annotate libaf/af_hrtf.c @ 34309:119af6360b00

Discard frames where the size does not match the AVCodecContext width/height. This avoids possible crashes on video size changes. The problem is that we reinitialize the vo on get_buffer but due to codec delay libavcodec might still return frames with the old size afterwards, which the vo might no longer be able to handle. Ideally libavcodec should not show this behaviour, since it requires that any application using DR1 can handle frames of different sizes simultaneously - which seems a bit extreme.

/*
 * Experimental audio filter that mixes 5.1 and 5.1 with matrix
 * encoded rear channels into headphone signal using FIR filtering
 * with HRTF.
 *
 * This file is part of MPlayer.
 *
 * MPlayer 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.
 *
 * MPlayer 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 MPlayer; if not, write to the Free Software Foundation, Inc.,
 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
 */

//#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <inttypes.h>

#include <math.h>

#include "mp_msg.h"
#include "af.h"
#include "dsp.h"

/* HRTF filter coefficients and adjustable parameters */
#include "af_hrtf.h"

typedef struct af_hrtf_s {
    /* Lengths */
    int dlbuflen, hrflen, basslen;
    /* L, C, R, Ls, Rs channels */
    float *lf, *rf, *lr, *rr, *cf, *cr;
    const float *cf_ir, *af_ir, *of_ir, *ar_ir, *or_ir, *cr_ir;
    int cf_o, af_o, of_o, ar_o, or_o, cr_o;
    /* Bass */
    float *ba_l, *ba_r;
    float *ba_ir;
    /* Whether to matrix decode the rear center channel */
    int matrix_mode;
    /* How to decode the input:
       0 = 5/5+1 channels
       1 = 2 channels
       2 = matrix encoded 2 channels */
    int decode_mode;
    /* Full wave rectified (FWR) amplitudes and gain used to steer the
       active matrix decoding of front channels (variable names
       lpr/lmr means Lt + Rt, Lt - Rt) */
    float l_fwr, r_fwr, lpr_fwr, lmr_fwr;
    float adapt_l_gain, adapt_r_gain, adapt_lpr_gain, adapt_lmr_gain;
    /* Matrix input decoding require special FWR buffer, since the
       decoding is done in place. */
    float *fwrbuf_l, *fwrbuf_r, *fwrbuf_lr, *fwrbuf_rr;
    /* Rear channel delay buffer for matrix decoding */
    float *rear_dlbuf;
    /* Full wave rectified amplitude and gain used to steer the active
       matrix decoding of center rear channel */
    float lr_fwr, rr_fwr, lrprr_fwr, lrmrr_fwr;
    float adapt_lr_gain, adapt_rr_gain;
    float adapt_lrprr_gain, adapt_lrmrr_gain;
    /* Cyclic position on the ring buffer */
    int cyc_pos;
    int print_flag;
} af_hrtf_t;

/* Convolution on a ring buffer
 *    nx:	length of the ring buffer
 *    nk:	length of the convolution kernel
 *    sx:	ring buffer
 *    sk:	convolution kernel
 *    offset:	offset on the ring buffer, can be
 */
static float conv(const int nx, const int nk, const float *sx, const float *sk,
		  const int offset)
{
    /* k = reminder of offset / nx */
    int k = offset >= 0 ? offset % nx : nx + (offset % nx);

    if(nk + k <= nx)
	return af_filter_fir(nk, sx + k, sk);
    else
	return af_filter_fir(nk + k - nx, sx, sk + nx - k) +
	    af_filter_fir(nx - k, sx + k, sk);
}

/* Detect when the impulse response starts (significantly) */
static int pulse_detect(const float *sx)
{
    /* nmax must be the reference impulse response length (128) minus
       s->hrflen */
    const int nmax = 128 - HRTFFILTLEN;
    const float thresh = IRTHRESH;
    int i;

    for(i = 0; i < nmax; i++)
	if(fabs(sx[i]) > thresh)
	    return i;
    return 0;
}

