. + + Some soft phones send speech signals with a lot of low frequency + energy, e.g. down to 20Hz. This can make the hybrid non-linear + which causes the echo canceller to fall over. This filter can help + by removing any low frequency before it gets to the tx port of the + hybrid. + + It can also help by removing and DC in the tx signal. DC is bad + for LMS algorithms. + + This is one of the classic DC removal filters, adjusted to provide + sufficient bass rolloff to meet the above requirement to protect hybrids + from things that upset them. The difference between successive samples + produces a lousy HPF, and then a suitably placed pole flattens things out. + The final result is a nicely rolled off bass end. The filtering is + implemented with extended fractional precision, which noise shapes things, + giving very clean DC removal. +*/ + +int16_t oslec_hpf_tx(struct oslec_state *ec, int16_t tx) +{ + int tmp; + int tmp1; + + if (ec->adaption_mode & ECHO_CAN_USE_TX_HPF) { + tmp = tx << 15; + + /* + * Make sure the gain of the HPF is 1.0. The first can still + * saturate a little under impulse conditions, and it might + * roll to 32768 and need clipping on sustained peak level + * signals. However, the scale of such clipping is small, and + * the error due to any saturation should not markedly affect + * the downstream processing. + */ + tmp -= (tmp >> 4); + + ec->tx_1 += -(ec->tx_1 >> DC_LOG2BETA) + tmp - ec->tx_2; + tmp1 = ec->tx_1 >> 15; + if (tmp1 > 32767) + tmp1 = 32767; + if (tmp1 < -32767) + tmp1 = -32767; + tx = tmp1; + ec->tx_2 = tmp; + } + + return tx; +} +EXPORT_SYMBOL_GPL(oslec_hpf_tx); + +MODULE_LICENSE("GPL"); +MODULE_AUTHOR("David Rowe"); +MODULE_DESCRIPTION("Open Source Line Echo Canceller"); +MODULE_VERSION("0.3.0"); diff --git a/drivers/misc/echo/echo.h b/drivers/misc/echo/echo.h new file mode 100644 index 000000000000..9b08c63e6369 --- /dev/null +++ b/drivers/misc/echo/echo.h @@ -0,0 +1,187 @@ +/* + * SpanDSP - a series of DSP components for telephony + * + * echo.c - A line echo canceller. This code is being developed + * against and partially complies with G168. + * + * Written by Steve Underwood <steveu@coppice.org> + * and David Rowe <david_at_rowetel_dot_com> + * + * Copyright (C) 2001 Steve Underwood and 2007 David Rowe + * + * All rights reserved. + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2, as + * published by the Free Software Foundation. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + */ + +#ifndef __ECHO_H +#define __ECHO_H + +/* +Line echo cancellation for voice + +What does it do? + +This module aims to provide G.168-2002 compliant echo cancellation, to remove +electrical echoes (e.g. from 2-4 wire hybrids) from voice calls. + +How does it work? + +The heart of the echo cancellor is FIR filter. This is adapted to match the +echo impulse response of the telephone line. It must be long enough to +adequately cover the duration of that impulse response. The signal transmitted +to the telephone line is passed through the FIR filter. Once the FIR is +properly adapted, the resulting output is an estimate of the echo signal +received from the line. This is subtracted from the received signal. The result +is an estimate of the signal which originated at the far end of the line, free +from echos of our own transmitted signal. + +The least mean squares (LMS) algorithm is attributed to Widrow and Hoff, and +was introduced in 1960. It is the commonest form of filter adaption used in +things like modem line equalisers and line echo cancellers. There it works very +well. However, it only works well for signals of constant amplitude. It works +very poorly for things like speech echo cancellation, where the signal level +varies widely. This is quite easy to fix. If the signal level is normalised - +similar to applying AGC - LMS can work as well for a signal of varying +amplitude as it does for a modem signal. This normalised least mean squares +(NLMS) algorithm is the commonest one used for speech echo cancellation. Many +other algorithms exist - e.g. RLS (essentially the same as Kalman filtering), +FAP, etc. Some perform significantly better than NLMS. However, factors such +as computational complexity and