demod_afsk.c

//
//    This file is part of Dire Wolf, an amateur radio packet TNC.
// 
//    Copyright (C) 2011, 2012, 2013, 2014, 2015, 2020  John Langner, WB2OSZ
//
//    This program is free software: you can redistribute it and/or modify
//    it under the terms of the GNU General Public License as published by
//    the Free Software Foundation, either version 2 of the License, or
//    (at your option) any later version.
//
//    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, see <http://www.gnu.org/licenses/>.
//


// #define DEBUG1 1     /* display debugging info */

// #define DEBUG3 1	/* print carrier detect changes. */

// #define DEBUG4 1	/* capture AFSK demodulator output to log files */
			/* Can be used to make nice plots. */

// #define DEBUG5 1	// Write just demodulated bit stream to file. */


/*------------------------------------------------------------------
 *
 * Module:      demod_afsk.c
 *
 * Purpose:   	Demodulator for Audio Frequency Shift Keying (AFSK).
 *		
 * Input:	Audio samples from either a file or the "sound card."
 *
 * Outputs:	Calls hdlc_rec_bit() for each bit demodulated.  
 *
 *---------------------------------------------------------------*/

#include "direwolf.h"

#include <stdlib.h>
#include <stdio.h>
#include <math.h>
#include <unistd.h>
#include <sys/stat.h>
#include <string.h>
#include <assert.h>
#include <ctype.h>

#include "audio.h"
#include "tune.h"
#include "fsk_demod_state.h"
#include "fsk_gen_filter.h"
#include "hdlc_rec.h"
#include "textcolor.h"
#include "demod_afsk.h"
#include "dsp.h"

#define MIN(a,b) ((a)<(b)?(a):(b))
#define MAX(a,b) ((a)>(b)?(a):(b))

#define TUNE(envvar,param,name,fmt) { 				\
	char *e = getenv(envvar);				\
	if (e != NULL) {					\
	  param = atof(e);					\
	  text_color_set (DW_COLOR_ERROR);			\
	  dw_printf ("TUNE: " name " = " fmt "\n", param);	\
	} }


// Cosine table indexed by unsigned byte.
static float fcos256_table[256];

#define fcos256(x) (fcos256_table[((x)>>24)&0xff])
#define fsin256(x) (fcos256_table[(((x)>>24)-64)&0xff])

static void nudge_pll (int chan, int subchan, int slice, float demod_out, struct demodulator_state_s *D, float amplitude);


/* Quick approximation to sqrt(x*x + y*y) */
/* No benefit for regular PC. */
/* Might help with microcomputer platform??? */

__attribute__((hot)) __attribute__((always_inline))
static inline float fast_hypot(float x, float y)
{
#if 0
        x = fabsf(x);
        y = fabsf(y);

        if (x > y) {
          return (x * .941246f + y * .41f);
        }
        else {
          return (y * .941246f + x * .41f);
        }
#else
	return (hypotf(x,y));
#endif
}


/* Add sample to buffer and shift the rest down. */

__attribute__((hot)) __attribute__((always_inline))
static inline void push_sample (float val, float *buff, int size)
{
	memmove(buff+1,buff,(size-1)*sizeof(float));
	buff[0] = val; 
}


/* FIR filter kernel. */

__attribute__((hot)) __attribute__((always_inline))
static inline float convolve (const float *__restrict__ data, const float *__restrict__ filter, int filter_taps)
{
	float sum = 0.0f;
	int j;

//#pragma GCC ivdep				// ignored until gcc 4.9
	for (j=0; j<filter_taps; j++) {
	    sum += filter[j] * data[j];
	}
	return (sum);
}

// Automatic Gain control - used when we have a single slicer.
//
// The first step is to create an envelope for the peak and valley
// of the mark or space amplitude.  We need to keep track of the valley
// because it does not go down to zero when the tone is not present.
// We want to find the difference between tone present and not.
//
// We use an IIR filter with fast attack and slow decay which only considers the past.
// Perhaps an improvement could be obtained by looking in the future as well.
//

// Result should settle down to 1 unit peak to peak.  i.e. -0.5 to +0.5 


__attribute__((hot)) __attribute__((always_inline))
static inline float agc (float in, float fast_attack, float slow_decay, float *ppeak, float *pvalley)
{
	if (in >= *ppeak) {
	  *ppeak = in * fast_attack + *ppeak * (1.0f - fast_attack);
	}
	else {
	  *ppeak = in * slow_decay + *ppeak * (1.0f - slow_decay);
	}

	if (in <= *pvalley) {
	  *pvalley = in * fast_attack + *pvalley * (1.0f - fast_attack);
	}
	else  {   
	  *pvalley = in * slow_decay + *pvalley * (1.0f - slow_decay);
	}

