#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <crblib/inc.h>

static void do_tdec_line(int * to,int *from,int len)
{
int x,*ptr,*low,*high,half;

	half = len>>1;

	assert( (len & 1) == 0 );
	assert( half >= 2 );

	low  = to; high = to + half; ptr = from;

	*high = ptr[1] - ((ptr[2] + ptr[0])>>1);
	*low  = ptr[0] + ((high[0] )>>1);
	ptr += 2; high++; low++;

	for(x=(half-2);x--;) {
		*high = ptr[1] - ((ptr[2] + ptr[0])>>1);
		*low =  ptr[0] + ((high[0] + high[-1])>>2);
		ptr += 2; high++; low++;
	}

	*high = ptr[1] - ptr[0];
	*low = *ptr + ((high[0] + high[-1])>>2);

}

static void un_tdec_line(int *to,int *from,int len)
{
int x,*ptr,*low,*high,half;

	half = len>>1;
	low  = from; high = from + half; ptr = to;
	
#if 1
	ptr[0] = low[0] - ((high[0])>>1);
	ptr += 2; high++; low++;
	for(x=(half-1);x--;) {
		ptr[ 0] = low[0]  - ((high[0] + high[-1])>>2);
		ptr[-1] = high[-1] + ((ptr[0] + ptr[-2])>>1);
		ptr += 2; high++; low++;
	}
	ptr[-1] = high[-1] + ptr[-2];
#else
	/*#*/{
	int hi,lh,cur,last;

	hi = *high++;
	ptr[0] = cur = (*low++) - (hi>>1);
	ptr += 2;
	for(x=(half-1);x--;) {
		last = cur;	lh = hi;
		hi = *high++; 
		ptr[ 0] = cur = (*low++)  - ((hi + lh)>>2);
		ptr[-1] = lh + ((cur + last)>>1);
		ptr += 2;
	}
	ptr[-1] = hi + cur;
	/*#*/}
#endif

}

void cdf22_2D(int **rows, int width, int height, int levels,bool inverse)
{
int x, y, w, h, l;
int *buffer,*tempbuf,*temprow;

    if (width%(1 << (levels+1)) || height%(1 << (levels+1)))
		errexit("width and height must be divisible by 2^(levels+1)");
  
    /* Allocate a work array (for transposing columns) */
    
   	if ( (buffer = newarray(int,height+max(width,height)+height)) == NULL )
		errexit("malloc failed");
	temprow = buffer+height;
	tempbuf = buffer+height+height;

	if ( !inverse ) {

		for (l = 0; l < levels; l++) {
			w = width >> l;
			h = height >> l;
      
			/* Rows */
	
			do_tdec_line(temprow,rows[h-1],w);
			for (y = h-2; y >=0; y--) {
				do_tdec_line(rows[y+1],rows[y],w);
			}
	
			/* Columns */
	
			for (x = 0; x < w; x++) {
					for (y = 1; y < h; y++) buffer[y-1] = rows[y][x];
					buffer[h-1] = temprow[x];
				do_tdec_line(tempbuf,buffer,h);
					for (y = 0; y < h; y++) rows[y][x] = tempbuf[y];
			}
		}

    } else {

		for (l = levels-1; l >= 0; l--) { /** backwards in scale **/
			w = width >> l;
			h = height >> l;

			/* Columns */
	
			for (x = 0; x < w; x++) {
					for (y = 0; y < h; y++) buffer[y] = rows[y][x];
				un_tdec_line(tempbuf,buffer,h);
					for (y = 0; y < h-1; y++) rows[y+1][x] = tempbuf[y];
					temprow[x] = tempbuf[h-1];
			}

			/* Rows */
			for (y = 0; y < h-1; y++) {
				un_tdec_line(rows[y],rows[y+1],w);
			} 
			un_tdec_line(rows[h-1],temprow,w);
		}
	}

	free(buffer);
}


void cdfQuad(int *band,int w,int h,int fullw,bool inverse)
{
int x, y;
int *buffer,*tempbuf,*bptr,*temprow;

   	if ( (buffer = newarray(int,h+h+max(w,h))) == NULL ) {
		errputs("malloc failed"); exit(10);
	}
	temprow = buffer+h;
	tempbuf = buffer+h+h;
  
	if ( !inverse ) { /* forward transform. */
	
		bptr = band + (h-1)*fullw;
		do_tdec_line(temprow,bptr,w);
		for (y = (h-1); y--;) {
			bptr -= fullw;
			do_tdec_line(bptr+fullw,bptr,w);
		}
    
		for (x = 0; x < w; x++) {
			bptr = band + x + fullw;
			for (y = 0; y < (h-1); y++) { buffer[y] = *bptr; bptr += fullw; }
			buffer[h-1] = temprow[x];
			do_tdec_line(tempbuf,buffer,h);
			bptr = band + x;
			for (y = 0; y < h; y++) { *bptr = tempbuf[y]; bptr += fullw; }
		}

    } else {

		for (x = 0; x < w; x++) {
			bptr = band + x;
			for (y = 0; y < h; y++) { buffer[y] = *bptr; bptr += fullw; }
			un_tdec_line(tempbuf,buffer,h);
			bptr = band + x + fullw;
			for (y = 0; y < h-1; y++) { *bptr = tempbuf[y]; bptr += fullw; }
			temprow[x] = tempbuf[h-1];
		}

		bptr = band;
		for (y = (h-1); y--; ){
			un_tdec_line(bptr,bptr+fullw,w);
			bptr += fullw;
		}
		un_tdec_line(bptr,temprow,w);

	}

	free(buffer);
}