path: root/vdr_sound.c

/*!
 * \file vdr_setup.c
 * \brief Sound manipulation classes for a VDR media plugin (muggle)
 *
 * \version $Revision: 1.2 $
 * \date    $Date: 2004/05/28 15:29:19 $
 * \author  Ralf Klueber, Lars von Wedel, Andreas Kellner
 * \author  Responsible author: $Author: lvw $
 *
 * $Id: vdr_sound.c,v 1.2 2004/05/28 15:29:19 lvw Exp $
 *
 * Adapted from 
 * MP3/MPlayer plugin to VDR (C++)
 * (C) 2001,2002 Stefan Huelswitt <huels@iname.com>
 */

// --- cResample ------------------------------------------------------------

// The resample code has been adapted from the madplay project
// (resample.c) found in the libmad distribution
   
class cResample
{
private:
  mad_fixed_t ratio;
  mad_fixed_t step;
  mad_fixed_t last;
  mad_fixed_t resampled[MAX_NSAMPLES];
public:
  bool SetInputRate(unsigned int oldrate, unsigned int newrate);
  unsigned int ResampleBlock(unsigned int nsamples, const mad_fixed_t *old);
  const mad_fixed_t *Resampled(void) { return resampled; }
  };

bool cResample::SetInputRate(unsigned int oldrate, unsigned int newrate)
{
  if(oldrate<8000 || oldrate>newrate*6) { // out of range
    esyslog("WARNING: samplerate %d out of range 8000-%d\n",oldrate,newrate*6);
    return 0;
    }
  ratio=mad_f_tofixed((double)oldrate/(double)newrate);
  step=0; last=0;
#ifdef DEBUG
  static mad_fixed_t oldratio=0;
  if(oldratio!=ratio) {
    printf("mad: new resample ratio %f (from %d kHz to %d kHz)\n",mad_f_todouble(ratio),oldrate,newrate);
    oldratio=ratio;
    }
#endif
  return ratio!=MAD_F_ONE;
}

unsigned int cResample::ResampleBlock(unsigned int nsamples, const mad_fixed_t *old)
{
  // This resampling algorithm is based on a linear interpolation, which is
  // not at all the best sounding but is relatively fast and efficient.
  //
  // A better algorithm would be one that implements a bandlimited
  // interpolation.

  mad_fixed_t *nsam=resampled;
  const mad_fixed_t *end=old+nsamples;
  const mad_fixed_t *begin=nsam;

  if(step < 0) {
    step = mad_f_fracpart(-step);

    while (step < MAD_F_ONE) {
      *nsam++ = step ? last+mad_f_mul(*old-last,step) : last;
      step += ratio;
      if(((step + 0x00000080L) & 0x0fffff00L) == 0)
	step = (step + 0x00000080L) & ~0x0fffffffL;
      }
    step -= MAD_F_ONE;
    }

  while (end - old > 1 + mad_f_intpart(step)) {
    old += mad_f_intpart(step);
    step = mad_f_fracpart(step);
    *nsam++ = step ? *old + mad_f_mul(old[1] - old[0], step) : *old;
    step += ratio;
    if (((step + 0x00000080L) & 0x0fffff00L) == 0)
      step = (step + 0x00000080L) & ~0x0fffffffL;
    }

  if (end - old == 1 + mad_f_intpart(step)) {
    last = end[-1];
    step = -step;
    }
  else step -= mad_f_fromint(end - old);

  return nsam-begin;
}

// --- cLevel ----------------------------------------------------------------

// The normalize algorithm and parts of the code has been adapted from the
// Normalize 0.7 project. (C) 1999-2002, Chris Vaill <cvaill@cs.columbia.edu>

// A little background on how normalize computes the volume
// of a wav file, in case you want to know just how your
// files are being munged:
//
// The volumes calculated are RMS amplitudes, which corre­
// spond (roughly) to perceived volume. Taking the RMS ampli­
// tude of an entire file would not give us quite the measure
// we want, though, because a quiet song punctuated by short
// loud parts would average out to a quiet song, and the
// adjustment we would compute would make the loud parts
// excessively loud.
//
// What we want is to consider the maximum volume of the
// file, and normalize according to that. We break up the
// signal into 100 chunks per second, and get the signal
// power of each chunk, in order to get an estimation of
// "instantaneous power" over time. This "instantaneous
// power" signal varies too much to get a good measure of the
// original signal's maximum sustained power, so we run a
// smoothing algorithm over the power signal (specifically, a
// mean filter with a window width of 100 elements). The max­
// imum point of the smoothed power signal turns out to be a
// good measure of the maximum sustained power of the file.
// We can then take the square root of the power to get maxi­
// mum sustained RMS amplitude.