#if 1
	float x = in;
	if (x > *ppeak) x = *ppeak;		// experiment: clip to envelope?
	if (x < *pvalley) x = *pvalley;
#endif
	if (*ppeak > *pvalley) {

	  return ((x - 0.5f * (*ppeak + *pvalley)) / (*ppeak - *pvalley)); // my original AGC

	  //return (( x - 0.5f * (*ppeak + *pvalley )) * ( *ppeak - *pvalley )); // see note below.
	  //return (x - 0.5f * (*ppeak + *pvalley));	// not as good either.
	}
	return (0.0f);
}


// K6JQ  pointed me to this wonderful article:
// Improved Automatic Threshold Correction Methods for FSK by Kok Chen, W7AY.
// http://www.w7ay.net/site/Technical/ATC/index.html
//
// The stated problem is a little different, selective fading for HF RTTY, but the
// general idea is the similar:  Compensating for imbalance of the two tones.
//
// The stronger tone probably has a better S/N ratio so we apply a larger
// weight to it.  Effectively it is comparing power rather than amplitude.
// This is the optimal method from the article referenced.
//
// Interesting idea but it did not work as well as the original AGC in this case.
// For VHF FM we are not dealing with rapid deep selective fading of one tone.
// Instead we have an imbalance which is the same for the whole frame.
// It might be interesting to try this with HF SSB packet which is much like RTTY.
//
// I use the term valley rather than noise floor.
// After a little algebra, it looks remarkably similar to the function above.
//
//	return (( x - valley ) * ( peak - valley ) - 0.5f * ( peak - valley ) * ( peak - valley ));
//	return (( x - valley ) - 0.5f * ( peak - valley )) * ( peak - valley ));
//	return (( x - 0.5f * (peak + valley )) * ( peak - valley ));



/*
 * for multi-slicer experiment.
 */

#define MIN_G 0.5f
#define MAX_G 4.0f

/* TODO: static */  float space_gain[MAX_SUBCHANS];



/*------------------------------------------------------------------
 *
 * Name:        demod_afsk_init
 *
 * Purpose:     Initialization for an AFSK demodulator.
 *		Select appropriate parameters and set up filters.
 *
 * Inputs:   	samples_per_sec
 *		baud
 *		mark_freq
 *		space_freq
 *	
 *		D		- Pointer to demodulator state for given channel.
 *
 * Outputs:	
 *
 * Returns:     None.
 *		
 * Bugs:	This doesn't do much error checking so don't give it
 *		anything crazy.
 *
 *----------------------------------------------------------------*/

void demod_afsk_init (int samples_per_sec, int baud, int mark_freq,
			int space_freq, char profile, struct demodulator_state_s *D)
{
	
	int j;

	for (j = 0; j < 256; j++) {
	  fcos256_table[j] = cosf((float)j * 2.0f * (float)M_PI / 256.0f);
	}
	
	memset (D, 0, sizeof(struct demodulator_state_s));
	D->num_slicers = 1;

#if DEBUG1
	dw_printf ("demod_afsk_init (rate=%d, baud=%d, mark=%d, space=%d, profile=%c\n",
		samples_per_sec, baud, mark_freq, space_freq, profile);
#endif
	D->profile = profile;

	switch (D->profile) {

	  case 'A':	// Official name
	  case 'E':	// For compatibility during transition

	    D->profile = 'A';

	    /* New in version 1.7 */
	    /* This is a simpler version of what has been used all along. */
	    /* Rather than convolving each sample with a pre-computed mark and */
	    /* space filter, we have two free running local oscillators.  */
	    /* Also see if we can do better with a Root Raised Cosine filter */
	    /* which supposedly reduces intersymbol interference. */

	    D->use_prefilter = 1;		/* first, a bandpass filter. */

	    if (baud > 600) {
	      D->prefilter_baud = 0.155;
						// Low cutoff below mark, high cutoff above space
						// as fraction of the symbol rate.
						// Intuitively you might expect this to be about
						// half the symbol rate, e.g. 600 Hz outside
						// the two tones of interest for 1200 baud.
						// It turns out that narrower is better.

	      D->pre_filter_len_sym = 383 * 1200. / 44100.;		// about 8 symbols
	      D->pre_window = BP_WINDOW_TRUNCATED;
	    }
	    else {
	      D->prefilter_baud = 0.87;			// TOTO: fine tune
	      D->pre_filter_len_sym = 1.857;
	      D->pre_window = BP_WINDOW_COSINE;
	    }
				    
	    // Local oscillators for Mark and Space tones.

