/*
	rebuffer.template -- this is a TEMPLATE for a rebuffer module

	since C has no templates I use this method to get a type-independent
	rebuffering module.  This avoids enormous amounts of duplicated code.

	to use this file, #include it with the following tokens defined:

	#define PRFX(word) XXXX ## word
	#define INITFUNC(word) initXXXX ## word
	#define NAMESTRING(word) "XXXX" #word
	#define TYPE YYYY

	where XXXX is replaced with the prefix you want for the module
	name and YYYY is replaced with the type the routine works with.
	All the other stuff above stays the same.  Here is an example:

	#define PRFX(word) int2 ## word
	#define INITFUNC(word) initint2 ## word
	#define NAMESTRING(word) "int2" #word
	#define TYPE signed short int
	#include "rebuffer.template"

	I would just have you set two variables and do the above myself,
	but the C preprocessor is so brain dead that it does not evaluate
	actual arguments to preprocessor functions until after they are
	substituted in! (and thus concatenated, and thus unable to be
	substituted).

	The alternative to this (other than to use C++, which would make this
	whole damn library so much easier to write and to read) is to have
	large numbers of almost identical copies of the rebuffering routines
	with a few type fields different.  This is to be avoided.

	This code will then define a module called (using above example)
	int2rebufferv2f and another called int2rebufferf2v.
*/

/* instance data of rebuffering modules */

typedef struct PRFX(rebuffer_data)
{
	bool multiple_of;	/* TRUE=buffer to multiple-of not exactly... */
	AFframecount nsamps;	/* # of fixed samples / multiple-of */
	TYPE *buf;		/* buf of nsamps samples */
	long offset;		/* see code for meaning */
	bool eof;		/* (pull only) end of input stream reached */
	bool sent_short_chunk;	/* (pull only) end of output stream indicated */

	TYPE *saved_buf;	/* (push only) saved buffer */
	long saved_offset;	/* (push only) saved buffer offset */
} PRFX(rebuffer_data);

#define REBUFFER_DATA PRFX(rebuffer_data)

/* REBUFFER variable to fixed -- PUSH ONLY */

static void PRFX(rebufferv2fmax_push)(struct _AFmoduleinst *i)
{
  REBUFFER_DATA *d = i->modspec;
  int nframes = d->nsamps/i->inc->f.channelCount;

  if (d->multiple_of)
    /* actually, it can be less, but this is ok */
    i->outc->nframes = i->inc->nframes + nframes;
  else
    i->outc->nframes = nframes;
}

static void PRFX(rebufferv2frun_push)(struct _AFmoduleinst *i)
{
  REBUFFER_DATA *d = i->modspec;
  long samples2push = i->inc->nframes * i->inc->f.channelCount;
  TYPE *inbufp = i->inc->buf;

  /* d->offset could be from 0 to nsamps. 0 is no saved samples,
     nsamps is all saved samples (should never be nsamps though)
     offset thus contains a count of the valid data samples */

  assert(d->offset >= 0 && d->offset < d->nsamps);

  /* check that we will be able to push even one block */
  if (d->offset + samples2push >= d->nsamps)
    {
      /* push the currently buffered samples through */
      if (d->offset != 0)
        memcpy(i->outc->buf, d->buf, sizeof(TYPE)*d->offset);

      if (d->multiple_of)
        {
          /* round down to nearest d->nsamps */
          int n = (((d->offset+samples2push) / d->nsamps) * d->nsamps);

          /*
             Example with d->nsamps==5, d->offset==3, and samples2push=10.
             B=a buffered sample
             N=a new sample (pushed into module on this call)
             Note that we group samples in groups of d->nsamps since they must
             be pushed that way (any # of groups at a time):

              offset  samples2push
                 |-||------------|
                 BBBNN  NNNNN  NNN....
                    |-------|
                     n-offset
                 |----------|
                       n

             n is the number of samples we will push on this run.  What's
             left over (3 samples in this case) will get buffered for use
             in the next run.
          */

          assert(n > d->offset);
          memcpy((TYPE *)i->outc->buf + d->offset, inbufp,
                 sizeof(TYPE)*(n - d->offset)); /* fill in rest of outbuf */

          _AFpush(i, n / i->outc->f.channelCount);

          inbufp += n - d->offset;
          samples2push -= n - d->offset;
          assert(samples2push >= 0);
          d->offset = 0;        /* the buffer is now empty */
        }
      else
        {
          while (d->offset + samples2push >= d->nsamps)
            {
              int n = d->nsamps - d->offset;
              /*
                 Same example as above.  Everything is the same except now
                 we can only push one group of d->nsamps samples at a time,
                 so we must loop through groups, and n takes on several
                 values:

                    offset  samples2push
                       |-||------------|
                       BBBNN  NNNNN  NNN....
                          ||  |---|
                           n    n

                 n says how many samples to take from the input chunk.
                 n first has the value d->nsamps - d->offset, and then it
                 always has the value d->nsamps.
              */
              memcpy((TYPE *)i->outc->buf + d->offset, inbufp,
                     sizeof(TYPE)*n); /* fill in rest of outbuf */