patents favour the use of NLMS. + +A simple refinement to NLMS can improve its performance with speech. NLMS tends +to adapt best to the strongest parts of a signal. If the signal is white noise, +the NLMS algorithm works very well. However, speech has more low frequency than +high frequency content. Pre-whitening (i.e. filtering the signal to flatten its +spectrum) the echo signal improves the adapt rate for speech, and ensures the +final residual signal is not heavily biased towards high frequencies. A very +low complexity filter is adequate for this, so pre-whitening adds little to the +compute requirements of the echo canceller. + +An FIR filter adapted using pre-whitened NLMS performs well, provided certain +conditions are met: + + - The transmitted signal has poor self-correlation. + - There is no signal being generated within the environment being + cancelled. + +The difficulty is that neither of these can be guaranteed. + +If the adaption is performed while transmitting noise (or something fairly +noise like, such as voice) the adaption works very well. If the adaption is +performed while transmitting something highly correlative (typically narrow +band energy such as signalling tones or DTMF), the adaption can go seriously +wrong. The reason is there is only one solution for the adaption on a near +random signal - the impulse response of the line. For a repetitive signal, +there are any number of solutions which converge the adaption, and nothing +guides the adaption to choose the generalised one. Allowing an untrained +canceller to converge on this kind of narrowband energy probably a good thing, +since at least it cancels the tones. Allowing a well converged canceller to +continue converging on such energy is just a way to ruin its generalised +adaption. A narrowband detector is needed, so adapation can be suspended at +appropriate times. + +The adaption process is based on trying to eliminate the received signal. When +there is any signal from within the environment being cancelled it may upset +the adaption process. Similarly, if the signal we are transmitting is small, +noise may dominate and disturb the adaption process. If we can ensure that the +adaption is only performed when we are transmitting a significant signal level, +and the environment is not, things will be OK. Clearly, it is easy to tell when +we are sending a significant signal. Telling, if the environment is generating +a significant signal, and doing it with sufficient speed that the adaption will +not have diverged too much more we stop it, is a little harder. + +The key problem in detecting when the environment is sourcing significant +energy is that we must do this very quickly. Given a reasonably long sample of +the received signal, there are a number of strategies which may be used to +assess whether that signal contains a strong far end component. However, by the +time that assessment is complete the far end signal will have already caused +major mis-convergence in the adaption process. An assessment algorithm is +needed which produces a fairly accurate result from a very short burst of far +end energy. + +How do I use it? + +The echo cancellor processes both the transmit and receive streams sample by +sample. The processing function is not declared inline. Unfortunately, +cancellation requires many operations per sample, so the call overhead is only +a minor burden. +*/ + +#include "fir.h" +#include "oslec.h" + +/* + G.168 echo canceller descriptor. This defines the working state for a line + echo canceller. +*/ +struct oslec_state { + int16_t tx; + int16_t rx; + int16_t clean; + int16_t clean_nlp; + + int nonupdate_dwell; + int curr_pos; + int taps; + int log2taps; + int adaption_mode; + + int cond_met; + int32_t pstates; + int16_t adapt; + int32_t factor; + int16_t shift; + + /* Average levels and averaging filter states */ + int ltxacc; + int lrxacc; + int lcleanacc; + int lclean_bgacc; + int ltx; + int lrx; + int lclean; + int lclean_bg; + int lbgn; + int lbgn_acc; + int lbgn_upper; + int lbgn_upper_acc; + + /* foreground and background filter states */ + struct fir16_state_t fir_state; + struct fir16_state_t fir_state_bg; + int16_t *fir_taps16[2]; + + /* DC