soundeditor/atsa/synthesis.c

/*
SYNTHESIS.C
Oscar Pablo Di Liscia / Juan Pampin

Extracted from atsh synth-funcs.c and modified by jpmeuret@free.fr
- moved to double computation everywhere
- do_synthesis no longer writes to a sound file, but simply allocates and returns
  a sample array => now independant from any sound IO library.
- improved synthesis algorithm, that precisely takes into account
  all the frames pieces (but suppose the time streching/expanding
  function is always increasing, contrary to original algorithm).
*/

#include <stdlib.h>
#include <stdio.h>
#include <math.h>
#include <string.h>
#include <time.h>

#include "atsa.h"


//CONSTANTS
#define TWO_PI (2*M_PI)

#define SINE_TABLE_LEN 16384

#define BW 0.1
#define BW_FREQ 500.

#define EPSILON 0.001

//MACROS
/*
;;; In these macros pha_1 and frq_1 refers to the instantaneous phase
;;; and frequency of the previous frame while pha and frq refer
;;; to the phase and frequency of the present frame.
*/
#define COMPUTE_M(pha_1, frq_1, pha, frq, dt) ((pha_1 + (frq_1 * dt) - pha) + ((frq - frq_1) * .5 * dt)) / TWO_PI   
#define COMPUTE_AUX(pha_1, pha, frq_1, dt, M) (pha + (TWO_PI * M)) - (pha_1 + (frq_1 * dt))
#define COMPUTE_ALPHA(aux, frq_1, frq, dt) ((3. / (dt * dt)) * aux ) - ((frq - frq_1) / dt)
#define COMPUTE_BETA(aux, frq_1, frq, dt) ((-2. / (dt * dt * dt)) * aux) + ((frq - frq_1) / (dt * dt))

// Linear interpolation with 3 given points (-.5,yp), (0.5,yc), (1.5, yn)
#define LIN3_INTERP(x, yp, yc, yn) \
 (((x) < .5) ? ((yc) + (yp)) / 2 + x * ((yc) - (yp)) : (3*(yc) - (yn)) / 2 + x * ((yn) - (yc)))

/*
;;; note that for this macro the values of i should go
;;; from 0 to dt. So in the case of having 220 samples
;;; within a frame the increment for i will be dt/200
*/
#define INTERP_PHASE(pha_1, frq_1, alpha, beta, i) (beta * i * i * i) + (alpha * i * i)+ (frq_1 * i) + pha_1

#define ENG_RMS(val, ws) sqrt((double)val/(ws * (double)ATSA_NOISE_VARIANCE))

#define fair_floor(value, epsilon) (int)(ceil(value) - (value) <= (epsilon) ? ceil(value) : floor(value))

//STRUCTURES
typedef struct { //the time data for each segment of the time function
  double beg_frame;   // 1st frame index in [0.0, sound.frames-1]
  double end_frame;   // last frame index in [1.0, sound.frames]
  double beg_time;    // start time (non streched) (relative to sparams->beg)
  double end_time;    // end time (non streched) (relative to sparams->beg)
  double time_factor; // time stretching factor (1=invariant)
} TIME_DATA;

typedef struct { //the data for the randi UG
  int    size; //size of the frame in samples this should be sr/freq.
  double a1;   //first amplitude value
  double a2;   //next  amplitude value
  int    cnt;  //sample position counter
} RANDI;


//GLOBAL VARIABLES
static int sine_table_ready = 0;
static double sine_table[SINE_TABLE_LEN];

///////////////////////////////////////////////////////////////////
//randi output random numbers in the range of 1,-1
//getting a new number at frequency freq and interpolating
//the intermediate values.
void randi_setup(double sr, double freq, RANDI *radat)
{
  // Initialize random seed
  srand((unsigned)time(0));

