/*!
 ***********************************************************************
 * \file image.c
 *
 * \brief
 *    Decode a Slice
 *
 * \author
 *    Main contributors (see contributors.h for copyright, address and affiliation details)
 *    - Inge Lille-Lang鴜               <inge.lille-langoy@telenor.com>
 *    - Rickard Sjoberg                 <rickard.sjoberg@era.ericsson.se>
 *    - Jani Lainema                    <jani.lainema@nokia.com>
 *    - Sebastian Purreiter             <sebastian.purreiter@mch.siemens.de>
 *    - Byeong-Moon Jeon                <jeonbm@lge.com>
 *    - Thomas Wedi                     <wedi@tnt.uni-hannover.de>
 *    - Gabi Blaettermann               <blaetter@hhi.de>
 *    - Ye-Kui Wang                     <wyk@ieee.org>
 *    - Antti Hallapuro                 <antti.hallapuro@nokia.com>
 *    - Alexis Tourapis                 <alexismt@ieee.org>
 ***********************************************************************
 */

#include "contributors.h"

#include <math.h>
#include <limits.h>
#include <stdlib.h>
#include <string.h>
#include <assert.h>

#ifdef WIN32
#include <io.h>
#else
#include <unistd.h>
#endif

#include "global.h"
#include "errorconcealment.h"
#include "image.h"
#include "mbuffer.h"
#include "fmo.h"
#include "nalu.h"
#include "parsetcommon.h"
#include "parset.h"
#include "header.h"
#include "rtp.h"
#include "sei.h"
#include "output.h"
#include "biaridecod.h"
#include "mb_access.h"
#include "memalloc.h"
#include "annexb.h"

#include "context_ini.h"
#include "cabac.h"
#include "loopfilter.h"

#include "vlc.h"

#include "erc_api.h"
extern objectBuffer_t *erc_object_list;
extern ercVariables_t *erc_errorVar;
extern frame erc_recfr;
extern int erc_mvperMB;
extern struct img_par *erc_img;

//extern FILE *p_out2;

extern StorablePicture **listX[6];
extern ColocatedParams *Co_located;

StorablePicture *dec_picture;

OldSliceParams old_slice;

void MbAffPostProc()
{
  imgpel temp[16][32];

  imgpel ** imgY  = dec_picture->imgY;
  imgpel ***imgUV = dec_picture->imgUV;

  int i, x, y, x0, y0, uv;
  for (i=0; i<(int)dec_picture->PicSizeInMbs; i+=2)
  {
    if (dec_picture->mb_field[i])
    {
      get_mb_pos(i, &x0, &y0);
      for (y=0; y<(2*MB_BLOCK_SIZE);y++)
        for (x=0; x<MB_BLOCK_SIZE; x++)
          temp[x][y] = imgY[y0+y][x0+x];

      for (y=0; y<MB_BLOCK_SIZE;y++)
        for (x=0; x<MB_BLOCK_SIZE; x++)
        {
          imgY[y0+(2*y)][x0+x]   = temp[x][y];
          imgY[y0+(2*y+1)][x0+x] = temp[x][y+MB_BLOCK_SIZE];
        }

      x0 = x0 / (16/img->mb_cr_size_x);
      y0 = y0 / (16/img->mb_cr_size_y);

      for (uv=0; uv<2; uv++)
      {
        for (y=0; y<(2*img->mb_cr_size_y);y++)
          for (x=0; x<img->mb_cr_size_x; x++)
            temp[x][y] = imgUV[uv][y0+y][x0+x];
          
        for (y=0; y<img->mb_cr_size_y;y++)
          for (x=0; x<img->mb_cr_size_x; x++)
          {
            imgUV[uv][y0+(2*y)][x0+x]   = temp[x][y];
            imgUV[uv][y0+(2*y+1)][x0+x] = temp[x][y+img->mb_cr_size_y];
          }
      }
    }
  }
}