	    D->u.afsk.m_osc_phase = 0;
	    D->u.afsk.m_osc_delta = round ( pow(2., 32.) * (double)mark_freq / (double)samples_per_sec );

	    D->u.afsk.s_osc_phase = 0;
	    D->u.afsk.s_osc_delta = round ( pow(2., 32.) * (double)space_freq / (double)samples_per_sec );

	    D->u.afsk.use_rrc = 1;
	    TUNE("TUNE_USE_RRC", D->u.afsk.use_rrc, "use_rrc", "%d")

	    if (D->u.afsk.use_rrc) {
	      D->u.afsk.rrc_width_sym = 2.80;
	      D->u.afsk.rrc_rolloff = 0.20;
	    }
	    else {
	      D->lpf_baud = 0.14;
	      D->lp_filter_width_sym = 1.388;	
	      D->lp_window = BP_WINDOW_TRUNCATED;
	    }

	    D->agc_fast_attack = 0.820;		
	    D->agc_slow_decay = 0.000214;
	    D->agc_fast_attack = 0.45;		
	    D->agc_slow_decay = 0.000195;
	    D->agc_fast_attack = 0.70;		
	    D->agc_slow_decay = 0.000090;

	    D->pll_locked_inertia = 0.74;
	    D->pll_searching_inertia = 0.50;
	    break;

	  case 'B':	// official name
	  case 'D':	// backward compatibility

	    D->profile = 'B';

	    // Experiment for version 1.7.
	    // Up to this point, I've always used separate mark and space
	    // filters and compared the amplitudes.
	    // Another technique for an FM demodulator is to mix with
	    // the center frequency and look for the rate of change of the phase.

	    D->use_prefilter = 1;		/* first, a bandpass filter. */

	    if (baud > 600) {
	      D->prefilter_baud = 0.19;
						// Low cutoff below mark, high cutoff above space
						// as fraction of the symbol rate.
						// Intuitively you might expect this to be about
						// half the symbol rate, e.g. 600 Hz outside
						// the two tones of interest for 1200 baud.
						// It turns out that narrower is better.
	
	      D->pre_filter_len_sym = 8.163;	// Filter length in symbol times.
	      D->pre_window = BP_WINDOW_TRUNCATED;
	    }
	    else {
	      D->prefilter_baud = 0.87;			// TOTO: fine tune
	      D->pre_filter_len_sym = 1.857;
	      D->pre_window = BP_WINDOW_COSINE;
	    }

	    // Local oscillator for Center frequency.

	    D->u.afsk.c_osc_phase = 0;
	    D->u.afsk.c_osc_delta = round ( pow(2., 32.) * 0.5 * (mark_freq + space_freq) / (double)samples_per_sec );

	    D->u.afsk.use_rrc = 1;
	    TUNE("TUNE_USE_RRC", D->u.afsk.use_rrc, "use_rrc", "%d")

	    if (D->u.afsk.use_rrc) {
	      D->u.afsk.rrc_width_sym = 2.00;
	      D->u.afsk.rrc_rolloff = 0.40;
	    }
	    else {
	      D->lpf_baud = 0.5;
	      D->lp_filter_width_sym = 1.714286;   //  63 * 1200. / 44100.;
	      D->lp_window = BP_WINDOW_TRUNCATED;
	    }

	    // For scaling phase shift into normallized -1 to +1 range for mark and space.
	    D->u.afsk.normalize_rpsam = 1.0 / (0.5 * abs(mark_freq - space_freq) * 2 * M_PI / samples_per_sec);

	    D->pll_locked_inertia = 0.74;
	    D->pll_searching_inertia = 0.50;

	    D->alevel_mark_peak = -1;		// FIXME:  disable display
	    D->alevel_space_peak = -1;
	    break;

	  default:

	    text_color_set(DW_COLOR_ERROR);
	    dw_printf ("Invalid AFSK demodulator profile = %c\n", profile);
	    exit (1);
	}


	TUNE("TUNE_PRE_BAUD", D->prefilter_baud, "prefilter_baud", "%.3f")
	TUNE("TUNE_PRE_WINDOW", D->pre_window, "pre_window", "%d")

	TUNE("TUNE_LPF_BAUD", D->lpf_baud, "lpf_baud", "%.3f")
	TUNE("TUNE_LP_WINDOW", D->lp_window, "lp_window", "%d")

	TUNE("TUNE_RRC_ROLLOFF", D->u.afsk.rrc_rolloff, "rrc_rolloff", "%.2f")
	TUNE("TUNE_RRC_WIDTH_SYM", D->u.afsk.rrc_width_sym, "rrc_width_sym", "%.2f")