  // Initialize 
  radat->size = (int) (sr / freq) - 1;
  radat->a1   = rand();
  radat->a2   = rand();
  radat->cnt  = 0;
}
///////////////////////////////////////////////////////////////////
double randi(RANDI *radat)
{
  double output;

  if (radat->cnt == radat->size) { //get a new random value
    radat->a1  = radat->a2; 
    radat->a2  = rand();
    radat->cnt = 0;
  }

  output = radat->a1 + (radat->a2 - radat->a1) * radat->cnt / radat->size;
  radat->cnt++;

  return 1. - 2. * output / RAND_MAX;
}
///////////////////////////////////////////////////////////////////
double randif(RANDI *radat, double freq, double sr)
{
  double output;

  if(radat->cnt == radat->size) { //get a new random value
    radat->a1  = radat->a2; 
    radat->a2  = rand();
    radat->cnt = 0;
    radat->size= (int) (sr / freq) - 1;
  }

  output= radat->a1 + (radat->a2 - radat->a1) * radat->cnt / radat->size;
  radat->cnt++;
  
  return 1. - 2. * output / RAND_MAX;
}
///////////////////////////////////////////////////////////////////
void make_sine_table() 
{
  static const double incr = TWO_PI / SINE_TABLE_LEN;
  int i;
  double theta = 0.;
  
  for(i=0; i < SINE_TABLE_LEN; i++) {
    sine_table[i] = sin(theta);
    theta += incr;
  }

  sine_table_ready = 1;
}
////////////////////////////////////////////////////////////////////
double ioscilator(double amp, double freq, double pha, double sr, double *oscpt)
{
  const double incr = freq * (double)SINE_TABLE_LEN / sr; 

  // Phase management = 3 options (uncomment only one) :
  // 1) No phase = 0 phase.
  const double osc = *oscpt; 

  // 2) Random phase.
  //static const int max_pha_shift = 300; //(SINE_TABLE_LEN / 2) / 1000;
  //const int pha_shift = rand() % (2 * max_pha_shift) - max_pha_shift;
  //const int pha_shift = 
  //    (int) ((2.0 * max_pha_shift + 1.0) * (rand() / (RAND_MAX + 1.0))) - max_pha_shift;
  //const double osc = fmod(*oscpt + SINE_TABLE_LEN + pha_shift, SINE_TABLE_LEN);

  // 3) Shipped phase (linear interpolation).
  //const int pha_shift = (int)(pha * (double)SINE_TABLE_LEN / TWO_PI);
  //const double osc = fmod(*oscpt + SINE_TABLE_LEN + pha_shift, SINE_TABLE_LEN);

  double output;
  int curr_ind, next_ind;

  if (!sine_table_ready)
    make_sine_table();

  // Linear interpolation of the amplitude from the sine table.
  curr_ind = (int)floor(osc);
  next_ind = (curr_ind + 1) % SINE_TABLE_LEN;
  output = amp * (sine_table[curr_ind]
                  + (sine_table[next_ind] - sine_table[curr_ind]) 
                    * (osc - curr_ind));

  // Update oscillator index.
  *oscpt = fmod(*oscpt + incr, SINE_TABLE_LEN);

  return output;
}
///////////////////////////////////////////////////////////////
double locate_frame(ATS_SOUND *ats_sound, double from_frame, double time)
{
  //Assuming that the duration of each frame may be different, we
  //do not have any other method to locate the frame for a given time
  double frame;
  int i_frame;

  if (from_frame < 0) 
    i_frame = 0;
  else if (from_frame > ats_sound->frames - 1) 
    i_frame = ats_sound->frames - 1;
  else
    i_frame = (int)floor(from_frame);

  while (i_frame < ats_sound->frames - 1
         && time > ats_sound->time[0][i_frame + 1])
  {
    fprintf(stderr, "locate_frame : i=%d, time[i+1]=%f\n",
            i_frame, ats_sound->time[0][i_frame + 1]);
    i_frame++;
  }

  if (i_frame == ats_sound->frames - 1)
    frame = (time - ats_sound->time[0][i_frame])
            / ((double)ats_sound->dur - ats_sound->time[0][i_frame]);
  else
    frame = (time - ats_sound->time[0][i_frame])
            / ((double)ats_sound->time[0][i_frame + 1] - ats_sound->time[0][i_frame]);
  frame += i_frame;
  if (frame >= ats_sound->frames)
    frame = ats_sound->frames;
  if (frame - floor(frame) < 1.0 / ats_sound->frame_size)
    // The time interval for a rame is open upward [beg, end[.
    frame *= (1.0 - 0.01 / ats_sound->frame_size);

  fprintf(stderr, "locate_frame(from=%f, time=%f) = %f\n", from_frame, time, frame);