class cLevel {
private:
  double maxpow;
  mad_fixed_t peak;
  struct Power {
    // smooth
    int npow, wpow;
    double powsum, pows[POW_WIN];
    // sum
    unsigned int nsum;
    double sum;
    } power[2];
  //
  inline void AddPower(struct Power *p, double pow);
public:
  void Init(void);
  void GetPower(struct mad_pcm *pcm);
  double GetLevel(void);
  double GetPeak(void);
  };

void cLevel::Init(void)
{
  for(int l=0 ; l<2 ; l++) {
    struct Power *p=&power[l];
    p->sum=p->powsum=0.0; p->wpow=p->npow=p->nsum=0;
    for(int i=POW_WIN-1 ; i>=0 ; i--) p->pows[i]=0.0;
    }
  maxpow=0.0; peak=0;
}

void cLevel::GetPower(struct mad_pcm *pcm)
{
  for(int i=0 ; i<pcm->channels ; i++) {
    struct Power *p=&power[i];
    mad_fixed_t *data=pcm->samples[i];
    for(int n=pcm->length ; n>0 ; n--) {
      if(*data < -peak) peak = -*data;
      if(*data >  peak) peak =  *data;
      double s=mad_f_todouble(*data++);
      p->sum+=(s*s);
      if(++(p->nsum)>=pcm->samplerate/100) {
        AddPower(p,p->sum/(double)p->nsum);
        p->sum=0.0; p->nsum=0;
        }
      }
    }
}

void cLevel::AddPower(struct Power *p, double pow)
{
  p->powsum+=pow;
  if(p->npow>=POW_WIN) {
    if(p->powsum>maxpow) maxpow=p->powsum;
    p->powsum-=p->pows[p->wpow];
    }
  else p->npow++;
  p->pows[p->wpow]=pow;
  p->wpow=(p->wpow+1) % POW_WIN;
}

double cLevel::GetLevel(void)
{
  if(maxpow<EPSILON) {
    // Either this whole file has zero power, or was too short to ever
    // fill the smoothing buffer.  In the latter case, we need to just
    // get maxpow from whatever data we did collect.

    if(power[0].powsum>maxpow) maxpow=power[0].powsum;
    if(power[1].powsum>maxpow) maxpow=power[1].powsum;
    }
  double level=sqrt(maxpow/(double)POW_WIN);     // adjust for the smoothing window size and root
  printf("norm: new volumen level=%f peak=%f\n",level,mad_f_todouble(peak));
  return level;
}

double cLevel::GetPeak(void)
{
  return mad_f_todouble(peak);
}

// --- cNormalize ------------------------------------------------------------

class cNormalize {
private:
  mad_fixed_t gain;
  double d_limlvl, one_limlvl;
  mad_fixed_t limlvl;
  bool dogain, dolimit;
#ifdef DEBUG
  // stats
  unsigned long limited, clipped, total;
  mad_fixed_t peak;
#endif
  // limiter
#ifdef USE_FAST_LIMITER
  mad_fixed_t *table, tablestart;
  int tablesize;
  inline mad_fixed_t FastLimiter(mad_fixed_t x);
#endif
  inline mad_fixed_t Limiter(mad_fixed_t x);
public:
  cNormalize(void);
  ~cNormalize();
  void Init(double Level, double Peak);
  void Stats(void);
  void AddGain(struct mad_pcm *pcm);
  };

cNormalize::cNormalize(void)
{
  d_limlvl = (double)the_setup.LimiterLevel / 100.0;
  one_limlvl = 1 - d_limlvl;
  limlvl = mad_f_tofixed(d_limlvl);
  printf( "norm: lim_lev=%f lim_acc=%d\n", d_limlvl, LIM_ACC );