              _AFpush(i, d->nsamps / i->outc->f.channelCount);

              inbufp += n;
              samples2push -= n;
              assert(samples2push >= 0);
              d->offset = 0;                /* clear out the buffer */
            }
        }

      /* if we pushed blocks, then we must have used all stored samples */
      assert(d->offset == 0);
    }

  /* at this point: guaranteed that  d->offset + samples2push < d->nsamps */
  assert(d->offset + samples2push < d->nsamps);

  /* save remaining samples in buffer */
  if ( samples2push != 0 )
    {
      memcpy(d->buf+d->offset, inbufp, sizeof(TYPE)*samples2push);
      d->offset += samples2push;
    }

  assert(d->offset >= 0 && d->offset < d->nsamps);
}

static void PRFX(rebufferv2fsync1)(struct _AFmoduleinst *i)
{
  REBUFFER_DATA *d = i->modspec;

  assert(d->offset >= 0 && d->offset < d->nsamps);

  /*
    save all the samples and the offset so we can
     restore our state later.
  */
  memcpy(d->saved_buf, d->buf, sizeof(TYPE)*d->nsamps);
  d->saved_offset = d->offset;
}

static void PRFX(rebufferv2fsync2)(struct _AFmoduleinst *i)
{
  REBUFFER_DATA *d = i->modspec;

  /* d->offset could be from 0 to nsamps. 0 is no saved samples,
     nsamps is all saved samples (should never be nsamps though)
     offset thus contains a count of the valid data samples */

  assert(d->offset >= 0 && d->offset < d->nsamps);

  /*
     push the currently buffered samples through--even if there
     are none!

     in other words, push a SHORT CHUNK -- see modules.c TOF for more info !
  */
  if (d->offset != 0)
    {
      memcpy(i->outc->buf, d->buf, sizeof(TYPE)*d->offset);
    }

  _AFpush(i, d->offset / i->outc->f.channelCount); /* SHORT CHUNK */

  /* restore state saved in sync1 */

  memcpy(d->buf, d->saved_buf, sizeof(TYPE)*d->nsamps);
  d->offset = d->saved_offset;

  assert(d->offset >= 0 && d->offset < d->nsamps);
}

static void PRFX(rebufferv2ffree)(struct _AFmoduleinst *i)
{
  REBUFFER_DATA *d = i->modspec;
  if (d->buf)
    free(d->buf);
  if (d->saved_buf)
    free(d->saved_buf);
}

static _AFmodule PRFX(rebufferv2f) =
{
  NAMESTRING(rebufferv2f),
  AF_NULL,
  AF_NULL, PRFX(rebufferv2fmax_push),
  AF_NULL, AF_NULL, AF_NULL,
  PRFX(rebufferv2frun_push), PRFX(rebufferv2fsync1), PRFX(rebufferv2fsync2),
  AF_NULL,
  PRFX(rebufferv2ffree)
};

_AFmoduleinst INITFUNC(rebufferv2f)(AFframecount nsamps, bool multiple_of)
{
  _AFmoduleinst ret = _AFnewmodinst(&PRFX(rebufferv2f));
  REBUFFER_DATA *d = _af_malloc(sizeof (REBUFFER_DATA));
  d->nsamps = nsamps;
  d->offset = 0;
  d->buf = _af_malloc(sizeof (TYPE) * nsamps);
  d->multiple_of = multiple_of;
  d->saved_buf = _af_malloc(sizeof (TYPE) * nsamps);
  ret.modspec = d;

  DEBG(printf("%s nsamps=%d multiple_of=%d\n", NAMESTRING(rebufferv2f),
              nsamps, multiple_of));

  return(ret);
}

/* ===== REBUFFER fixed to variable -- PULL ONLY */

static void PRFX(rebufferf2vmax_pull)(struct _AFmoduleinst *i)
{
  REBUFFER_DATA *d = i->modspec;
  long nframes = d->nsamps / i->inc->f.channelCount;

  if (d->multiple_of)
    /* actually, it can be less, but this is ok */
    i->inc->nframes = i->outc->nframes + nframes;
  else
    i->inc->nframes = nframes;
}

static void PRFX(rebufferf2vrun_pull)(struct _AFmoduleinst *i)
{
  REBUFFER_DATA *d = i->modspec;
  long samples2pull = i->outc->nframes * i->outc->f.channelCount;
  TYPE *outbufp = i->outc->buf;

