/*
	File:		ScaleBlit.cpp

	Contains:	fast scaling blitter

	Written by:	Strout

	Copyright:	Copyright ©1999 the Salk Institute

	Change History (most recent first):

         <4>    06/17/99    JJS     bug fix in yscale calculation
         <3>    06/17/99    JJS     enable non-square scaling
         <2>    06/11/99    JJS     revert to CopyBits when scale factor is odd
         <1>    06/11/99    JJS     initial checkin to StimulusLib2 project
*/


#include "ScaleBlit.h"

#ifndef __ERRORS__
#include <Errors.h>
#endif

OSErr ScaleBlitOffscreen(PixMapPtr srcpm, PixMapPtr destpm, Rect *srcR, Rect *destR)
{
	// note scale ("scale" means X scale; we'll note Y scale separately below)
	long scale = (destR->right-destR->left) / (srcR->right-srcR->left);
	
	// if scaling is odd, then we can't do it -- just use CopyBits
	if (scale & 1) {
		CopyBits( (BitMapPtr)srcpm, (BitMapPtr)destpm, srcR, destR, srcCopy, NULL);
		return noErr;
	}
		
	// get destination address
	short destRowBytes = destpm->rowBytes & 0x3FFF;
	double* destBits = (double*)(destpm->baseAddr);
	if (!destBits) return paramErr;
	long destRowDoubles = destRowBytes / sizeof(double);

	// find how many bytes to iterate over the dest buffer while drawing
	long pixelBytes = (srcpm->pixelSize == 16 ? 2 : 4);
	long drawBytes = (destR->right - destR->left) * pixelBytes;
	long drawDoubles = drawBytes / sizeof(double);

	// get source memory address
	long *srcBits = (long*)srcpm->baseAddr;
	long srcRowLongs = (srcpm->rowBytes & 0x3FFF) / sizeof(long);

	// stuff therein
	double *destRowStart = destBits;
	long *srcRowStart = srcBits;
	long yscale = (destR->bottom-destR->top) / (srcR->bottom-srcR->top);
	long row = yscale;		// remaining dest rows for this source row
	long colDoublesInScale = scale * pixelBytes / sizeof(double);
	long col = colDoublesInScale;
	int zero = 0;
	for (long y=destR->top; y<destR->bottom; y++) {
		// copy source row to dest row
		long *srcp = srcRowStart;
		double *destp = destRowStart;
		union {
			double pixd;
			long pixl[2];
		} pixel;

		// get initial source pixels (2 at a time, or 8 bytes)
		pixel.pixl[0] = pixel.pixl[1] = *srcp;

		for (long x=drawDoubles; x; x--) {
			// OFI: unroll this loop, to avoid so much branching and checking!
			// (But then you must be very careful of boundary conditions.)
			if (not ((long)destp & 31)) __dcbz(destp, zero);	// clear cache line
			double foo = pixel.pixd;
			*destp++ = pixel.pixd;
			if ((col=col-1)==0) {
				// advance to next source pixel
				srcp++;
				pixel.pixl[0] = pixel.pixl[1] = *srcp;
				col = colDoublesInScale;
			}
		}

		// advance to next dest row
		destRowStart += destRowDoubles;
		if (--row == 0) {
			// advance to next source row
			srcRowStart += srcRowLongs;
			row = yscale;
		}
	}
	return noErr;
}

//	Use this function when scaling to the screen:
OSErr ScaleBlitToScreen(PixMapPtr srcpm, PixMapPtr destpm, PixMapPtr tempm, 
						Rect *srcR, Rect *destR)
{
	ScaleBlitOffscreen( srcpm, tempm, srcR, destR );
	CopyBits( (BitMapPtr)tempm, (BitMapPtr)destpm, destR, destR, srcCopy, NULL );
	return noErr;
}