/*!
 ***********************************************************************
 * \brief
 *    decodes one I- or P-frame
 *
 ***********************************************************************
 */

int decode_one_frame(struct img_par *img,struct inp_par *inp, struct snr_par *snr)
{
  int current_header;
  Slice *currSlice = img->currentSlice;

  img->current_slice_nr = 0;
  img->current_mb_nr = -4711;     // initialized to an impossible value for debugging -- correct value is taken from slice header
  currSlice->next_header = -8888; // initialized to an impossible value for debugging -- correct value is taken from slice header
  img->num_dec_mb = 0;
  img->newframe = 1;

  while ((currSlice->next_header != EOS && currSlice->next_header != SOP))
  {
    current_header = read_new_slice();

    if (current_header == EOS)
    {
      exit_picture();
      return EOS;
    }

    decode_slice(img, inp, current_header);

    img->newframe = 0;
    img->current_slice_nr++;
  }

  exit_picture();

  return (SOP);
}


/*!
************************************************************************
* \brief
*    Find PSNR for all three components.Compare decoded frame with
*    the original sequence. Read inp->jumpd frames to reflect frame skipping.
************************************************************************
*/
void find_snr(
              struct snr_par  *snr,   //!< pointer to snr parameters
              StorablePicture *p,     //!< picture to be compared
              int p_ref)              //!< open reference YUV file
{
  int i,j;
  int64 diff_y,diff_u,diff_v;
  int uv;
  int64  status;
  int symbol_size_in_bytes = img->pic_unit_bitsize_on_disk/8;
  int64 framesize_in_bytes = (((int64)p->size_y*p->size_x) + ((int64)p->size_y_cr*p->size_x_cr)*2) * symbol_size_in_bytes;
  unsigned int max_pix_value_sqd = img->max_imgpel_value * img->max_imgpel_value;
  unsigned int max_pix_value_sqd_uv = img->max_imgpel_value_uv * img->max_imgpel_value_uv; // WSKIM SAMSUNG AIT
  Boolean rgb_output = (active_sps->vui_seq_parameters.matrix_coefficients==0);  // S.SUN 18072004
  
  char *buf;

// calculate frame number
  int  psnrPOC = active_sps->mb_adaptive_frame_field_flag ? p->poc /(input->poc_scale) : p->poc/(3-input->poc_scale);
//  int  psnrPOC = p->MbaffFrameFlag ? p->poc /(input->poc_scale) : p->poc/(3-input->poc_scale);
  
  // KS: Code below might work better if you have fields and a large (>1) poc offset between them
//  int  poc_diff=max(1,(p->bottom_poc - p->top_poc));
//  int  psnrPOC = active_sps->mb_adaptive_frame_field_flag ? p->poc /(input->poc_scale*poc_diff) : p->poc/((3-input->poc_scale)*poc_diff);

  if (psnrPOC==0 && img->psnr_number)
    img->idr_psnr_number=img->psnr_number + 1;
  img->psnr_number=max(img->psnr_number,img->idr_psnr_number+psnrPOC);

  frame_no = img->idr_psnr_number+psnrPOC;

  // KS: this buffer should actually be allocated only once, but this is still much faster than the previous version
  buf = malloc (p->size_y*p->size_x*symbol_size_in_bytes);

  if (NULL == buf)
  {
    no_mem_exit("find_snr: buf");
  }

  status = lseek (p_ref, framesize_in_bytes*frame_no, SEEK_SET);
  if (status == -1)
    {
      snprintf(errortext, ET_SIZE, "Error in seeking frame number: %d", frame_no);
      fprintf(stderr, "%s", errortext);
      return;
  }

  if(rgb_output)
    lseek (p_ref, framesize_in_bytes/3, SEEK_CUR);  // S.SUN 18072004

  read(p_ref, buf, p->size_y*p->size_x*symbol_size_in_bytes);

  for (j=0; j < p->size_y; j++)
    for (i=0; i < p->size_x; i++)
    {
      imgY_ref[j][i]=0;
      memcpy(&(imgY_ref[j][i]), buf +((j*p->size_x+i)*symbol_size_in_bytes), symbol_size_in_bytes);
    }