/* Fuzzy matrix coefficient transfer function to "lock" the matrix on
   a effectively passive mode if the gain is approximately 1 */
static inline float passive_lock(float x)
{
   const float x1 = x - 1;
   const float ax1s = fabs(x - 1) * (1.0 / MATAGCLOCK);

   return x1 - x1 / (1 + ax1s * ax1s) + 1;
}

/* Unified active matrix decoder for 2 channel matrix encoded surround
   sources */
static inline void matrix_decode(short *in, const int k, const int il,
			  const int ir, const int decode_rear,
			  const int dlbuflen,
			  float l_fwr, float r_fwr,
			  float lpr_fwr, float lmr_fwr,
			  float *adapt_l_gain, float *adapt_r_gain,
			  float *adapt_lpr_gain, float *adapt_lmr_gain,
			  float *lf, float *rf, float *lr,
			  float *rr, float *cf)
{
   const int kr = (k + MATREARDELAY) % dlbuflen;
   float l_gain = (l_fwr + r_fwr) /
      (1 + l_fwr + l_fwr);
   float r_gain = (l_fwr + r_fwr) /
      (1 + r_fwr + r_fwr);
   /* The 2nd axis has strong gain fluctuations, and therefore require
      limits.  The factor corresponds to the 1 / amplification of (Lt
      - Rt) when (Lt, Rt) is strongly correlated. (e.g. during
      dialogues).  It should be bigger than -12 dB to prevent
      distortion. */
   float lmr_lim_fwr = lmr_fwr > M9_03DB * lpr_fwr ?
      lmr_fwr : M9_03DB * lpr_fwr;
   float lpr_gain = (lpr_fwr + lmr_lim_fwr) /
      (1 + lpr_fwr + lpr_fwr);
   float lmr_gain = (lpr_fwr + lmr_lim_fwr) /
      (1 + lmr_lim_fwr + lmr_lim_fwr);
   float lmr_unlim_gain = (lpr_fwr + lmr_fwr) /
      (1 + lmr_fwr + lmr_fwr);
   float lpr, lmr;
   float l_agc, r_agc, lpr_agc, lmr_agc;
   float f, d_gain, c_gain, c_agc_cfk;

#if 0
   static int counter = 0;
   static FILE *fp_out;

   if(counter == 0)
      fp_out = fopen("af_hrtf.log", "w");
   if(counter % 240 == 0)
      fprintf(fp_out, "%g %g %g %g %g ", counter * (1.0 / 48000),
	      l_gain, r_gain, lpr_gain, lmr_gain);
#endif

   /*** AXIS NO. 1: (Lt, Rt) -> (C, Ls, Rs) ***/
   /* AGC adaption */
   d_gain = (fabs(l_gain - *adapt_l_gain) +
	     fabs(r_gain - *adapt_r_gain)) * 0.5;
   f = d_gain * (1.0 / MATAGCTRIG);
   f = MATAGCDECAY - MATAGCDECAY / (1 + f * f);
   *adapt_l_gain = (1 - f) * *adapt_l_gain + f * l_gain;
   *adapt_r_gain = (1 - f) * *adapt_r_gain + f * r_gain;
   /* Matrix */
   l_agc = in[il] * passive_lock(*adapt_l_gain);
   r_agc = in[ir] * passive_lock(*adapt_r_gain);
   cf[k] = (l_agc + r_agc) * M_SQRT1_2;
   if(decode_rear) {
      lr[kr] = rr[kr] = (l_agc - r_agc) * M_SQRT1_2;
      /* Stereo rear channel is steered with the same AGC steering as
	 the decoding matrix. Note this requires a fast updating AGC
	 at the order of 20 ms (which is the case here). */
      lr[kr] *= (l_fwr + l_fwr) /
	 (1 + l_fwr + r_fwr);
      rr[kr] *= (r_fwr + r_fwr) /
	 (1 + l_fwr + r_fwr);
   }