  return frame;
}
////////////////////////////////////////////////////////////////////
//Synthesizes a Buffer using phase interpolation (not used for the moment)
/*void synth_buffer_phint(double a1, double a2, double f1, double f2, double p1, double p2, double dt, double frame_samps, double* frbuf)
{
  double t_inc, a_inc, M, aux, alpha, beta, time, amp, scale, new_phase;
  int k, index;
  double out=0., phase=0.;

  if (!sine_table)
    make_sine_table();

  f1  *=TWO_PI;
  f2  *=TWO_PI;
  t_inc= dt  / frame_samps;
  a_inc= (a2 - a1) / frame_samps;
  M    = COMPUTE_M(p1, f1, p2, f2,dt);
  aux  = COMPUTE_AUX(p1, p2, f1, dt, M);
  alpha= COMPUTE_ALPHA(aux,f1,f2,dt);
  beta = COMPUTE_BETA(aux,f1,f2,dt);
  time = 0.;
  amp  = a1;
  scale = TWO_PI / (SINE_TABLE_LEN - 1); // must take it out from here...

  for(k = 0; k < (int)frame_samps; k++) { 
 
    phase = INTERP_PHASE(p1,f1,alpha,beta,time);
    new_phase = (phase >= TWO_PI  ? phase - TWO_PI : phase);
    index=(int)((new_phase / TWO_PI)*(double)SINE_TABLE_LEN - 1.);
    while ( index >= SINE_TABLE_LEN ) {
      index -=SINE_TABLE_LEN;
    }
     while ( index < 0 ) {
      index +=SINE_TABLE_LEN;
    }
    out = sine_table[index] * amp;

    /////////////////////////////////////////////////////////
    time +=t_inc;
    amp  +=a_inc;
    frbuf[k] +=out; //buffer adds each partial at each pass

  }
}
*/
////////////////////////////////////////////////////////////////////
//Synthesizes a Buffer NOT using phase interpolation
int synth_deterministic_only(double ampl_p, double ampl_c, double ampl_n,
                             double freq_p, double freq_c, double freq_n, 
                             double pha_p, double pha_c, double pha_n, 
                             double time_offset, double duration, double sample_rate, 
                             double *oscpt, double* sample_buf)
{
  int s, frame_samps;
  double ampl, freq, pha, frame_offset;

  frame_samps = fair_floor(sample_rate * duration, EPSILON);

  if (ampl_p == 0. && ampl_c == 0. && ampl_n == 0.) 
      return frame_samps; //nothing to do if no amplitude

  for(s = 0; s < frame_samps; s++) {
    frame_offset = (time_offset + s / sample_rate) / duration;
    ampl = LIN3_INTERP(frame_offset, ampl_p, ampl_c, ampl_n);
    freq = LIN3_INTERP(frame_offset, freq_p, freq_c, freq_n);
    pha  = 0; // Temporary: LIN3_INTERP(frame_offset, pha_p,   pha_c,   pha_n);
    sample_buf[s] += ioscilator(ampl, freq, pha, sample_rate, oscpt);
  }

  return frame_samps;
}
////////////////////////////////////////////////////////////////////
int synth_residual_only(double ampl_p, double ampl_c, double ampl_n,
                        double freq,
                        double time_offset, double duration, double sample_rate,
                        double *oscpt, RANDI* rdata, double* sample_buf)
{
  int s, frame_samps;
  double ampl, pha, frame_offset;

  frame_samps = fair_floor(sample_rate * duration, EPSILON);

  if(ampl_p==0. && ampl_c==0. && ampl_n==0.) 
      return frame_samps; //nothing to do if no amplitude

  for(s = 0; s < frame_samps; s++) { 
    frame_offset = (time_offset + s / sample_rate) / duration;
    ampl = LIN3_INTERP(frame_offset, ampl_p, ampl_c, ampl_n);
    pha = 0.; // Temporary. Should we use a random phase here ?
    sample_buf[s] += ioscilator(ampl, freq, pha, sample_rate, oscpt) * randi(rdata);  
  }

  return frame_samps;
 }
////////////////////////////////////////////////////////////////////
int synth_both(double ampl_p, double ampl_c, double ampl_n,
               double freq_p, double freq_c, double freq_n, 
               double pha_p, double pha_c, double pha_n, 
               double resid_p, double resid_c, double resid_n, 
               double time_offset, double duration, double sample_rate, 
               double *oscpt, RANDI* rdata, double* sample_buf)
{
  int s, frame_samps;
  double ampl, freq, pha, resid, rfreq, frame_offset;
  double rfreq_p, rfreq_c, rfreq_n; 