  if(rgb_output)
    lseek (p_ref, -framesize_in_bytes*2/3, SEEK_CUR);  // S.SUN 18072004

  for (uv=0; uv < 2; uv++)
  {
    if(rgb_output && uv==1)
      lseek (p_ref, framesize_in_bytes/3, SEEK_CUR);  // S.SUN 18072004

    read(p_ref, buf, p->size_y_cr*p->size_x_cr*symbol_size_in_bytes);

    for (j=0; j < p->size_y_cr ; j++)
      for (i=0; i < p->size_x_cr; i++)
        memcpy(&(imgUV_ref[uv][j][i]), buf +((j*p->size_x_cr+i)*symbol_size_in_bytes), symbol_size_in_bytes);
  }
  
  free (buf);

  img->quad[0]=0;
  diff_y=0;
  for (j=0; j < p->size_y; ++j)
  {
    for (i=0; i < p->size_x; ++i)
    {
      diff_y += img->quad[abs(p->imgY[j][i]-imgY_ref[j][i])];
    }
  }

  // Chroma
  diff_u=0;
  diff_v=0;

  for (j=0; j < p->size_y_cr; ++j)
  {
    for (i=0; i < p->size_x_cr; ++i)
    {
      diff_u += img->quad[abs(imgUV_ref[0][j][i]-p->imgUV[0][j][i])];
      diff_v += img->quad[abs(imgUV_ref[1][j][i]-p->imgUV[1][j][i])];
    }
  }

/*  if (diff_y == 0)
      diff_y = 1;
  if (diff_u == 0)
      diff_u = 1;
  if (diff_v == 0)
      diff_v = 1; */

  // Collecting SNR statistics
  if (diff_y != 0)
    snr->snr_y=(float)(10*log10(max_pix_value_sqd*(double)((double)(img->width)*(img->height) / diff_y)));        // luma snr for current frame
  else
    snr->snr_y=0;
  if (diff_u != 0)
    snr->snr_u=(float)(10*log10(max_pix_value_sqd_uv*(double)((double)(img->width_cr)*(img->height_cr) / (diff_u))));    //  chroma snr for current frame
  else
    snr->snr_u=0;
  if (diff_v != 0)
    snr->snr_v=(float)(10*log10(max_pix_value_sqd_uv*(double)((double)(img->width_cr)*(img->height_cr) / (diff_v))));    //  chroma snr for current frame
  else
    snr->snr_v=0;

  if (img->number == 0) // first
  {
    snr->snr_ya=snr->snr_y1=snr->snr_y;                                                        // keep luma snr for first frame
    snr->snr_ua=snr->snr_u1=snr->snr_u;                                                        // keep chroma snr for first frame
    snr->snr_va=snr->snr_v1=snr->snr_v;                                                        // keep chroma snr for first frame
  
  }
  else
  {
    snr->snr_ya=(float)(snr->snr_ya*(img->number+Bframe_ctr)+snr->snr_y)/(img->number+Bframe_ctr+1); // average snr chroma for all frames
    snr->snr_ua=(float)(snr->snr_ua*(img->number+Bframe_ctr)+snr->snr_u)/(img->number+Bframe_ctr+1); // average snr luma for all frames
    snr->snr_va=(float)(snr->snr_va*(img->number+Bframe_ctr)+snr->snr_v)/(img->number+Bframe_ctr+1); // average snr luma for all frames
  } 
}


/*!
 ************************************************************************
 * \brief
 *    Interpolation of 1/4 subpixel
 ************************************************************************
 */
void get_block(int ref_frame, StorablePicture **list, int x_pos, int y_pos, struct img_par *img, int block[BLOCK_SIZE][BLOCK_SIZE])
{

  int dx, dy;
  int x, y;
  int i, j;
  int maxold_x,maxold_y;
  int result;
  int pres_x;
  int pres_y; 
  int tmp_res[4][9];
  static const int COEF[6] = {    1, -5, 20, 20, -5, 1  };