   /*** AXIS NO. 2: (Lt + Rt, Lt - Rt) -> (L, R) ***/
   lpr = (in[il] + in[ir]) * M_SQRT1_2;
   lmr = (in[il] - in[ir]) * M_SQRT1_2;
   /* AGC adaption */
   d_gain = fabs(lmr_unlim_gain - *adapt_lmr_gain);
   f = d_gain * (1.0 / MATAGCTRIG);
   f = MATAGCDECAY - MATAGCDECAY / (1 + f * f);
   *adapt_lpr_gain = (1 - f) * *adapt_lpr_gain + f * lpr_gain;
   *adapt_lmr_gain = (1 - f) * *adapt_lmr_gain + f * lmr_gain;
   /* Matrix */
   lpr_agc = lpr * passive_lock(*adapt_lpr_gain);
   lmr_agc = lmr * passive_lock(*adapt_lmr_gain);
   lf[k] = (lpr_agc + lmr_agc) * M_SQRT1_2;
   rf[k] = (lpr_agc - lmr_agc) * M_SQRT1_2;

   /*** CENTER FRONT CANCELLATION ***/
   /* A heuristic approach exploits that Lt + Rt gain contains the
      information about Lt, Rt correlation.  This effectively reshapes
      the front and rear "cones" to concentrate Lt + Rt to C and
      introduce Lt - Rt in L, R. */
   /* 0.67677 is the emprical lower bound for lpr_gain. */
   c_gain = 8 * (*adapt_lpr_gain - 0.67677);
   c_gain = c_gain > 0 ? c_gain : 0;
   /* c_gain should not be too high, not even reaching full
      cancellation (~ 0.50 - 0.55 at current AGC implementation), or
      the center will s0und too narrow. */
   c_gain = MATCOMPGAIN / (1 + c_gain * c_gain);
   c_agc_cfk = c_gain * cf[k];
   lf[k] -= c_agc_cfk;
   rf[k] -= c_agc_cfk;
   cf[k] += c_agc_cfk + c_agc_cfk;
#if 0
   if(counter % 240 == 0)
      fprintf(fp_out, "%g %g %g %g %g\n",
	      *adapt_l_gain, *adapt_r_gain,
	      *adapt_lpr_gain, *adapt_lmr_gain,
	      c_gain);
   counter++;
#endif
}

static inline void update_ch(af_hrtf_t *s, short *in, const int k)
{
    const int fwr_pos = (k + FWRDURATION) % s->dlbuflen;
    /* Update the full wave rectified total amplitude */
    /* Input matrix decoder */
    if(s->decode_mode == HRTF_MIX_MATRIX2CH) {
       s->l_fwr += abs(in[0]) - fabs(s->fwrbuf_l[fwr_pos]);
       s->r_fwr += abs(in[1]) - fabs(s->fwrbuf_r[fwr_pos]);
       s->lpr_fwr += abs(in[0] + in[1]) -
	  fabs(s->fwrbuf_l[fwr_pos] + s->fwrbuf_r[fwr_pos]);
       s->lmr_fwr += abs(in[0] - in[1]) -
	  fabs(s->fwrbuf_l[fwr_pos] - s->fwrbuf_r[fwr_pos]);
    }
    /* Rear matrix decoder */
    if(s->matrix_mode) {
       s->lr_fwr += abs(in[2]) - fabs(s->fwrbuf_lr[fwr_pos]);
       s->rr_fwr += abs(in[3]) - fabs(s->fwrbuf_rr[fwr_pos]);
       s->lrprr_fwr += abs(in[2] + in[3]) -
	  fabs(s->fwrbuf_lr[fwr_pos] + s->fwrbuf_rr[fwr_pos]);
       s->lrmrr_fwr += abs(in[2] - in[3]) -
	  fabs(s->fwrbuf_lr[fwr_pos] - s->fwrbuf_rr[fwr_pos]);
    }