  frame_samps = fair_floor(sample_rate * duration, EPSILON);

  if(ampl_p==0. && ampl_c==0. && ampl_n==0.
     && resid_p==0. && resid_c==0. && resid_n==0.)
    return frame_samps; //nothing to do if no amplitude

  rfreq_p = BW * (freq_p < BW_FREQ ? BW_FREQ : freq_p / 2);
  rfreq_c = BW * (freq_c < BW_FREQ ? BW_FREQ : freq_c / 2);
  rfreq_n = BW * (freq_n < BW_FREQ ? BW_FREQ : freq_n / 2);

  for(s = 0; s < frame_samps; s++) {
    frame_offset = (time_offset + s / sample_rate) / duration;
    ampl  = LIN3_INTERP(frame_offset, ampl_p,  ampl_c,  ampl_n);
    freq  = LIN3_INTERP(frame_offset, freq_p,  freq_c,  freq_n);
    pha   = 0; // Temporary: LIN3_INTERP(frame_offset, pha_p,   pha_c,   pha_n);
    resid = LIN3_INTERP(frame_offset, resid_p, resid_c, resid_n);
    rfreq = LIN3_INTERP(frame_offset, rfreq_p, rfreq_c, rfreq_n);
    sample_buf[s] += 
      ioscilator(1.0, freq, pha, sample_rate, oscpt)
      * (ampl + resid * randif(rdata, rfreq, sample_rate));  
  }

  return frame_samps;
}

////////////////////////////////////////////////////////////////////
////////////////////////////////////////////////////////////////////
/////////THIS IS THE MAIN SYNTHESIS LOOP////////////////////////////
////////////////////////////////////////////////////////////////////
void do_synthesis(ATS_SOUND *ats_sound, SPARAMS* sparams, CURVE* timenv, int *selected,
                  double** out_samps, int* n_out_samps)
{
  double dt=0., rfreq;
  int frame_samps=0;
  double bframe, eframe;
  double dur, dy, bxval, byval, exval, eyval, difx, dify;
  TIME_DATA *tdata;
  int nbp;
  double *ospt=0;
  RANDI *rarray=0;
  static double res_band_centers[ATSA_CRITICAL_BANDS];
  int todo;
  double ampl_p, ampl_c, ampl_n;
  double freq_p, freq_c, freq_n;
  double pha_p, pha_c, pha_n;
  double resid_p, resid_c, resid_n;
  int n_samps;
  int first, last;
  int f, f_p, f_n;
  double frame_dur, time_offset, duration;
  int p, d, b;

  // Fix special params values
  if (sparams->end <= sparams->beg)
      sparams->end = sparams->beg + ats_sound->dur;

  // Transfer residual data if partial noise is present
  if (ats_sound->band_energy) {
    for(f=0; f<ats_sound->frames; f++) {
      band_energy_to_res(ats_sound, f);
    }
  }

  // Build envelop segments descriptor and compute number of samples to generate
  nbp   = get_nbp(timenv);
  tdata = (TIME_DATA*)malloc(nbp * sizeof(TIME_DATA));
  dy    = get_maxy_value(timenv) - get_miny_value(timenv);
  dur   = sparams->end - sparams->beg;
  todo=0;
  bframe = 0.;
  fprintf(stdout, "do_synthesis: nbp=%d, dur=%f, dy=%f\n", nbp, dur, dy);

  for(p=0; p < nbp - 1; ++p){

    //get the data from the time envelope and convert it to time
    bxval= dur * get_x_value(timenv, p); // We assume xmin=0, xmax=1
    byval= dur * get_y_value(timenv, p) * dy; // We assume ymin=0
    exval= dur * get_x_value(timenv, p+1); // We assume xmin=0, xmax=1
    eyval= dur * get_y_value(timenv, p+1) * dy; // We assume ymin=0
    
    fprintf(stdout, "do_synthesis: seg %d : bxval=%f, byval=%f, exval=%f, eyval=%f\n",
            p, bxval, byval, exval, eyval);