#ifdef USE_FAST_LIMITER
  mad_fixed_t start=limlvl & ~(F_LIM_JMP-1);
  tablestart=start;
  tablesize=(unsigned int)(F_LIM_MAX-start)/F_LIM_JMP + 2;
  table=new mad_fixed_t[tablesize];
  if(table) {
    printf("norm: table size=%d start=%08x jump=%08x\n",tablesize,start,F_LIM_JMP);
    for(int i=0 ; i<tablesize ; i++) {
      table[i]=Limiter(start);
      start+=F_LIM_JMP;
      }
    tablesize--; // avoid a -1 in FastLimiter()

    // do a quick accuracy check, just to be sure that FastLimiter() is working
    // as expected :-)
#ifdef ACC_DUMP
    FILE *out=fopen("/tmp/limiter","w");
#endif
    mad_fixed_t maxdiff=0;
    for(mad_fixed_t x=F_LIM_MAX ; x>=limlvl ; x-=mad_f_tofixed(1e-4)) {
      mad_fixed_t diff=mad_f_abs(Limiter(x)-FastLimiter(x));
      if(diff>maxdiff) maxdiff=diff;
#ifdef ACC_DUMP
      fprintf(out,"%0.10f\t%0.10f\t%0.10f\t%0.10f\t%0.10f\n",
        mad_f_todouble(x),mad_f_todouble(Limiter(x)),mad_f_todouble(FastLimiter(x)),mad_f_todouble(diff),mad_f_todouble(maxdiff));
      if(ferror(out)) break;
#endif
      }
#ifdef ACC_DUMP
    fclose(out);
#endif
    printf("norm: accuracy %.12f\n",mad_f_todouble(maxdiff));
    if(mad_f_todouble(maxdiff)>1e-6) 
      {
	esyslog("ERROR: accuracy check failed, normalizer disabled");
	delete table; table=0;
      }
    }
  else esyslog("ERROR: no memory for lookup table, normalizer disabled");
#endif // USE_FAST_LIMITER
}

cNormalize::~cNormalize()
{
#ifdef USE_FAST_LIMITER
  delete table;
#endif
}

void cNormalize::Init(double Level, double Peak)
{
  double Target=(double)the_setup.TargetLevel/100.0;
  double dgain=Target/Level;
  if(dgain>MAX_GAIN) dgain=MAX_GAIN;
  gain=mad_f_tofixed(dgain);
  // Check if we actually need to apply a gain
  dogain=(Target>0.0 && fabs(1-dgain)>MIN_GAIN);
#ifdef USE_FAST_LIMITER
  if(!table) dogain=false;
#endif
  // Check if we actually need to do limiting:
  // we have to if limiter is enabled, if gain>1 and if the peaks will clip.
  dolimit=(d_limlvl<1.0 && dgain>1.0 && Peak*dgain>1.0);
#ifdef DEBUG
  printf("norm: gain=%f dogain=%d dolimit=%d (target=%f level=%f peak=%f)\n",dgain,dogain,dolimit,Target,Level,Peak);
  limited=clipped=total=0; peak=0;
#endif
}

void cNormalize::Stats(void)
{
#ifdef DEBUG
  if(total)
    printf("norm: stats tot=%ld lim=%ld/%.3f%% clip=%ld/%.3f%% peak=%.3f\n",
           total,limited,(double)limited/total*100.0,clipped,(double)clipped/total*100.0,mad_f_todouble(peak));
#endif
}

mad_fixed_t cNormalize::Limiter(mad_fixed_t x)
{
// Limiter function:
//
//        / x                                                (for x <= lev)
//   x' = |
//        \ tanh((x - lev) / (1-lev)) * (1-lev) + lev        (for x > lev)
//
// call only with x>=0. For negative samples, preserve sign outside this function
//
// With limiter level = 0, this is equivalent to a tanh() function;
// with limiter level = 1, this is equivalent to clipping.

  if(x>limlvl) {
#ifdef DEBUG
    if(x>MAD_F_ONE) clipped++;
    limited++;
#endif
    x=mad_f_tofixed(tanh((mad_f_todouble(x)-d_limlvl) / one_limlvl) * one_limlvl + d_limlvl);
    }
  return x;
}