	TUNE("TUNE_AGC_FAST", D->agc_fast_attack, "agc_fast_attack", "%.3f")
	TUNE("TUNE_AGC_SLOW", D->agc_slow_decay, "agc_slow_decay", "%.6f")
	  
	TUNE("TUNE_PLL_LOCKED", D->pll_locked_inertia, "pll_locked_inertia", "%.2f")
	TUNE("TUNE_PLL_SEARCHING", D->pll_searching_inertia, "pll_searching_inertia", "%.2f")


/*
 * Calculate constants used for timing.
 * The audio sample rate must be at least a few times the data rate.
 *
 * Baud is an integer so we hack in a fine adjustment for EAS.
 * Probably makes no difference because the DPLL keeps it in sync.
 *
 * A fraction if a Hz would make no difference for the filters.
 */
	if (baud == 521) {
	  D->pll_step_per_sample = (int) round((TICKS_PER_PLL_CYCLE * (double)520.83) / ((double)samples_per_sec));
	}
	else {
	  D->pll_step_per_sample = (int) round((TICKS_PER_PLL_CYCLE * (double)baud) / ((double)samples_per_sec));
	}

/* 
 * Optionally apply a bandpass ("pre") filter to attenuate
 * frequencies outside the range of interest.
 */

	if (D->use_prefilter) {

	  // odd number is a little better
	  D->pre_filter_taps = ((int)( D->pre_filter_len_sym * (float)samples_per_sec / (float)baud )) | 1;

	  TUNE("TUNE_PRE_FILTER_TAPS", D->pre_filter_taps, "pre_filter_taps", "%d")

// TODO:  Size comes out to 417 for 1200 bps with 48000 sample rate.
// The message is upsetting.  Can we handle this better?

	  if (D->pre_filter_taps > MAX_FILTER_SIZE) {
	    text_color_set (DW_COLOR_ERROR);
	    dw_printf ("Warning: Calculated pre filter size of %d is too large.\n", D->pre_filter_taps);
	    dw_printf ("Decrease the audio sample rate or increase the decimation factor.\n");
	    dw_printf ("You can use -D2 or -D3, on the command line, to down-sample the audio rate\n");
	    dw_printf ("before demodulating.  This greatly decreases the CPU requirements with little\n");
	    dw_printf ("impact on the decoding performance.  This is useful for a slow ARM processor,\n");
	    dw_printf ("such as with a Raspberry Pi model 1.\n");
	    D->pre_filter_taps = (MAX_FILTER_SIZE - 1) | 1;
	  }

	  float f1 = MIN(mark_freq,space_freq) - D->prefilter_baud * baud;
	  float f2 = MAX(mark_freq,space_freq) + D->prefilter_baud * baud;
#if 0
	  text_color_set(DW_COLOR_DEBUG);
	  dw_printf ("Generating prefilter %.0f to %.0f Hz.\n", f1, f2);
#endif
	  f1 = f1 / (float)samples_per_sec;
	  f2 = f2 / (float)samples_per_sec;
	  
	  gen_bandpass (f1, f2, D->pre_filter, D->pre_filter_taps, D->pre_window);
	}

/*
 * Now the lowpass filter.
 * In version 1.7 a Root Raised Cosine filter is added as an alternative
 * to the generic low pass filter.
 * In both cases, lp_filter and lp_filter_taps are used but the 
 * contents will be generated differently.  Later code does not care.
 */
	if (D->u.afsk.use_rrc) {

	  assert (D->u.afsk.rrc_width_sym >= 1 && D->u.afsk.rrc_width_sym <= 16);
	  assert (D->u.afsk.rrc_rolloff >= 0. && D->u.afsk.rrc_rolloff <= 1.);

	  D->lp_filter_taps =  ((int) (D->u.afsk.rrc_width_sym * (float)samples_per_sec / baud)) | 1;  // odd works better

	  TUNE("TUNE_LP_FILTER_TAPS", D->lp_filter_taps, "lp_filter_taps (RRC)", "%d")

	  if (D->lp_filter_taps > MAX_FILTER_SIZE) {
	    text_color_set(DW_COLOR_ERROR);
	    dw_printf ("Calculated RRC low pass filter size of %d is too large.\n", D->lp_filter_taps);
	    dw_printf ("Decrease the audio sample rate or increase the decimation factor or\n");
	    dw_printf ("recompile the application with MAX_FILTER_SIZE larger than %d.\n", MAX_FILTER_SIZE);
	    D->lp_filter_taps = (MAX_FILTER_SIZE - 1) | 1;
	  }

	  assert (D->lp_filter_taps > 8 && D->lp_filter_taps <= MAX_FILTER_SIZE);
	  (void)gen_rrc_lowpass (D->lp_filter, D->lp_filter_taps, D->u.afsk.rrc_rolloff, (float)samples_per_sec / baud);
	}
	else {
	  D->lp_filter_taps = (int) round( D->lp_filter_width_sym * (float)samples_per_sec / (float)baud );