    switch (s->decode_mode) {
    case HRTF_MIX_51:
       /* 5/5+1 channel sources */
       s->lf[k] = in[0];
       s->cf[k] = in[4];
       s->rf[k] = in[1];
       s->fwrbuf_lr[k] = s->lr[k] = in[2];
       s->fwrbuf_rr[k] = s->rr[k] = in[3];
       break;
    case HRTF_MIX_MATRIX2CH:
       /* Matrix encoded 2 channel sources */
       s->fwrbuf_l[k] = in[0];
       s->fwrbuf_r[k] = in[1];
       matrix_decode(in, k, 0, 1, 1, s->dlbuflen,
		     s->l_fwr, s->r_fwr,
		     s->lpr_fwr, s->lmr_fwr,
		     &(s->adapt_l_gain), &(s->adapt_r_gain),
		     &(s->adapt_lpr_gain), &(s->adapt_lmr_gain),
		     s->lf, s->rf, s->lr, s->rr, s->cf);
       break;
    case HRTF_MIX_STEREO:
       /* Stereo sources */
       s->lf[k] = in[0];
       s->rf[k] = in[1];
       s->cf[k] = s->lr[k] = s->rr[k] = 0;
       break;
    }

    /* We need to update the bass compensation delay line, too. */
    s->ba_l[k] = in[0] + in[4] + in[2];
    s->ba_r[k] = in[4] + in[1] + in[3];
}

/* Initialization and runtime control */
static int control(struct af_instance_s *af, int cmd, void* arg)
{
    af_hrtf_t *s = af->setup;
    int test_output_res;
    char mode;

    switch(cmd) {
    case AF_CONTROL_REINIT:
	af->data->rate   = ((af_data_t*)arg)->rate;
	if(af->data->rate != 48000) {
	    // automatic samplerate adjustment in the filter chain
	    // is not yet supported.
	    mp_msg(MSGT_AFILTER, MSGL_ERR,
		   "[hrtf] ERROR: Sampling rate is not 48000 Hz (%d)!\n",
		   af->data->rate);
	    return AF_ERROR;
	}
	af->data->nch    = ((af_data_t*)arg)->nch;
	    if(af->data->nch == 2) {
 	       /* 2 channel input */
 	       if(s->decode_mode != HRTF_MIX_MATRIX2CH) {
   		  /* Default behavior is stereo mixing. */
 		  s->decode_mode = HRTF_MIX_STEREO;
	       }
	    }
	    else if (af->data->nch < 5)
	      af->data->nch = 5;
	af->data->format = AF_FORMAT_S16_NE;
	af->data->bps    = 2;
	test_output_res = af_test_output(af, (af_data_t*)arg);
	af->mul = 2.0 / af->data->nch;
	// after testing input set the real output format
	af->data->nch = 2;
	s->print_flag = 1;
	return test_output_res;
    case AF_CONTROL_COMMAND_LINE:
	sscanf((char*)arg, "%c", &mode);
	switch(mode) {
	case 'm':
	    /* Use matrix rear decoding. */
	    s->matrix_mode = 1;
	    break;
	case 's':
	    /* Input needs matrix decoding. */
	    s->decode_mode = HRTF_MIX_MATRIX2CH;
	    break;
	case '0':
	    s->matrix_mode = 0;
	    break;
	default:
	    mp_msg(MSGT_AFILTER, MSGL_ERR,
		   "[hrtf] Mode is neither 'm', 's', nor '0' (%c).\n",
		   mode);
	    return AF_ERROR;
	}
	s->print_flag = 1;
	return AF_OK;
    }

    return AF_UNKNOWN;
}

/* Deallocate memory */
static void uninit(struct af_instance_s *af)
{
    if(af->setup) {
	af_hrtf_t *s = af->setup;

	free(s->lf);
	free(s->rf);
	free(s->lr);
	free(s->rr);
	free(s->cf);
	free(s->cr);
	free(s->ba_l);
	free(s->ba_r);
	free(s->ba_ir);
	free(s->fwrbuf_l);
	free(s->fwrbuf_r);
	free(s->fwrbuf_lr);
	free(s->fwrbuf_rr);
	free(af->setup);
    }
    if(af->data)
	free(af->data->audio);
    free(af->data);
}

/* Filter data through filter

Two "tricks" are used to compensate the "color" of the KEMAR data:

1. The KEMAR data is refiltered to ensure that the front L, R channels
on the same side of the ear are equalized (especially in the high
frequencies).