#ifdef USE_FAST_LIMITER
mad_fixed_t cNormalize::FastLimiter(mad_fixed_t x)
{
// The fast algorithm is based on a linear interpolation between the
// the values in the lookup table. Relays heavly on libmads fixed point format.

  if(x>limlvl) {
    int i=(unsigned int)(x-tablestart)/F_LIM_JMP;
#ifdef DEBUG
    if(x>MAD_F_ONE) clipped++;
    limited++;
    if(i>=tablesize) printf("norm: overflow x=%f x-ts=%f i=%d tsize=%d\n",
                            mad_f_todouble(x),mad_f_todouble(x-tablestart),i,tablesize);
#endif
    mad_fixed_t r=x & (F_LIM_JMP-1);
    x=MAD_F_ONE;
    if(i<tablesize) {
      mad_fixed_t *ptr=&table[i];
      x=*ptr;
      mad_fixed_t d=*(ptr+1)-x;
      //x+=mad_f_mul(d,r)<<LIM_ACC;                // this is not accurate as mad_f_mul() does >>MAD_F_FRACBITS
                                                   // which is senseless in the case of following <<LIM_ACC.
      x+=((long long)d*(long long)r)>>LIM_SHIFT;   // better, don't know if works on all machines
      }
    }
  return x;
}
#endif

#ifdef USE_FAST_LIMITER
#define LIMITER_FUNC FastLimiter
#else
#define LIMITER_FUNC Limiter
#endif

void cNormalize::AddGain(struct mad_pcm *pcm)
{
  if(dogain) {
    for(int i=0 ; i<pcm->channels ; i++) {
      mad_fixed_t *data=pcm->samples[i];
#ifdef DEBUG
      total+=pcm->length;
#endif
      if(dolimit) {
        for(int n=pcm->length ; n>0 ; n--) {
          mad_fixed_t s=mad_f_mul(*data,gain);
          if(s<0) {
            s=-s;
#ifdef DEBUG
            if(s>peak) peak=s;
#endif
            s=LIMITER_FUNC(s);
            s=-s;
            }
          else {
#ifdef DEBUG
            if(s>peak) peak=s;
#endif
            s=LIMITER_FUNC(s);
            }
          *data++=s;
          }
        }
      else {
        for(int n=pcm->length ; n>0 ; n--) {
          mad_fixed_t s=mad_f_mul(*data,gain);
#ifdef DEBUG
          if(s>peak) peak=s;
          else if(-s>peak) peak=-s;
#endif
          if(s>MAD_F_ONE) s=MAD_F_ONE;   // do clipping
          if(s<-MAD_F_ONE) s=-MAD_F_ONE;
          *data++=s;
          }
        }
      }
    }
}

// --- cScale ----------------------------------------------------------------

// The dither code has been adapted from the madplay project
// (audio.c) found in the libmad distribution

enum eAudioMode { amRound, amDither };

class cScale {
private:
  enum { MIN=-MAD_F_ONE, MAX=MAD_F_ONE - 1 };
#ifdef DEBUG
  // audio stats
  unsigned long clipped_samples;
  mad_fixed_t peak_clipping;
  mad_fixed_t peak_sample;
#endif
  // dither
  struct dither {
    mad_fixed_t error[3];
    mad_fixed_t random;
    } leftD, rightD;
  //
  inline mad_fixed_t Clip(mad_fixed_t sample, bool stats=true);
  inline signed long LinearRound(mad_fixed_t sample);
  inline unsigned long Prng(unsigned long state);
  inline signed long LinearDither(mad_fixed_t sample, struct dither *dither);
public:
  void Init(void);
  void Stats(void);
  unsigned int ScaleBlock(unsigned char *data, unsigned int size, unsigned int &nsamples, const mad_fixed_t * &left, const mad_fixed_t * &right, eAudioMode mode);
  };

void cScale::Init(void)
{
#ifdef DEBUG
  clipped_samples=0; peak_clipping=peak_sample=0;
#endif
  memset(&leftD,0,sizeof(leftD));
  memset(&rightD,0,sizeof(rightD));
}

void cScale::Stats(void)
{
#ifdef DEBUG
  printf("mp3: scale stats clipped=%ld peak_clip=%f peak=%f\n",
         clipped_samples,mad_f_todouble(peak_clipping),mad_f_todouble(peak_sample));
#endif
}

