#include <stdio.h>
#ifndef NO_LPC_FIX
#include <stdlib.h>
#endif
#include <math.h>
#include <sys/types.h>
#include <netinet/in.h>

#include "lpc.h"
#include "../ulaw2linear.h"

#if defined(Solaris) || defined(HEWLETT_PACKARD) || defined(__FreeBSD__)
#define bcopy(a, b, n)	  memmove(b, a, n)
#endif

#define MAXWINDOW	1000	/* Max analysis window length */
#define FS		8000.0	/* Sampling rate */

#define DOWN		5	/* Decimation for pitch analyzer */
#define PITCHORDER	4	/* Model order for pitch analyzer */
#define FC		600.0	/* Pitch analyzer filter cutoff */
#define MINPIT		50.0	/* Minimum pitch */
#define MAXPIT		300.0	/* Maximum pitch */

#define MINPER		(int)(FS/(DOWN*MAXPIT)+.5)	/* Minimum period */
#define MAXPER		(int)(FS/(DOWN*MINPIT)+.5)	/* Maximum period */

#define WSCALE		1.5863	/* Energy loss due to windowing */

#define BUFLEN		((FRAMESIZE * 3) / 2)

#define SILENCEFIX		/* Enable absolute silence fix */

/*  The LPC coder does truly awful things when driven into clipping.
    If you set GAIN_ADJUST to a number less than 1.0, samples will
    be scaled by that factor to avoid overdriving the coder.  */

#define GAIN_ADJUST 0.9

static struct lpcwork {
    float w_s[MAXWINDOW], w_y[MAXWINDOW], w_h[MAXWINDOW], w_w[MAXWINDOW];
} *work = NULL;

static float fa[6], u, u1, yp1, yp2;

#define lin_to_ulaw(x) audio_s2u(x)
#define ulaw_to_lin(x) audio_u2s(x)

static void
auto_correl(w, n, p, r)
  float *w;
  int n;
  int p;
  float *r;
{
	int i, k, nk;

	for (k = 0; k <= p; k++) {
		nk = n - k;
		r[k] = 0.0;
		for (i = 0; i < nk; i++)
			r[k] += w[i] * w[i + k];
	}
}

static void
durbin(r, p, k, g)
  float *r;
  int p;
  float *k;
  float *g;
{
	int i, j;
	float a[LPC_FILTORDER + 1], at[LPC_FILTORDER + 1], e;

	for (i = 0; i <= p; i++)
		a[i] = at[i] = 0.0;

	e = r[0];
	for (i = 1; i <= p; i++) {
		k[i] = -r[i];
		for (j = 1; j < i; j++) {
			at[j] = a[j];
			k[i] -= a[j] * r[i - j];
		}
#ifdef SILENCEFIX
if (e == 0) {
    *g = 0;
    return;
}
#endif
		k[i] /= e;
		a[i] = k[i];
		for (j = 1; j < i; j++)
			a[j] = at[j] + k[i] * at[i - j];
		e *= 1.0 - k[i] * k[i];
	}

#ifdef SILENCEFIX
if (e < 0) {
    e = 0;
}
#endif
	*g = sqrt(e);
}

static void
inverse_filter(w, k)
  float *w;
  float *k;
{
	int i, j;
	float b[PITCHORDER + 1], bp[PITCHORDER + 1], f[PITCHORDER + 1];

	for (i = 0; i <= PITCHORDER; i++)
		b[i] = f[i] = bp[i] = 0.0;

	for (i = 0; i < BUFLEN / DOWN; i++) {
		f[0] = b[0] = w[i];
		for (j = 1; j <= PITCHORDER; j++) {
			f[j] = f[j - 1] + k[j] * bp[j - 1];
			b[j] = k[j] * f[j - 1] + bp[j - 1];
			bp[j - 1] = b[j - 1];
		}
		w[i] = f[PITCHORDER];
	}
}

static void
calc_pitch(w, per)
  float *w;
  float *per;
{
	int i, j, rpos;
	float d[MAXWINDOW / DOWN], k[PITCHORDER + 1], r[MAXPER + 1], g, rmax;
	float rval, rm, rp;
	float a, b, c, x, y;
	static int vuv = 0;