2. A bass compensation is introduced to ensure that 0-200 Hz are not
damped (without any real 3D acoustical image, however).
*/
static af_data_t* play(struct af_instance_s *af, af_data_t *data)
{
    af_hrtf_t *s = af->setup;
    short *in = data->audio; // Input audio data
    short *out = NULL; // Output audio data
    short *end = in + data->len / sizeof(short); // Loop end
    float common, left, right, diff, left_b, right_b;
    const int dblen = s->dlbuflen, hlen = s->hrflen, blen = s->basslen;

    if(AF_OK != RESIZE_LOCAL_BUFFER(af, data))
	return NULL;

    if(s->print_flag) {
	s->print_flag = 0;
	switch (s->decode_mode) {
	case HRTF_MIX_51:
	  mp_msg(MSGT_AFILTER, MSGL_INFO,
		 "[hrtf] Using HRTF to mix %s discrete surround into "
		 "L, R channels\n", s->matrix_mode ? "5+1" : "5");
	  break;
	case HRTF_MIX_STEREO:
	  mp_msg(MSGT_AFILTER, MSGL_INFO,
		 "[hrtf] Using HRTF to mix stereo into "
		 "L, R channels\n");
	  break;
	case HRTF_MIX_MATRIX2CH:
	  mp_msg(MSGT_AFILTER, MSGL_INFO,
		 "[hrtf] Using active matrix to decode 2 channel "
		 "input, HRTF to mix %s matrix surround into "
		 "L, R channels\n", "3/2");
	  break;
	default:
	  mp_msg(MSGT_AFILTER, MSGL_WARN,
		 "[hrtf] bogus decode_mode: %d\n", s->decode_mode);
	  break;
	}

       if(s->matrix_mode)
	  mp_msg(MSGT_AFILTER, MSGL_INFO,
		 "[hrtf] Using active matrix to decode rear center "
		 "channel\n");
    }

    out = af->data->audio;

    /* MPlayer's 5 channel layout (notation for the variable):
     *
     * 0: L (LF), 1: R (RF), 2: Ls (LR), 3: Rs (RR), 4: C (CF), matrix
     * encoded: Cs (CR)
     *
     * or: L = left, C = center, R = right, F = front, R = rear
     *
     * Filter notation:
     *
     *      CF
     * OF        AF
     *      Ear->
     * OR        AR
     *      CR
     *
     * or: C = center, A = same side, O = opposite, F = front, R = rear
     */

    while(in < end) {
	const int k = s->cyc_pos;

	update_ch(s, in, k);

	/* Simulate a 7.5 ms -20 dB echo of the center channel in the
	   front channels (like reflection from a room wall) - a kind of
	   psycho-acoustically "cheating" to focus the center front
	   channel, which is normally hard to be perceived as front */
	s->lf[k] += CFECHOAMPL * s->cf[(k + CFECHODELAY) % s->dlbuflen];
	s->rf[k] += CFECHOAMPL * s->cf[(k + CFECHODELAY) % s->dlbuflen];