    //diff=0. is a special case we must take in account
    //here all we do is to set it to one millisecond (arbitrarly)
    difx= exval - bxval; 
    if(difx == 0.)
        difx=.001;
    dify= eyval - byval;
    if(dify == 0.)
        dify=.001;

    //locate the frame for the begining and end of segments
    bframe= locate_frame(ats_sound, bframe, byval);
    eframe= locate_frame(ats_sound, bframe, eyval);

    //collect the data to be used
    tdata[p].beg_frame   = bframe;
    tdata[p].end_frame   = eframe;
    tdata[p].beg_time    = byval;
    tdata[p].end_time    = eyval;
    tdata[p].time_factor = fabs(difx/dify);

    // update the number of samples to synthesise
    dt = fabs(eyval - byval);
    todo += fair_floor(dt * sparams->sr * tdata[p].time_factor, EPSILON);

    bframe=eframe;
  }
  
  // Allocate and zero output sample array
  *out_samps = (double*)malloc(todo*sizeof(double));
  memset(*out_samps, 0, todo*sizeof(double));
  //fprintf(stdout, "do_synthesis: %d samples to be generated\n", todo);

  // Allocate space for oscilators and noise generator
  if(sparams->ramp == 0.) { 
        //deterministic synthesis only
    ospt = (double*)malloc(ats_sound->partials * sizeof(double));
        memset(ospt, 0, ats_sound->partials * sizeof(double));
  }  else if(sparams->amp == 0.) { 
        //residual synthesis only
    ospt = (double*)malloc(ATSA_CRITICAL_BANDS * sizeof(double));
        memset(ospt, 0, ATSA_CRITICAL_BANDS * sizeof(double));
    rarray= (RANDI*)malloc(ATSA_CRITICAL_BANDS * sizeof(RANDI));
    for(b=0; b<ATSA_CRITICAL_BANDS; ++b) {
      res_band_centers[b]= 
                (ATSA_CRITICAL_BAND_EDGES[b+1] + ATSA_CRITICAL_BAND_EDGES[b]) / 2;
      randi_setup(sparams->sr, ATSA_CRITICAL_BAND_EDGES[b+1] - ATSA_CRITICAL_BAND_EDGES[b],
                                  &rarray[b]);
    }
  } else { 
        //residual and deterministic synthesis
    ospt = (double*)malloc(ats_sound->partials * sizeof(double));
        memset(ospt, 0, ats_sound->partials * sizeof(double));
    rarray= (RANDI*)malloc(ats_sound->partials * sizeof(RANDI));
    for(p=0; p<ats_sound->partials; p++) {
          rfreq=BW * (ats_sound->frq[p][(int)floor(tdata[0].beg_frame)] < BW_FREQ ?
                      BW_FREQ : ats_sound->frq[p][(int)floor(tdata[0].beg_frame)]);
      randi_setup(sparams->sr,rfreq,&rarray[p]);
    }
  }

  // Generate samples :
  // For each time function control point/segment :
  n_samps=0;
  for(p = 0; p < nbp - 1; p++) { 

    first=(int)floor(tdata[p].beg_frame);
    last=(int)floor(tdata[p].end_frame);
 
    fprintf(stdout, "do_synthesis: seg %d : begf=%d, endf=%d, fact=%f, begt=%f, endt=%f\n",
            p, first, last, tdata[p].time_factor, tdata[p].beg_time, tdata[p].end_time);

    // For each frame inside the segment :
    for(f = first; f <= last; f++) {

      // Determine index of "previous" and "next" frame for interpolation.
      f_p = f < 1 ? 0 : f - 1;
      f_n = f >= ats_sound->frames - 1 ? ats_sound->frames - 1 : f + 1;

      // Determine frame duration.
      if (f < ats_sound->frames - 1)
          frame_dur = ats_sound->time[0][f+1] - ats_sound->time[0][f];
      else
          frame_dur = ats_sound->dur - ats_sound->time[0][f];
          
      // Determine time origin in frame and duration for sample generation.
      if (f == first && f != tdata[p].beg_frame)
      {
        time_offset = tdata[p].beg_time - ats_sound->time[0][f];
        duration = frame_dur - time_offset;
      }
      else if (f == last && f != tdata[p].end_frame)
      {
        duration = tdata[p].end_time - ats_sound->time[0][f];
        time_offset = 0.0;
      }
      else
      {
          duration = frame_dur;
          time_offset = 0.0;
      }