blocking filter states */ + int tx_1; + int tx_2; + int rx_1; + int rx_2; + + /* optional High Pass Filter states */ + int32_t xvtx[5]; + int32_t yvtx[5]; + int32_t xvrx[5]; + int32_t yvrx[5]; + + /* Parameters for the optional Hoth noise generator */ + int cng_level; + int cng_rndnum; + int cng_filter; + + /* snapshot sample of coeffs used for development */ + int16_t *snapshot; +}; + +#endif /* __ECHO_H */ diff --git a/drivers/misc/echo/fir.h b/drivers/misc/echo/fir.h new file mode 100644 index 000000000000..7b9fabf1fea5 --- /dev/null +++ b/drivers/misc/echo/fir.h @@ -0,0 +1,216 @@ +/* + * SpanDSP - a series of DSP components for telephony + * + * fir.h - General telephony FIR routines + * + * Written by Steve Underwood <steveu@coppice.org> + * + * Copyright (C) 2002 Steve Underwood + * + * All rights reserved. + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2, as + * published by the Free Software Foundation. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + */ + +#if !defined(_FIR_H_) +#define _FIR_H_ + +/* + Blackfin NOTES & IDEAS: + + A simple dot product function is used to implement the filter. This performs + just one MAC/cycle which is inefficient but was easy to implement as a first + pass. The current Blackfin code also uses an unrolled form of the filter + history to avoid 0 length hardware loop issues. This is wasteful of + memory. + + Ideas for improvement: + + 1/ Rewrite filter for dual MAC inner loop. The issue here is handling + history sample offsets that are 16 bit aligned - the dual MAC needs + 32 bit aligmnent. There are some good examples in libbfdsp. + + 2/ Use the hardware circular buffer facility tohalve memory usage. + + 3/ Consider using internal memory. + + Using less memory might also improve speed as cache misses will be + reduced. A drop in MIPs and memory approaching 50% should be + possible. + + The foreground and background filters currenlty use a total of + about 10 MIPs/ch as measured with speedtest.c on a 256 TAP echo + can. +*/ + +/* + * 16 bit integer FIR descriptor. This defines the working state for a single + * instance of an FIR filter using 16 bit integer coefficients. + */ +struct fir16_state_t { + int taps; + int curr_pos; + const int16_t *coeffs; + int16_t *history; +}; + +/* + * 32 bit integer FIR descriptor. This defines the working state for a single + * instance of an FIR filter using 32 bit integer coefficients, and filtering + * 16 bit integer data. + */ +struct fir32_state_t { + int taps; + int curr_pos; + const int32_t *coeffs; + int16_t *history; +}; + +/* + * Floating point FIR descriptor. This defines the working state for a single + * instance of an FIR filter using floating point coefficients and data. + */ +struct fir_float_state_t { + int taps; + int curr_pos; + const float *coeffs; + float *history; +}; + +static inline const int16_t *fir16_create(struct fir16_state_t *fir, + const int16_t *coeffs, int taps) +{ + fir->taps = taps; + fir->curr_pos = taps - 1; + fir->coeffs = coeffs; +#if defined(__bfin__) + fir->history = kcalloc(2 * taps, sizeof(int16_t), GFP_KERNEL); +#else + fir->history = kcalloc(taps, sizeof(int16_t), GFP_KERNEL); +#endif + return fir->history; +} + +static inline void fir16_flush(struct fir16_state_t *fir) +{ +#if defined(__bfin__) + memset(fir->history, 0, 2 * fir->taps * sizeof(int16_t)); +#else + memset(fir->history, 0, fir->taps * sizeof(int16_t)); +#endif +} + +static inline void fir16_free(struct fir16_state_t *fir) +{ + kfree(fir->history); +} + +#ifdef __bfin__ +static inline int32_t dot_asm(short *x, short *y, int len) +{ + int dot; + + len--; + + __asm__("I0 = %1;\n\t" + "I1 = %2;\n\t" + "A0 = 0;\n\t" + "R0.L = W[I0++] || R1.L = W[I1++];\n\t" + "LOOP dot%= LC0 = %3;\n\t" + "LOOP_BEGIN dot%=;\n\t" + "A0 += R0.L * R1.L (IS) || R0.L = W[I0++] || R1.L = W[I1++];\n\t" + "LOOP_END dot%=;\n\t" + "A0 += R0.L*R1.L (IS);\n\t" + "R0 = A0;\n\t" + "%0 = R0;\n\t" + : "=&d"(dot) + : "a"(x), "a"(y), "a"(len) + : "I0", "I1", "A1", "A0", "R0", "R1" + ); + + return dot; +} +#endif + +static inline int16_t fir16(struct fir16_state_t *fir, int16_t sample) +{ + int32_t y; +#if defined(__bfin__) + fir->history[fir->curr_pos] = sample; + fir->history[fir->curr_pos + fir->taps] = sample; + y = dot_asm((int16_t *) fir->coeffs, &fir->history[fir->curr_pos], + fir->taps); +#else + int i; + int offset1; + int offset2; + + fir->history[fir->curr_pos] = sample; + + offset2 = fir->curr_pos; + offset1 = fir->taps - offset2; + y = 0; + for (i = fir->taps - 1; i >= offset1; i--) + y += fir->coeffs[i] * fir->history[i - offset1]; + for (; i >= 0; i--) + y += fir->coeffs[i] * fir->history[i + offset2]; +#endif + if (fir->curr_pos <= 0) + fir->curr_pos = fir->taps; + fir->curr_pos--; + return (int16_t) (y >> 15); +} + +static inline const int16_t *fir32_create(struct fir32_state_t *fir, + const int32_t *coeffs, int taps) +{ + fir->taps = taps; + fir->curr_pos = taps - 1; + fir->coeffs = coeffs; + fir->history = kcalloc(taps, sizeof(int16_t), GFP_KERNEL); + return fir->history; +} + +static inline void fir32_flush(struct fir32_state_t *fir) +{ + memset(fir->history, 0, fir->taps * sizeof(int16_t)); +} + +static inline void fir32_free(struct fir32_state_t *fir) +{ + kfree(fir->history); +} + +static inline int16_t fir32(struct fir32_state_t *fir, int16_t sample) +{ + int i; + int32_t y; + int offset1; + int offset2; + + fir->history[fir->curr_pos] = sample; + offset2 = fir->curr_pos; + offset1 = fir->taps - offset2; + y = 0; + for (i = fir->taps - 1; i >= offset1; i--) + y += fir->coeffs[i] * fir->history[i - offset1]; + for (; i >= 0; i--) + y += fir->coeffs[i] * fir->history[i + offset2]; + if (fir->curr_pos <= 0) + fir->curr_pos = fir->taps; + fir->curr_pos--; + return (int16_t) (y >> 15); +} + +#endif diff --git a/drivers/misc/echo/oslec.h b/drivers/misc/echo/oslec.h new file mode 100644 index 000000000000..f4175360ce27 --- /dev/null +++ b/drivers/misc/echo/oslec.h @@ -0,0 +1,94 @@ +/* + * OSLEC - A line echo canceller. This code is being developed + * against and partially complies with G168. Using code from SpanDSP + * + * Written by Steve Underwood <steveu@coppice.org> + * and David Rowe <david_at_rowetel_dot_com> + * + * Copyright (C) 2001 Steve Underwood and 2007-2008 David Rowe + * + * All rights reserved. + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2, as + * published by the Free Software Foundation. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * + */ + +#ifndef __OSLEC_H +#define __OSLEC_H + +/* Mask bits for the adaption mode */ +#define ECHO_CAN_USE_ADAPTION 0x01 +#define ECHO_CAN_USE_NLP 0x02 +#define ECHO_CAN_USE_CNG 0x04 +#define ECHO_CAN_USE_CLIP 0x08 +#define ECHO_CAN_USE_TX_HPF 0x10 +#define ECHO_CAN_USE_RX_HPF 0x20 +#define ECHO_CAN_DISABLE 0x40 + +/** + * oslec_state: G.168 echo canceller descriptor. + * + * This defines the working state for a line echo canceller. + */ +struct oslec_state; + +/** + * oslec_create - Create a voice echo canceller context. + * @len: The length of the canceller, in samples. + * @return: The new canceller context, or NULL if the canceller could not be + * created. + */ +struct oslec_state *oslec_create(int len, int adaption_mode); + +/** + * oslec_free - Free a voice echo canceller context. + * @ec: The echo canceller context. + */ +void oslec_free(struct oslec_state *ec); + +/** + * oslec_flush - Flush (reinitialise) a voice echo canceller context. + * @ec: The echo canceller context. + */ +void oslec_flush(struct oslec_state *ec); + +/** + * oslec_adaption_mode - set the adaption mode of a voice echo canceller context. + * @ec The echo canceller context. + * @adaption_mode: The mode. + */ +void oslec_adaption_mode(struct oslec_state *ec, int adaption_mode); + +void oslec_snapshot(struct oslec_state *ec); + +/** + * oslec_update: Process a sample through a voice echo canceller. + * @ec: The echo canceller context. + * @tx: The transmitted audio sample. + * @rx: The received audio sample. + * + * The return value is the clean (echo cancelled) received sample. + */ +int16_t oslec_update(struct oslec_state *ec, int16_t tx, int16_t rx); + +/** + * oslec_hpf_tx: Process to high pass filter the tx signal. + * @ec: The echo canceller context. + * @tx: The transmitted auio sample. + * + * The return value is the HP filtered transmit sample, send this to your D/A. + */ +int16_t oslec_hpf_tx(struct oslec_state *ec, int16_t tx); + +#endif /* __OSLEC_H */ |