	switch (s->decode_mode) {
	case HRTF_MIX_51:
	case HRTF_MIX_MATRIX2CH:
	   /* Mixer filter matrix */
	   common = conv(dblen, hlen, s->cf, s->cf_ir, k + s->cf_o);
	   if(s->matrix_mode) {
	      /* In matrix decoding mode, the rear channel gain must be
		 renormalized, as there is an additional channel. */
	      matrix_decode(in, k, 2, 3, 0, s->dlbuflen,
			    s->lr_fwr, s->rr_fwr,
			    s->lrprr_fwr, s->lrmrr_fwr,
			    &(s->adapt_lr_gain), &(s->adapt_rr_gain),
			    &(s->adapt_lrprr_gain), &(s->adapt_lrmrr_gain),
			    s->lr, s->rr, NULL, NULL, s->cr);
	      common +=
		 conv(dblen, hlen, s->cr, s->cr_ir, k + s->cr_o) *
		 M1_76DB;
	      left    =
		 ( conv(dblen, hlen, s->lf, s->af_ir, k + s->af_o) +
		   conv(dblen, hlen, s->rf, s->of_ir, k + s->of_o) +
		   (conv(dblen, hlen, s->lr, s->ar_ir, k + s->ar_o) +
		    conv(dblen, hlen, s->rr, s->or_ir, k + s->or_o)) *
		   M1_76DB + common);
	      right   =
		 ( conv(dblen, hlen, s->rf, s->af_ir, k + s->af_o) +
		   conv(dblen, hlen, s->lf, s->of_ir, k + s->of_o) +
		   (conv(dblen, hlen, s->rr, s->ar_ir, k + s->ar_o) +
		    conv(dblen, hlen, s->lr, s->or_ir, k + s->or_o)) *
		   M1_76DB + common);
	   } else {
	      left    =
		 ( conv(dblen, hlen, s->lf, s->af_ir, k + s->af_o) +
		   conv(dblen, hlen, s->rf, s->of_ir, k + s->of_o) +
		   conv(dblen, hlen, s->lr, s->ar_ir, k + s->ar_o) +
		   conv(dblen, hlen, s->rr, s->or_ir, k + s->or_o) +
		   common);
	      right   =
		 ( conv(dblen, hlen, s->rf, s->af_ir, k + s->af_o) +
		   conv(dblen, hlen, s->lf, s->of_ir, k + s->of_o) +
		   conv(dblen, hlen, s->rr, s->ar_ir, k + s->ar_o) +
		   conv(dblen, hlen, s->lr, s->or_ir, k + s->or_o) +
		   common);
	   }
	   break;
	case HRTF_MIX_STEREO:
	   left    =
	      ( conv(dblen, hlen, s->lf, s->af_ir, k + s->af_o) +
		conv(dblen, hlen, s->rf, s->of_ir, k + s->of_o));
	   right   =
	      ( conv(dblen, hlen, s->rf, s->af_ir, k + s->af_o) +
		conv(dblen, hlen, s->lf, s->of_ir, k + s->of_o));
	   break;
	default:
	    /* make gcc happy */
	    left = 0.0;
	    right = 0.0;
	    break;
	}

	/* Bass compensation for the lower frequency cut of the HRTF.  A
	   cross talk of the left and right channel is introduced to
	   match the directional characteristics of higher frequencies.
	   The bass will not have any real 3D perception, but that is
	   OK (note at 180 Hz, the wavelength is about 2 m, and any
	   spatial perception is impossible). */
	left_b  = conv(dblen, blen, s->ba_l, s->ba_ir, k);
	right_b = conv(dblen, blen, s->ba_r, s->ba_ir, k);
	left  += (1 - BASSCROSS) * left_b  + BASSCROSS * right_b;
	right += (1 - BASSCROSS) * right_b + BASSCROSS * left_b;
	/* Also mix the LFE channel (if available) */
	if(data->nch >= 6) {
	    left  += in[5] * M3_01DB;
	    right += in[5] * M3_01DB;
	}

	/* Amplitude renormalization. */
	left  *= AMPLNORM;
	right *= AMPLNORM;

	switch (s->decode_mode) {
	case HRTF_MIX_51:
	case HRTF_MIX_STEREO:
	   /* "Cheating": linear stereo expansion to amplify the 3D
	      perception.  Note: Too much will destroy the acoustic space
	      and may even result in headaches. */
	   diff = STEXPAND2 * (left - right);
	   out[0] = (int16_t)(left  + diff);
	   out[1] = (int16_t)(right - diff);
	   break;
	case HRTF_MIX_MATRIX2CH:
	   /* Do attempt any stereo expansion with matrix encoded
	      sources.  The L, R channels are already stereo expanded
	      by the steering, any further stereo expansion will sound
	      very unnatural. */
	   out[0] = (int16_t)left;
	   out[1] = (int16_t)right;
	   break;
	}

	/* Next sample... */
	in = &in[data->nch];
	out = &out[af->data->nch];
	(s->cyc_pos)--;
	if(s->cyc_pos < 0)
	    s->cyc_pos += dblen;
    }