#ifdef NO_LPC_FIX
	/* Old decimation sometimes fails to recognise voiced. */
	for (i = 0, j = 0; i < BUFLEN; i += DOWN)
	    d[j++] = w[i];
#else
	/* New: average rather than decimating. */
	for (i = 0, j = 0; i < BUFLEN; j++) {
	    d[j] = 0;
	    for (rpos = 0; rpos < DOWN; rpos++) {
		d[j] += w[i++];
	    }
	    /* d[j] /- DOWN;	   Not actually necessary. */	   
	}
#endif
	auto_correl(d, BUFLEN / DOWN, PITCHORDER, r);
	durbin(r, PITCHORDER, k, &g);
	inverse_filter(d, k);
#ifdef NO_LPC_FIX
	auto_correl(d, BUFLEN / DOWN, MAXPER + 1, r);
#else
	auto_correl(d, BUFLEN / DOWN, MAXPER, r);
#endif
	rpos = 0;
	rmax = 0.0;
	for (i = MINPER; i <= MAXPER; i++) {
		if (r[i] > rmax) {
			rmax = r[i];
			rpos = i;
		}
	}

	rm = r[rpos - 1];
	rp = r[rpos + 1];
#ifdef OLDWAY
	rval = rmax / r[0];
#endif

	a = 0.5 * rm - rmax + 0.5 * rp;
	b = -0.5 * rm * (2.0 * rpos + 1.0) + 
	    2.0 * rpos * rmax + 0.5 * rp * (1.0 - 2.0 * rpos);
	c = 0.5 * rm * (rpos * rpos + rpos) +
	    rmax * (1.0 - rpos * rpos) + 0.5 * rp * (rpos * rpos - rpos);

	x = -b / (2.0 * a);
	y = a * x * x + b * x + c;
	x *= DOWN;

	rmax = y;
#ifdef OLDWAY
	rval = rmax / r[0];
#else
	if (r[0] == 0.0) {
	    rval = 1.0;
	} else {
	    rval = rmax / r[0];
	}
#endif
#ifdef NO_LPC_FIX
	if (rval >= 0.4 || (vuv == 3 && rval >= 0.3)) {
#else
	if ((rval >= 0.4 || (vuv == 3 && rval >= 0.3)) && (x > 0)) {
#endif
		*per = x;
		vuv = (vuv & 1) * 2 + 1;
	} else {
		*per = 0.0;
		vuv = (vuv & 1) * 2;
	}
}

#define s   work->w_s
#define y   work->w_y
#define h   work->w_h

void
lpc_init(state)
  lpcstate_t* state;
{
	int	i;
	float	r, v, w, wcT;

#ifdef HEXDUMP
        /*  Let's make sure the compiler hasn't inserted
	    padding in the lpcparams_t structure which will
	    make it incompatible with other machines.  */

	{
	    lpcparams_t *p = 0;

	    if ((&p->gain != ((unsigned char *) 2)) ||
		(p->k != ((signed char *) 4))) {
                fprintf(stderr, "Alignment problem in lpcparams.h structure.\n");
                fprintf(stderr, "Add definitions or compiler options in lpc directory\n");
                fprintf(stderr, "to guarantee this structure is packed.\n");
	    }
	}
#endif

	for (i = 0; i < BUFLEN; i++) {
		s[i] = 0.0;
		h[i] = WSCALE * (0.54 - 0.46 *
		       cos(2 * M_PI * i / (BUFLEN - 1.0)));
	}
	wcT = 2 * M_PI * FC / FS;
	r = 0.36891079 * wcT;
	v = 0.18445539 * wcT;
	w = 0.92307712 * wcT;
	fa[1] = -exp(-r);
	fa[2] = 1.0 + fa[1];
	fa[3] = -2.0 * exp(-v) * cos(w);
	fa[4] = exp(-2.0 * v);
	fa[5] = 1.0 + fa[3] + fa[4];

	u1 = 0.0;
	yp1 = 0.0;
	yp2 = 0.0;

	state->Oldper = 0.0;
	state->OldG = 0.0;
	for (i = 0; i <= LPC_FILTORDER; i++) {
		state->Oldk[i] = 0.0;
		state->bp[i] = 0.0;
	}
	state->pitchctr = 0;
}

void
lpc_analyze(buf, params)
  const unsigned char *buf;
  lpcparams_t *params;
{
	int	i, j;
#define w   work->w_w
	float	r[LPC_FILTORDER + 1];
	float	per, G, k[LPC_FILTORDER + 1];

	for (i = 0, j = BUFLEN - FRAMESIZE; i < FRAMESIZE; i++, j++) {
		s[j] = (float) (GAIN_ADJUST * ((ulaw_to_lin(*buf++)) / 32768.));
		u = fa[2] * s[j] - fa[1] * u1;
		y[j] = fa[5] * u1 - fa[3] * yp1 - fa[4] * yp2;
		u1 = u;
		yp2 = yp1;
		yp1 = y[j];
	}

