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

#define SHIFT(x)	((x)>>(12))

#define ROOT3	7094
#define ROOT2O2	2896
#define ROOT3O4	1774
#define ROOT3M2O4	(-275)

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

	if ( len <= 2 ) errexit("special-cased to len > 2 , scumbag");
	if ( len&1 ) errexit("len shouldn't be odd");

	half = len>>1;

	ptr = from;	high = to + half;
	for(x=0;x<half;x++) {
		*high++ = ptr[1] - SHIFT( ROOT3 * ptr[0] );
		ptr += 2;
	}

	ptr = from; low = to; high = to + half;
	for(x=0;x<half;x++) {
		if ( x == half-1 ) {
			*low = ptr[0] + SHIFT( ROOT3O4 * high[0] + ROOT3M2O4 * high[0] );
		} else {
			*low = ptr[0] + SHIFT( ROOT3O4 * high[0] + ROOT3M2O4 * high[1] );
		}
		if ( x == 0 )	*high += low[0] ;
		else			*high += low[-1] ;
		ptr += 2; high++; low++;
	}
}

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

	half = len>>1;

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

	low  = from;	high = from + half;
	for(x=0;x<half;x++) {
		if ( x == 0 )	*high -= low[0] ;
		else			*high -= low[-1] ;
		high++; low++;
	}

	ptr = to;	low  = from;	high = from + half;
	for(x=0;x<half;x++) {
		if ( x == half-1 ) {
			ptr[0] = *low - SHIFT( ROOT3O4 * high[0] + ROOT3M2O4 * high[0] );
		} else {
			ptr[0] = *low - SHIFT( ROOT3O4 * high[0] + ROOT3M2O4 * high[1] );
		}
		ptr += 2; high++; low++;
	}

	ptr = to;	high = from + half;
	for(x=0;x<half;x++) {
		ptr[1] = (*high++) + SHIFT( ROOT3 * ptr[0] );
		ptr += 2;
	}
}

void d4_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 d4Quad(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);
}