  /* d->offset could be from 0 to nsamps. 0=saved all samples,
     nsamps=saved no samples (should never be 0 though)
     offset is thus offset of first valid sample in buffer
  */
  assert(d->offset > 0 && d->offset <= d->nsamps);

  /*
     as explained TOF modules.c, a module should not pull more
     frames from its input after receiving the short chunk,
     otherwise it is an error.  This checks if the next module
     in the chain is attempting to do so:
  */
  assert(!d->sent_short_chunk);

  /*
     Here is a way of thinking about the stream of data being pulled
     from this module (say d->nsamps=5, d->offset=2, samples2pull=11).
     We arrange samples in groups of d->nsamps as they arrived from the
     input stream (since the chunk size constraint comes only from
     that side):

     X=old data long gone
     B=data we got from d->buf
     N=new data we have to pull from our input module

                  samples2pull=11
                >-------------|
       XXXXX  XXBBB  NNNNN  NNNNN
       |---|  |---|  |---|  |---|
       nsamps nsamps nsamps nsamps

     Think of samples2pull as a distance which shrinks as we satisfy the
     request.  The way it shrinks is that the left side of it (shown as '>'
     above) moves gradually left until it meets the fixed right side
     (shown as '|'), at which time we have completely satisfied the
     output's request.

     As the diagram indicates, some of those samples are buffered up from
     the last run (BBB), and some of those samples have to be pulled
     from our input right now.

     If the '>' moves PAST the '|' to the right, then this means we have pulled
     more samples that we needed for the output's request, and so we should
     buffer up those samples for the next run_pull.
  */


  /* ----- first try and use currently buffered samples */

  CHNK(printf("%*.*s%s buffer holds %d samples\n",
              i->margin, i->margin, "", i->mod->name,
              d->nsamps-d->offset));

  if (d->offset != d->nsamps)
    {
      int buffered = d->nsamps - d->offset;
      int n = min(samples2pull, buffered);
      memcpy(outbufp, d->buf+d->offset, sizeof(TYPE)*n);
      CHNK(printf("%*.*s%s taking %d currently buffered samples\n",
                  i->margin, i->margin, "", i->mod->name,
                  n));
      /*
        this may make samples2pull negative and outbufp invalid:
        that's ok, it means we have some samples left over after
        satisfying the output's request
      */
      outbufp += buffered;
      samples2pull -= buffered;
      /*
         we have taken n samples from the buffer so we indicate
         that the buffer is that much smaller
      */
      d->offset += n;
    }

  CHNK(printf("%*.*s%s now samples2pull=%d\n",
              i->margin, i->margin, "", i->mod->name,
              samples2pull));

  /*
     at this point the picture from above looks something like:

                   samples2pull=8
                     >--------|
       XXXXX  XXBBB  NNNNN  NNNNN
       |---|  |---|  |---|  |---|
       nsamps nsamps nsamps nsamps

     note that samples2pull has shrunk by the number of samples we
     had buffered up.

     If it happened that samples2pull was originally less than the
     number of samples we have buffered up, as in

            samples2pull=2
                >|
       XXXXX  XXBBB
       |---|  |---|
       nsamps nsamps

     Then at this point things look like:

            samples2pull=-1
                 |>
       XXXXX  XXBBB
       |---|  |---|
       nsamps nsamps

     Which is just fine.  It means we should re-save that 1 sample for
     later use.
  */

  /* ----- then try and pull more samples from the source */

  while (!d->eof && samples2pull > 0)
    {
      int req, got;

      /*
         req is the number of "N" frames (from the pictures above) which
         we want to pull at a time.
      */
      if (d->multiple_of)
        /* round samples2pull up to nearest d->nsamps */
        req = ( ( ((samples2pull-1)/d->nsamps) + 1) * d->nsamps);
      else
        req = d->nsamps;

      assert(req > 0);

      /* pull chunk(s) of data from source */

      _AFpull(i, req / i->inc->f.channelCount);

      got = i->inc->nframes * i->inc->f.channelCount; /* may be short chunk */

      if (got != req)       /* short chunk on input */
        d->eof = AF_TRUE;

      memcpy(outbufp, i->inc->buf, sizeof(TYPE)*min(samples2pull, got));

      /*
        this may make samples2pull negative and outbufp invalid:
        that's ok, it means we have some samples left over after
        satisfying the output's request
      */
      outbufp += got;
      samples2pull -= got;