	  TUNE("TUNE_LP_FILTER_TAPS", D->lp_filter_taps, "lp_filter_taps (FIR)", "%d")

	  if (D->lp_filter_taps > MAX_FILTER_SIZE) {
	    text_color_set (DW_COLOR_ERROR);
	    dw_printf ("Calculated FIR low pass filter size of %d is too large.\n", D->lp_filter_taps);
	    dw_printf ("Decrease the audio sample rate or increase the decimation factor or\n");
	    dw_printf ("recompile the application with MAX_FILTER_SIZE larger than %d.\n", MAX_FILTER_SIZE);
	    D->lp_filter_taps = (MAX_FILTER_SIZE - 1) | 1;
	  }

	  assert (D->lp_filter_taps > 8 && D->lp_filter_taps <= MAX_FILTER_SIZE);

	  float fc = baud * D->lpf_baud / (float)samples_per_sec;
	  gen_lowpass (fc, D->lp_filter, D->lp_filter_taps, D->lp_window);
	}


/*
 * Starting with version 1.2
 * try using multiple slicing points instead of the traditional AGC.
 */
	space_gain[0] = MIN_G;
	float step = powf(10.0, log10f(MAX_G/MIN_G) / (MAX_SUBCHANS-1));
	for (j=1; j<MAX_SUBCHANS; j++) {
	  space_gain[j] = space_gain[j-1] * step;
	}

}  /* demod_afsk_init */



/*-------------------------------------------------------------------
 *
 * Name:        demod_afsk_process_sample
 *
 * Purpose:     (1) Demodulate the AFSK signal.
 *		(2) Recover clock and data.
 *
 * Inputs:	chan	- Audio channel.  0 for left, 1 for right.
 *		subchan - modem of the channel.
 *		sam	- One sample of audio.
 *			  Should be in range of -32768 .. 32767.
 *
 * Returns:	None 
 *
 * Descripion:	First demodulate the AFSK signal.
 *
 *		A digital phase locked loop (PLL) recovers the symbol
 *		clock and picks out data bits at the proper rate.
 *
 *		For each recovered data bit, we call:
 *
 *			  hdlc_rec (channel, demodulated_bit);
 *
 *		to decode HDLC frames from the stream of bits.
 *
 * Future:	This could be generalized by passing in the name
 *		of the function to be called for each bit recovered
 *		from the demodulator.  For now, it's simply hard-coded.
 *
 * Evolution:	The simple version works less well when there is a substantial difference
 *		in amplitude of the two tones.  e.g. When de-emphasis cuts the
 *		higher tone down to about half the amplitude. We overcome that
 *		by boosting the space amplitude by varying amounts before slicing.
 *
 *		In version 1.7 an entirely different approach is added, an FM
 *		discriminator which produces a result proportional to the 
 *		frequency.
 *
 *--------------------------------------------------------------------*/
/* 
 * Which tone is stronger?
 *
 * In an ideal world, simply compare.  In my first naive attempt, that
 * worked well with perfect signals. In the real world, we don't
 * have too many perfect signals.
 *
 * Here is an excellent explanation:
 * http://www.febo.com/packet/layer-one/transmit.html
 *
 * Under real conditions, we find that the higher tone usually has a
 * considerably smaller amplitude due to the passband characteristics
 * of the transmitter and receiver.  To make matters worse, it
 * varies considerably from one station to another.
 *
 * The two filters also have different amounts of DC bias.
 *
 * My solution was to apply automatic gain control (AGC) to the mark and space 
 * levels.  This works by looking at the minimum and maximum outputs
 * for each filter and scaling the results to be roughly in the -0.5 to +0.5 range.
 * Results were excellent after tweaking the attack and decay times.
 *
 * 4X6IZ took a different approach.  See QEX Jul-Aug 2012.
 *
 * He ran two different demodulators in parallel.  One of them boosted the higher
 * frequency tone by 6 dB.  Any duplicates were removed.  This produced similar results.
 * He also used a bandpass filter before the mark/space filters.  
 * I haven't tried this combination yet for 1200 baud.
 *
 * First, let's take a look at Track 1 of the TNC test CD.  Here the receiver
 * has a flat response.  We find the mark/space strength ratios very from 0.53 to 1.38
 * with a median of 0.81.  This is in line with expectations because most
 * transmitters add pre-emphasis to boost the higher audio frequencies.
 * Track 2 should more closely resemble what comes out of the speaker on a typical
 * transceiver.  Here we see a ratio from 1.73 to 3.81 with a median of 2.48.
 * 
 * This is similar to my observations of local signals, from the speaker.
 * The amplitude ratio varies from 1.48 to 3.41 with a median of 2.70. 
 */