  dx = x_pos&3;
  dy = y_pos&3;
  x_pos = (x_pos-dx)/4;
  y_pos = (y_pos-dy)/4;

  maxold_x = dec_picture->size_x-1;
  maxold_y = dec_picture->size_y-1;

  if (dec_picture->mb_field[img->current_mb_nr])
    maxold_y = dec_picture->size_y/2 - 1;

  if (dx == 0 && dy == 0) {  /* fullpel position */
    for (j = 0; j < BLOCK_SIZE; j++)
      for (i = 0; i < BLOCK_SIZE; i++)
        block[i][j] = list[ref_frame]->imgY[max(0,min(maxold_y,y_pos+j))][max(0,min(maxold_x,x_pos+i))];
  }
  else { /* other positions */

    if (dy == 0) { /* No vertical interpolation */

      for (j = 0; j < BLOCK_SIZE; j++) {
        for (i = 0; i < BLOCK_SIZE; i++) {
          for (result = 0, x = -2; x < 4; x++)
            result += list[ref_frame]->imgY[max(0,min(maxold_y,y_pos+j))][max(0,min(maxold_x,x_pos+i+x))]*COEF[x+2];
          block[i][j] = max(0, min(img->max_imgpel_value, (result+16)/32));
        }
      }

      if ((dx&1) == 1) {
        for (j = 0; j < BLOCK_SIZE; j++)
          for (i = 0; i < BLOCK_SIZE; i++)
            block[i][j] = (block[i][j] + list[ref_frame]->imgY[max(0,min(maxold_y,y_pos+j))][max(0,min(maxold_x,x_pos+i+dx/2))] +1 )/2;
      }
    }
    else if (dx == 0) {  /* No horizontal interpolation */

      for (j = 0; j < BLOCK_SIZE; j++) {
        for (i = 0; i < BLOCK_SIZE; i++) {
          for (result = 0, y = -2; y < 4; y++)
            result += list[ref_frame]->imgY[max(0,min(maxold_y,y_pos+j+y))][max(0,min(maxold_x,x_pos+i))]*COEF[y+2];
          block[i][j] = max(0, min(img->max_imgpel_value, (result+16)/32));
        }
      }

      if ((dy&1) == 1) {
        for (j = 0; j < BLOCK_SIZE; j++)
          for (i = 0; i < BLOCK_SIZE; i++)
           block[i][j] = (block[i][j] + list[ref_frame]->imgY[max(0,min(maxold_y,y_pos+j+dy/2))][max(0,min(maxold_x,x_pos+i))] +1 )/2;
      }
    }
    else if (dx == 2) {  /* Vertical & horizontal interpolation */

      for (j = -2; j < BLOCK_SIZE+3; j++) {
        for (i = 0; i < BLOCK_SIZE; i++)
          for (tmp_res[i][j+2] = 0, x = -2; x < 4; x++)
            tmp_res[i][j+2] += list[ref_frame]->imgY[max(0,min(maxold_y,y_pos+j))][max(0,min(maxold_x,x_pos+i+x))]*COEF[x+2];
      }

      for (j = 0; j < BLOCK_SIZE; j++) {
        for (i = 0; i < BLOCK_SIZE; i++) {
          for (result = 0, y = -2; y < 4; y++)
            result += tmp_res[i][j+y+2]*COEF[y+2];
          block[i][j] = max(0, min(img->max_imgpel_value, (result+512)/1024));
        } 
      }

      if ((dy&1) == 1) {
        for (j = 0; j < BLOCK_SIZE; j++)
          for (i = 0; i < BLOCK_SIZE; i++)
            block[i][j] = (block[i][j] + max(0, min(img->max_imgpel_value, (tmp_res[i][j+2+dy/2]+16)/32)) +1 )/2;
      }
    }
    else if (dy == 2) {  /* Horizontal & vertical interpolation */

      for (j = 0; j < BLOCK_SIZE; j++) {
        for (i = -2; i < BLOCK_SIZE+3; i++)
          for (tmp_res[j][i+2] = 0, y = -2; y < 4; y++)