	calc_pitch(y, &per);

	for (i = 0; i < BUFLEN; i++)
		w[i] = s[i] * h[i];
	auto_correl(w, BUFLEN, LPC_FILTORDER, r);
	durbin(r, LPC_FILTORDER, k, &G);

	params->period = (unsigned short) htons((unsigned short) (per * (1 << 8)));
#ifdef NO_LPC_FIX
	params->gain = G * (1 << 8);
#else
	i = (int) (G * (1 << 8));
	if (i > 255) {
	    i = 255;
	}
	params->gain = i;
#endif
	for (i = 0; i < LPC_FILTORDER; i++) {
#ifdef NO_LPC_FIX
	    params->k[i] = k[i + 1] * (1 << 7);
#else
	    float u = k[i + 1];

	    if (u < -0.9999) {
		u = -0.9999f;
	    } else if (u > 0.9999) {
		u = 0.9999f;
	    }
	    params->k[i] = (signed char) (0.5 + (127.0 * k[i + 1]));
#endif
	}

	bcopy(s + FRAMESIZE, s, (BUFLEN - FRAMESIZE) * sizeof(s[0]));
	bcopy(y + FRAMESIZE, y, (BUFLEN - FRAMESIZE) * sizeof(y[0]));
#undef w
}

void
lpc_synthesize(buf, params, state)
  unsigned char *buf;
  lpcparams_t *params;
  lpcstate_t* state;
{
	int i, j;
	register double u, f, per, G, NewG, Ginc, Newper, perinc;
	double k[LPC_FILTORDER + 1], Newk[LPC_FILTORDER + 1],
	       kinc[LPC_FILTORDER + 1];

	per = (double) ((unsigned short) ntohs(params->period)) / 256.;
	G = (double) params->gain / 256.;
	k[0] = 0.0;
	for (i = 0; i < LPC_FILTORDER; i++)
		k[i + 1] = (double) (params->k[i]) / 128.;

	G /= sqrt(BUFLEN / (per == 0.0? 3.0 : per));
	Newper = state->Oldper;
	NewG = state->OldG;
	for (i = 1; i <= LPC_FILTORDER; i++)
		Newk[i] = state->Oldk[i];

	if (state->Oldper != 0 && per != 0) {
		perinc = (per - state->Oldper) / (double)FRAMESIZE;
		Ginc = (G - state->OldG) / (double)FRAMESIZE;
		for (i = 1; i <= LPC_FILTORDER; i++)
			kinc[i] = (k[i] - state->Oldk[i]) / (double)FRAMESIZE;
	} else {
		perinc = 0.0;
		Ginc = 0.0;
		for (i = 1; i <= LPC_FILTORDER; i++)
			kinc[i] = 0.0;
	}

	if (Newper == 0)
		state->pitchctr = 0;

	for (i = 0; i < FRAMESIZE; i++) {
		if (Newper == 0) {
#ifdef NO_LPC_FIX
			u = ((double)random() / 2147483648.0) * NewG;
#else
			u = ((rand() / (1.0 + RAND_MAX)) - 0.5) * 1.5874 * NewG;
#endif
		} else {
			if (state->pitchctr == 0) {
				u = NewG;
				state->pitchctr = (int) Newper;
			} else {
				u = 0.0;
				state->pitchctr--;
			}
		}

		f = u;
		for (j = LPC_FILTORDER; j >= 1; j--) {
			register double b = state->bp[j - 1];
			register double kj = Newk[j];
			Newk[j] = kj + kinc[j];
			f -= b * kj;
			b += f * kj;
			state->bp[j] = b;
		}
		state->bp[0] = f;

#ifdef NO_LPC_FIX
		*buf++ = lin_to_ulaw((int)(f * 32768.0) & 0xffff);
#else
		u = f;
		if (u < -0.9999) {
		    u = -0.9999;
		} else if (u > 0.9999) {
		    u = 0.9999;
		}
		*buf++ = lin_to_ulaw((int) (u * 32767.0));
#endif

		Newper += perinc;
		NewG += Ginc;
	}

	state->Oldper = per;
	state->OldG = G;
	for (i = 1; i <= LPC_FILTORDER; i++)
		state->Oldk[i] = k[i];
}

/*  LPC_START  --  Allocate working storage for LPC coder.  */

int lpc_start()
{
    if (work == NULL) {
	work = (struct lpcwork *) malloc(sizeof(struct lpcwork));
    }
    return work != NULL;
}

/*  LPC_END  --  Release working storage for LPC coder.  */

void lpc_end()
{
    if (work != NULL) {
	free(work);
	work = NULL;
    }
}