__attribute__((hot))
void demod_afsk_process_sample (int chan, int subchan, int sam, struct demodulator_state_s *D)
{
#if DEBUG4
	static FILE *demod_log_fp = NULL;
	static int seq = 0;			/* for log file name */
#endif

	assert (chan >= 0 && chan < MAX_CHANS);
	assert (subchan >= 0 && subchan < MAX_SUBCHANS);

/* 
 * Filters use last 'filter_taps' samples.
 *
 * First push the older samples down. 
 *
 * Finally, put the most recent at the beginning.
 *
 * Future project?  Can we do better than shifting each time?
 */

	/* Scale to nice number. */

	float fsam = (float)sam / 16384.0f;

	switch (D->profile) {

	  case 'E':
	  default:
	  case 'A': {
				/* ========== New in Version 1.7 ========== */

				//	Cleaner & simpler than earlier 'A' thru 'E'

	    if (D->use_prefilter) {
	      push_sample (fsam, D->raw_cb, D->pre_filter_taps);
	      fsam = convolve (D->raw_cb, D->pre_filter, D->pre_filter_taps);
	    }

	    push_sample (fsam * fcos256(D->u.afsk.m_osc_phase), D->u.afsk.m_I_raw, D->lp_filter_taps);
	    push_sample (fsam * fsin256(D->u.afsk.m_osc_phase), D->u.afsk.m_Q_raw, D->lp_filter_taps);
	    D->u.afsk.m_osc_phase += D->u.afsk.m_osc_delta;

	    push_sample (fsam * fcos256(D->u.afsk.s_osc_phase), D->u.afsk.s_I_raw, D->lp_filter_taps);
	    push_sample (fsam * fsin256(D->u.afsk.s_osc_phase), D->u.afsk.s_Q_raw, D->lp_filter_taps);
	    D->u.afsk.s_osc_phase += D->u.afsk.s_osc_delta;

	    float m_I = convolve (D->u.afsk.m_I_raw, D->lp_filter, D->lp_filter_taps);
	    float m_Q = convolve (D->u.afsk.m_Q_raw, D->lp_filter, D->lp_filter_taps);
	    float m_amp = fast_hypot(m_I, m_Q);

	    float s_I = convolve (D->u.afsk.s_I_raw, D->lp_filter, D->lp_filter_taps);
	    float s_Q = convolve (D->u.afsk.s_Q_raw, D->lp_filter, D->lp_filter_taps);
	    float s_amp = fast_hypot(s_I, s_Q);

/*
 * Capture the mark and space peak amplitudes for display.
 * It uses fast attack and slow decay to get an idea of the
 * overall amplitude.
 */
	    if (m_amp >= D->alevel_mark_peak) {
	      D->alevel_mark_peak = m_amp * D->quick_attack + D->alevel_mark_peak * (1.0f - D->quick_attack);
	    }
	    else {
	      D->alevel_mark_peak = m_amp * D->sluggish_decay + D->alevel_mark_peak * (1.0f - D->sluggish_decay);
	    }

	    if (s_amp >= D->alevel_space_peak) {
	      D->alevel_space_peak = s_amp * D->quick_attack + D->alevel_space_peak * (1.0f - D->quick_attack);
	    }
	      else {
	      D->alevel_space_peak = s_amp * D->sluggish_decay + D->alevel_space_peak * (1.0f - D->sluggish_decay);
	    }

	    if (D->num_slicers <= 1) {

	      // Which tone is stonger?  That's simple with an ideal signal.
	      // However, we don't see too many ideal signals.
	      // Due to mismatching pre-emphasis and de-emphasis, the two
	      // tones will often have greatly different amplitudes so we use
	      // automatic gain control (AGC) to scale each to the same range
	      // before comparing.
	      // This is probably over complicated and could be combined with
	      // the signal amplitude measurement, above.
	      // It works so let's move along to other topics.

	      float m_norm = agc (m_amp, D->agc_fast_attack, D->agc_slow_decay, &(D->m_peak), &(D->m_valley));
	      float s_norm = agc (s_amp, D->agc_fast_attack, D->agc_slow_decay, &(D->s_peak), &(D->s_valley));

	      // The normalized values should be around -0.5 to +0.5 so the difference
	      // should work out to be around -1 to +1.
	      // This is important because nudge_pll uses the demod_out amplitude to assign
	      // a quality or confidence score to the symbol.

	      float demod_out = m_norm - s_norm;

	      // Tested and it looks good.  Range of about -1 to +1.
	      //printf ("JWL DEBUG demod A with agc = %6.2f\n", demod_out);

	      nudge_pll (chan, subchan, 0, demod_out, D, 1.0);

	    }
	    else {
	      // Multiple slice case.
	      // Rather than trying to find the best threshold location, use multiple 
	      // slicer thresholds in parallel.
	      // The best slicing point will vary from packet to packet but should
	      // remain about the same for a given packet.