      // Apply the time factor.
      duration *= tdata[p].time_factor;      
      time_offset *= tdata[p].time_factor;      

      // Do the sample generation.
      if(sparams->ramp == 0.) {
          //deterministic synthesis only
          for(d = 0; d < ats_sound->partials; d++) {
              if (selected && !selected[d])
                  continue;
              ampl_p = ats_sound->amp[d][f_p] * sparams->amp;
              ampl_c = ats_sound->amp[d][f]   * sparams->amp;
              ampl_n = ats_sound->amp[d][f_n] * sparams->amp;
              freq_p = ats_sound->frq[d][f_p] * sparams->frec;
              freq_c = ats_sound->frq[d][f]   * sparams->frec;
              freq_n = ats_sound->frq[d][f_n] * sparams->frec;
              pha_p = ats_sound->pha[d][f_p];
              pha_c = ats_sound->pha[d][f];
              pha_n = ats_sound->pha[d][f_n];
              frame_samps = synth_deterministic_only(ampl_p, ampl_c, ampl_n, freq_p, freq_c, freq_n,
                                                     pha_p, pha_c, pha_n,
                                                     time_offset, duration, sparams->sr,
                                                     ospt+d, *out_samps + n_samps);
          }
      } else if(sparams->amp == 0.) {
          //residual synthesis only
          for(b = 0; b < ATSA_CRITICAL_BANDS; b++) {
              ampl_p  = ENG_RMS(ats_sound->band_energy[b][f_p], ats_sound->window_size)
                  * sparams->ramp;
              ampl_c  = ENG_RMS(ats_sound->band_energy[b][f], ats_sound->window_size)
                  * sparams->ramp;
              ampl_n  = ENG_RMS(ats_sound->band_energy[b][f_n], ats_sound->window_size)
                  * sparams->ramp;
              freq_c  = res_band_centers[b] * sparams->frec;
              frame_samps = synth_residual_only(ampl_p, ampl_c, ampl_n, res_band_centers[p],
                                                time_offset, duration, sparams->sr,
                                                ospt+b, &rarray[p], *out_samps + n_samps);
          }
      } else {
          //residual and deterministic synthesis
          for(d = 0; d < ats_sound->partials; d++) {
              if (selected && !selected[d])
                  continue;
              ampl_p  = ats_sound->amp[d][f_p] * sparams->amp;
              ampl_c  = ats_sound->amp[d][f]   * sparams->amp;
              ampl_n  = ats_sound->amp[d][f_n] * sparams->amp;
              freq_p  = ats_sound->frq[d][f_p] * sparams->frec;
              freq_c  = ats_sound->frq[d][f]   * sparams->frec;
              freq_n  = ats_sound->frq[d][f_n] * sparams->frec;
              pha_p = ats_sound->pha[d][f_p];
              pha_c = ats_sound->pha[d][f];
              pha_n = ats_sound->pha[d][f_n];
              resid_p = ENG_RMS(ats_sound->res[d][f_p] * sparams->ramp, ats_sound->window_size)
                  * sparams->ramp;
              resid_c = ENG_RMS(ats_sound->res[d][f] * sparams->ramp, ats_sound->window_size)
                  * sparams->ramp;
              resid_n = ENG_RMS(ats_sound->res[d][f_n] * sparams->ramp, ats_sound->window_size)
                  * sparams->ramp;
              frame_samps = synth_both(ampl_p, ampl_c, ampl_n, freq_p, freq_c, freq_n,
                                       pha_p, pha_c, pha_n, resid_p, resid_c, resid_n,
                                       time_offset, duration, sparams->sr,
                                       ospt+d, &rarray[p], *out_samps + n_samps);
          }
      }
        
      // Jump into sample buffer for next frame.
      n_samps += (int)frame_samps;
      
      fprintf(stdout, "    frame#%d, samps=%d (%d/%d)\n", f, frame_samps, n_samps, todo);
    }
  }
  
  *n_out_samps = n_samps;
  
  fprintf(stdout, "%d samples generated\n", *n_out_samps);

  if (ospt)
      free(ospt);
  if (rarray)
      free(rarray);
  free(tdata);

  return;
}