// gather signal statistics while clipping
mad_fixed_t cScale::Clip(mad_fixed_t sample, bool stats)
{
#ifndef DEBUG
  if (sample > MAX) sample = MAX;
  if (sample < MIN) sample = MIN;
#else
  if(!stats) {
    if (sample > MAX) sample = MAX;
    if (sample < MIN) sample = MIN;
    }
  else {
    if (sample >= peak_sample) {
      if (sample > MAX) {
        ++clipped_samples;
        if (sample - MAX > peak_clipping)
	  peak_clipping = sample - MAX;
        sample = MAX;
        }
      peak_sample = sample;
      }
    else if (sample < -peak_sample) {
      if (sample < MIN) {
        ++clipped_samples;
        if (MIN - sample > peak_clipping)
	  peak_clipping = MIN - sample;
        sample = MIN;
        }
      peak_sample = -sample;
      }
    }
#endif
  return sample;
}

// generic linear sample quantize routine
signed long cScale::LinearRound(mad_fixed_t sample)
{
  // round
  sample += (1L << (MAD_F_FRACBITS - OUT_BITS));
  // clip
  sample=Clip(sample);
  // quantize and scale
  return sample >> (MAD_F_FRACBITS + 1 - OUT_BITS);
}

// 32-bit pseudo-random number generator
unsigned long cScale::Prng(unsigned long state)
{
  return (state * 0x0019660dL + 0x3c6ef35fL) & 0xffffffffL;
}

// generic linear sample quantize and dither routine
signed long cScale::LinearDither(mad_fixed_t sample, struct dither *dither)
{
  unsigned int scalebits;
  mad_fixed_t output, mask, random;

  // noise shape
  sample += dither->error[0] - dither->error[1] + dither->error[2];
  dither->error[2] = dither->error[1];
  dither->error[1] = dither->error[0] / 2;
  // bias
  output = sample + (1L << (MAD_F_FRACBITS + 1 - OUT_BITS - 1));
  scalebits = MAD_F_FRACBITS + 1 - OUT_BITS;
  mask = (1L << scalebits) - 1;
  // dither
  random  = Prng(dither->random);
  output += (random & mask) - (dither->random & mask);
  dither->random = random;
  // clip
  output=Clip(output);
  sample=Clip(sample,false);
  // quantize
  output &= ~mask;
  // error feedback
  dither->error[0] = sample - output;
  // scale
  return output >> scalebits;
}

// write a block of signed 16-bit big-endian PCM samples
unsigned int cScale::ScaleBlock(unsigned char *data, unsigned int size, unsigned int &nsamples, const mad_fixed_t * &left, const mad_fixed_t * &right, eAudioMode mode)
{
  signed int sample;
  unsigned int len, res;

  len=size/OUT_FACT; res=size;
  if(len>nsamples) { len=nsamples; res=len*OUT_FACT; }
  nsamples-=len;

  if(right) {  // stereo
    switch (mode) {
      case amRound:
        while (len--) {
          sample  = LinearRound(*left++);
          *data++ = sample >> 8;
          *data++ = sample >> 0;
          sample  = LinearRound(*right++);
          *data++ = sample >> 8;
          *data++ = sample >> 0;
          }
        break;
      case amDither:
        while (len--) {
	  sample  = LinearDither(*left++,&leftD);
	  *data++ = sample >> 8;
	  *data++ = sample >> 0;
	  sample  = LinearDither(*right++,&rightD);
	  *data++ = sample >> 8;
	  *data++ = sample >> 0;
          }
        break;
      }
    }
  else {  // mono, duplicate left channel
    switch (mode) {
      case amRound:
        while (len--) {
	  sample  = LinearRound(*left++);
	  *data++ = sample >> 8;
	  *data++ = sample >> 0;
	  *data++ = sample >> 8;
	  *data++ = sample >> 0;
          }
        break;
      case amDither:
        while (len--) {
	  sample  = LinearDither(*left++,&leftD);
	  *data++ = sample >> 8;
	  *data++ = sample >> 0;
	  *data++ = sample >> 8;
	  *data++ = sample >> 0;
          }
        break;
      }
    }
 return res;
}