	      // We are not performing the AGC step here but still want the envelope
	      // for caluculating the confidence level (or quality) of the sample.

	      (void) agc (m_amp, D->agc_fast_attack, D->agc_slow_decay, &(D->m_peak), &(D->m_valley));
	      (void) agc (s_amp, D->agc_fast_attack, D->agc_slow_decay, &(D->s_peak), &(D->s_valley));

	      for (int slice=0; slice<D->num_slicers; slice++) {
	        float demod_out = m_amp - s_amp * space_gain[slice];
	        float amp = 0.5f * (D->m_peak - D->m_valley + (D->s_peak - D->s_valley) * space_gain[slice]);
	        if (amp < 0.0000001f) amp = 1;	// avoid divide by zero with no signal.

	        // Tested and it looks good.  Range of about -1 to +1 relative to amp.
		// Biased one way or the other depending on the space gain.
	        //printf ("JWL DEBUG demod A with slicer %d: %6.2f / %6.2f = %6.2f\n", slice, demod_out, amp, demod_out/amp);

	        nudge_pll (chan, subchan, slice, demod_out, D, amp);
	      }
	    }
	  }
	  break;

	case 'D':
	case 'B': {
				/* ========== Version 1.7 Experiment ========== */

				// New - Convert frequency to a value proportional to frequency.

	  if (D->use_prefilter) {
	    push_sample (fsam, D->raw_cb, D->pre_filter_taps);
	    fsam = convolve (D->raw_cb, D->pre_filter, D->pre_filter_taps);
	  }

	  push_sample (fsam * fcos256(D->u.afsk.c_osc_phase), D->u.afsk.c_I_raw, D->lp_filter_taps);
	  push_sample (fsam * fsin256(D->u.afsk.c_osc_phase), D->u.afsk.c_Q_raw, D->lp_filter_taps);
	  D->u.afsk.c_osc_phase += D->u.afsk.c_osc_delta;

	  float c_I = convolve (D->u.afsk.c_I_raw, D->lp_filter, D->lp_filter_taps);
	  float c_Q = convolve (D->u.afsk.c_Q_raw, D->lp_filter, D->lp_filter_taps);

	  float phase = atan2f (c_Q, c_I);
	  float rate = phase - D->u.afsk.prev_phase; 
	  if (rate > M_PI) rate -= 2 * M_PI;
	  else if (rate < -M_PI) rate += 2 * M_PI;
	  D->u.afsk.prev_phase = phase;

	  // Rate is radians per audio sample interval or something like that.
	  // Scale scale that into -1 to +1 for expected tones.

	  float norm_rate = rate * D->u.afsk.normalize_rpsam;

	  // We really don't have mark and space amplitudes available in this case.

	  if (D->num_slicers <= 1) {

	    float demod_out = norm_rate;
	    // Tested and it looks good.  Range roughly -1 to +1.
	    //printf ("JWL DEBUG demod B single = %6.2f\n", demod_out);

	    nudge_pll (chan, subchan, 0, demod_out, D, 1.0);

	  }
	  else {

	    // This would be useful for HF SSB where a tuning error
	    // would shift the frequency.  Multiple slicing points would
	    // then compensate for differences in transmit/receive frequencies.
	    // 
	    // Where should we set the thresholds?
	    // I'm thinking something like:
	    // 	-.5	-.375	-.25	-.125	0	.125	.25	.375	.5
	    //
	    // Assuming a 300 Hz shift, this would put slicing thresholds up
	    // to +-75 Hz from the center.

	    for (int slice=0; slice<D->num_slicers; slice++) {

	      float offset = -0.5 + slice * (1. / (D->num_slicers - 1));
	      float demod_out = norm_rate + offset;

	      //printf ("JWL DEBUG demod B slice %d, offset = %6.3f, demod_out = %6.2f\n", slice, offset, demod_out);

	      nudge_pll (chan, subchan, slice, demod_out, D, 1.0);
	    }
	  }
	  }
	  break;
	}
		