      /* we should only do loop once for multiple_of */

      if (d->multiple_of)
        assert(d->eof || samples2pull <= 0);

      if (samples2pull < 0)
        {
          /*
             we pulled more frames than needed for the user's request
             (and the loop is about to exit) so save the remaining samples
             in d->buf.
          */
          assert(d->offset==d->nsamps); /* if we pulled, buffer must be used up */

          /* samples2pull is -(the number of samples the next mod didn't want) */
          d->offset = d->nsamps + samples2pull;

          assert(d->offset > 0 && d->offset <= d->nsamps);

          memcpy(d->buf+d->offset, (TYPE *)i->inc->buf+d->offset,
                 sizeof(TYPE)*(d->nsamps-d->offset));
        }
      else
        {
          assert(d->offset==d->nsamps); /* if we pulled, buffer must be used up */
        }
    }

  CHNK(printf("%*.*s%s ... and now samples2pull=%d\n",
              i->margin, i->margin, "", i->mod->name,
              samples2pull));

  /* -----
     at this point we have done all we can to try and satisfy the output's
     request.  We may have hit eof on input.

     If samples2pull <= 0, we have pulled enough samples to satisfy the
     request.

     If d->eof==AF_TRUE, then we hit EOF on input.

     If we did not hit end of file, then we must have satisfied the request.
     If we did hit end of file, then we may or may not have satisfied the request.

     If we hit eof and we did not satisfy the request, then we push the
     SHORT CHUNK, after which the module on our output is not allowed to
     pull any more samples.

     Otherwise we save any samples left over into d->buf.

     Note that just because we hit EOF on input doesn't mean we're going to
     push the short chunk on output right away.  Because we buffer up samples
     internally, there could be any number of run_push calls between hitting
     eof on input and sending the short chunk on output.  d->eof indicates
     the first condition, d->send_short_chunk indicates the second.
  */

  if (d->eof && samples2pull > 0) /* EOF reached and not enough data */
    {
      i->outc->nframes -= samples2pull / i->inc->f.channelCount; /* SHORT CHUNK out */
      d->sent_short_chunk = AF_TRUE;
      assert(d->offset == d->nsamps); /* we must have used all buffered frames */
    }
  else
    {
      /* !d->eof && samples2pull > 0 is impossible */
      assert(samples2pull <= 0);

      /* samples2pull is -(the number of samples the next mod didn't want) */
      assert(d->offset == d->nsamps + samples2pull);
    }
  assert(d->offset > 0 && d->offset <= d->nsamps);

  CHNK(printf("%*.*s%s ... and now buffer holds %d samples\n",
              i->margin, i->margin, "", i->mod->name,
              d->nsamps-d->offset));
}

static void PRFX(rebufferf2vreset1)(struct _AFmoduleinst *i)
{
  REBUFFER_DATA *d = i->modspec;
  d->offset = d->nsamps;
  d->eof = AF_FALSE;
  d->sent_short_chunk = AF_FALSE;
  assert(d->offset > 0 && d->offset <= d->nsamps);
}

static void PRFX(rebufferf2vreset2)(struct _AFmoduleinst *i)
{
#ifdef AF_DEBUG
  REBUFFER_DATA *d = i->modspec;
  assert(d->offset > 0 && d->offset <= d->nsamps);
#endif
}

static void PRFX(rebufferf2vfree)(struct _AFmoduleinst *i)
{
  REBUFFER_DATA *d = i->modspec;
  if (d->buf)
    free(d->buf);
}

static _AFmodule PRFX(rebufferf2v) =
{
  NAMESTRING(rebufferf2v),
  AF_NULL,
  PRFX(rebufferf2vmax_pull), AF_NULL,
  PRFX(rebufferf2vrun_pull), PRFX(rebufferf2vreset1), PRFX(rebufferf2vreset2),
  AF_NULL, AF_NULL, AF_NULL,
  AF_NULL,
  PRFX(rebufferf2vfree)
};

_AFmoduleinst INITFUNC(rebufferf2v)(AFframecount nsamps, bool multiple_of)
{
  _AFmoduleinst ret = _AFnewmodinst(&PRFX(rebufferf2v));
  REBUFFER_DATA *d = _af_malloc(sizeof (REBUFFER_DATA));
  d->nsamps = nsamps;
  d->offset = d->nsamps;
  d->buf = _af_malloc(sizeof (TYPE) * nsamps);
  d->multiple_of = multiple_of;
  d->eof = AF_FALSE;
  d->sent_short_chunk = AF_FALSE;
  ret.modspec = d;

  DEBG(printf("%s nsamps=%d multiple_of=%d\n", NAMESTRING(rebufferf2v),
              nsamps, multiple_of));

  return(ret);
}