    /* Set output data */
    data->audio = af->data->audio;
    data->len   = data->len / data->nch * 2;
    data->nch   = 2;

    return data;
}

static int allocate(af_hrtf_t *s)
{
    if ((s->lf = malloc(s->dlbuflen * sizeof(float))) == NULL) return -1;
    if ((s->rf = malloc(s->dlbuflen * sizeof(float))) == NULL) return -1;
    if ((s->lr = malloc(s->dlbuflen * sizeof(float))) == NULL) return -1;
    if ((s->rr = malloc(s->dlbuflen * sizeof(float))) == NULL) return -1;
    if ((s->cf = malloc(s->dlbuflen * sizeof(float))) == NULL) return -1;
    if ((s->cr = malloc(s->dlbuflen * sizeof(float))) == NULL) return -1;
    if ((s->ba_l = malloc(s->dlbuflen * sizeof(float))) == NULL) return -1;
    if ((s->ba_r = malloc(s->dlbuflen * sizeof(float))) == NULL) return -1;
    if ((s->fwrbuf_l =
	 malloc(s->dlbuflen * sizeof(float))) == NULL) return -1;
    if ((s->fwrbuf_r =
	 malloc(s->dlbuflen * sizeof(float))) == NULL) return -1;
    if ((s->fwrbuf_lr =
	 malloc(s->dlbuflen * sizeof(float))) == NULL) return -1;
    if ((s->fwrbuf_rr =
	 malloc(s->dlbuflen * sizeof(float))) == NULL) return -1;
    return 0;
}

/* Allocate memory and set function pointers */
static int af_open(af_instance_t* af)
{
    int i;
    af_hrtf_t *s;
    float fc;

    af->control = control;
    af->uninit = uninit;
    af->play = play;
    af->mul = 1;
    af->data = calloc(1, sizeof(af_data_t));
    af->setup = calloc(1, sizeof(af_hrtf_t));
    if((af->data == NULL) || (af->setup == NULL))
	return AF_ERROR;

    s = af->setup;

    s->dlbuflen = DELAYBUFLEN;
    s->hrflen = HRTFFILTLEN;
    s->basslen = BASSFILTLEN;

    s->cyc_pos = s->dlbuflen - 1;
    /* With a full (two axis) steering matrix decoder, s->matrix_mode
       should not be enabled lightly (it will also steer the Ls, Rs
       channels). */
    s->matrix_mode = 0;
    s->decode_mode = HRTF_MIX_51;

    s->print_flag = 1;

    if (allocate(s) != 0) {
 	mp_msg(MSGT_AFILTER, MSGL_ERR, "[hrtf] Memory allocation error.\n");
	return AF_ERROR;
    }

    for(i = 0; i < s->dlbuflen; i++)
	s->lf[i] = s->rf[i] = s->lr[i] = s->rr[i] = s->cf[i] =
	    s->cr[i] = 0;

    s->lr_fwr =
	s->rr_fwr = 0;

    s->cf_ir = cf_filt + (s->cf_o = pulse_detect(cf_filt));
    s->af_ir = af_filt + (s->af_o = pulse_detect(af_filt));
    s->of_ir = of_filt + (s->of_o = pulse_detect(of_filt));
    s->ar_ir = ar_filt + (s->ar_o = pulse_detect(ar_filt));
    s->or_ir = or_filt + (s->or_o = pulse_detect(or_filt));
    s->cr_ir = cr_filt + (s->cr_o = pulse_detect(cr_filt));

    if((s->ba_ir = malloc(s->basslen * sizeof(float))) == NULL) {
 	mp_msg(MSGT_AFILTER, MSGL_ERR, "[hrtf] Memory allocation error.\n");
	return AF_ERROR;
    }
    fc = 2.0 * BASSFILTFREQ / (float)af->data->rate;
    if(af_filter_design_fir(s->basslen, s->ba_ir, &fc, LP | KAISER, 4 * M_PI) ==
       -1) {
	mp_msg(MSGT_AFILTER, MSGL_ERR, "[hrtf] Unable to design low-pass "
	       "filter.\n");
	return AF_ERROR;
    }
    for(i = 0; i < s->basslen; i++)
	s->ba_ir[i] *= BASSGAIN;

    return AF_OK;
}

/* Description of this filter */
af_info_t af_info_hrtf = {
    "HRTF Headphone",
    "hrtf",
    "ylai",
    "",
    AF_FLAGS_REENTRANT,
    af_open
};