#if DEBUG4

	if (chan == 0) {
	if (D->slicer[slice].data_detect) {
	  char fname[30];

	  
	  if (demod_log_fp == NULL) {
	    seq++;
	    snprintf (fname, sizeof(fname), "demod/%04d.csv", seq);
	    if (seq == 1) mkdir ("demod", 0777);

	    demod_log_fp = fopen (fname, "w");
	    text_color_set(DW_COLOR_DEBUG);
	    dw_printf ("Starting demodulator log file %s\n", fname);
	    fprintf (demod_log_fp, "Audio, Mark, Space, Demod, Data, Clock\n");
	  }
	  fprintf (demod_log_fp, "%.3f, %.3f, %.3f, %.3f, %.2f, %.2f\n", fsam + 3.5, m_norm + 2, s_norm + 2, 
			(m_norm - s_norm) / 2 + 1.5,
			demod_data ? .9 : .55,  
			(D->data_clock_pll & 0x80000000) ? .1 : .45);
	}
	else {
	  if (demod_log_fp != NULL) {
	    fclose (demod_log_fp);
	    demod_log_fp = NULL;
	  }
	}
	}

#endif


} /* end demod_afsk_process_sample */



/*
 * Finally, a PLL is used to sample near the centers of the data bits.
 *
 * D points to a demodulator for a channel/subchannel pair so we don't
 * have to keep recalculating it.
 *
 * D->data_clock_pll is a SIGNED 32 bit variable.
 * When it overflows from a large positive value to a negative value, we 
 * sample a data bit from the demodulated signal.
 *
 * Ideally, the the demodulated signal transitions should be near
 * zero we we sample mid way between the transitions.
 *
 * Nudge the PLL by removing some small fraction from the value of 
 * data_clock_pll, pushing it closer to zero.
 * 
 * This adjustment will never change the sign so it won't cause
 * any erratic data bit sampling.
 *
 * If we adjust it too quickly, the clock will have too much jitter.
 * If we adjust it too slowly, it will take too long to lock on to a new signal.
 *
 * Be a little more aggressive about adjusting the PLL
 * phase when searching for a signal.  Don't change it as much when
 * locked on to a signal.
 *
 * I don't think the optimal value will depend on the audio sample rate
 * because this happens for each transition from the demodulator.
 */

__attribute__((hot))
static void nudge_pll (int chan, int subchan, int slice, float demod_out, struct demodulator_state_s *D, float amplitude)
{
	D->slicer[slice].prev_d_c_pll = D->slicer[slice].data_clock_pll;

	// Perform the add as unsigned to avoid signed overflow error.
	D->slicer[slice].data_clock_pll = (signed)((unsigned)(D->slicer[slice].data_clock_pll) + (unsigned)(D->pll_step_per_sample));

	  //text_color_set(DW_COLOR_DEBUG);
	  // dw_printf ("prev = %lx, new data clock pll = %lx\n" D->prev_d_c_pll, D->data_clock_pll);

	if (D->slicer[slice].data_clock_pll < 0 && D->slicer[slice].prev_d_c_pll > 0) {

	  /* Overflow - this is where we sample. */
	  // Assign it a confidence level or quality, 0 to 100, based on the amplitude.
	  // Those very close to 0 are suspect.  We'll get back to this later.

	  int quality = fabsf(demod_out) * 100.0f / amplitude;
	  if (quality > 100) quality = 100;

#if DEBUG5
	  // Write bit stream to a file.

	  static FILE *bsfp = NULL;
	  static int bcount = 0;
	  if (chan == 0 && subchan == 0 && slice == 0) {
	    if (bsfp == NULL) {
	       bsfp = fopen ("bitstream.txt", "w");
	    }
	    fprintf (bsfp, "%d", demod_out > 0);
	    bcount++;
	    if (bcount % 64 == 0) {
	      fprintf (bsfp, "\n");
	    }
	  }

#endif

	  hdlc_rec_bit (chan, subchan, slice, demod_out > 0, 0, quality);
	  pll_dcd_each_symbol2 (D, chan, subchan, slice);
	}

	// Transitions nudge the DPLL phase toward the incoming signal.

	int demod_data = demod_out > 0;
        if (demod_data != D->slicer[slice].prev_demod_data) {

	  pll_dcd_signal_transition2 (D, slice, D->slicer[slice].data_clock_pll);

	  if (D->slicer[slice].data_detect) {
	    D->slicer[slice].data_clock_pll = (int)(D->slicer[slice].data_clock_pll * D->pll_locked_inertia);
	  }
	  else {
	    D->slicer[slice].data_clock_pll = (int)(D->slicer[slice].data_clock_pll * D->pll_searching_inertia);
	  }
	}

/*
 * Remember demodulator output so we can compare next time.
 */
	D->slicer[slice].prev_demod_data = demod_data;

} /* end nudge_pll */


/* end demod_afsk.c */