2xsai.cpp

著名SFC模拟器Snes9x的源代码。

/*
 * Snes9x - Portable Super Nintendo Entertainment System (TM) emulator.
 *
 * (c) Copyright 1996 - 2001 Gary Henderson ([e-mail]) and
 *                           Jerremy Koot ([e-mail])
 *
 * Super FX C emulator code 
 * (c) Copyright 1997 - 1999 Ivar ([e-mail]) and
 *                           Gary Henderson.
 * Super FX assembler emulator code (c) Copyright 1998 zsKnight and _Demo_.
 *
 * DSP1 emulator code (c) Copyright 1998 Ivar, _Demo_ and Gary Henderson.
 * C4 asm and some C emulation code (c) Copyright 2000 zsKnight and _Demo_.
 * C4 C code (c) Copyright 2001 Gary Henderson ([e-mail]).
 *
 * DOS port code contains the works of other authors. See headers in
 * individual files.
 *
 * Snes9x homepage: www.snes9x.com
 *
 * Permission to use, copy, modify and distribute Snes9x in both binary and
 * source form, for non-commercial purposes, is hereby granted without fee,
 * providing that this license information and copyright notice appear with
 * all copies and any derived work.
 *
 * This software is provided 'as-is', without any express or implied
 * warranty. In no event shall the authors be held liable for any damages
 * arising from the use of this software.
 *
 * Snes9x is freeware for PERSONAL USE only. Commercial users should
 * seek permission of the copyright holders first. Commercial use includes
 * charging money for Snes9x or software derived from Snes9x.
 *
 * The copyright holders request that bug fixes and improvements to the code
 * should be forwarded to them so everyone can benefit from the modifications
 * in future versions.
 *
 * Super NES and Super Nintendo Entertainment System are trademarks of
 * Nintendo Co., Limited and its subsidiary companies.
 */
#ifdef __DJGPP
#include <allegro.h>
#endif

#include "snes9x.h"
#include "port.h"
#include "gfx.h"

extern "C"
{

#ifdef MMX
    void _2xSaILine (uint8 *srcPtr, uint8 *deltaPtr, uint32 srcPitch,
		     uint32 width, uint8 *dstPtr, uint32 dstPitch);
    void _2xSaISuperEagleLine (uint8 *srcPtr, uint8 *deltaPtr,
			       uint32 srcPitch, uint32 width,
			       uint8 *dstPtr, uint32 dstPitch);
    void _2xSaISuper2xSaILine (uint8 *srcPtr, uint8 *deltaPtr,
			       uint32 srcPitch, uint32 width,
			       uint8 *dstPtr, uint32 dstPitch);
    void Init_2xSaIMMX (uint32 BitFormat);
    void BilinearMMX (uint16 * A, uint16 * B, uint16 * C, uint16 * D,
		      uint16 * dx, uint16 * dy, uint8 *dP);
    void BilinearMMXGrid0 (uint16 * A, uint16 * B, uint16 * C, uint16 * D,
			   uint16 * dx, uint16 * dy, uint8 *dP);
    void BilinearMMXGrid1 (uint16 * A, uint16 * B, uint16 * C, uint16 * D,
			   uint16 * dx, uint16 * dy, uint8 *dP);
    void EndMMX ();

static bool8 cpu_mmx = 1;
#endif
} 

static uint32 colorMask = 0xF7DEF7DE;
static uint32 lowPixelMask = 0x08210821;
static uint32 qcolorMask = 0xE79CE79C;
static uint32 qlowpixelMask = 0x18631863;
static uint32 redblueMask = 0xF81F;
static uint32 greenMask = 0x7E0;

int Init_2xSaI (uint32 BitFormat)
{
    if (BitFormat == 565)
    {
	colorMask = 0xF7DEF7DE;
	lowPixelMask = 0x08210821;
	qcolorMask = 0xE79CE79C;
	qlowpixelMask = 0x18631863;
	redblueMask = 0xF81F;
	greenMask = 0x7E0;
    }
    else if (BitFormat == 555)
    {
	colorMask = 0x7BDE7BDE;
	lowPixelMask = 0x04210421;
	qcolorMask = 0x739C739C;
	qlowpixelMask = 0x0C630C63;
	redblueMask = 0x7C1F;
	greenMask = 0x3E0;
    }
    else
    {
	return 0;
    }

#ifdef MMX
    Init_2xSaIMMX (BitFormat);
#endif

    return 1;
}

static inline int GetResult1 (uint32 A, uint32 B, uint32 C, uint32 D,
			      uint32 /* E */)
{
    int x = 0;
    int y = 0;
    int r = 0;

    if (A == C)
	x += 1;
    else if (B == C)
	y += 1;
    if (A == D)
	x += 1;
    else if (B == D)
	y += 1;
    if (x <= 1)
	r += 1;
    if (y <= 1)
	r -= 1;
    return r;
}

static inline int GetResult2 (uint32 A, uint32 B, uint32 C, uint32 D,
			      uint32 /* E */)
{
    int x = 0;
    int y = 0;
    int r = 0;

    if (A == C)
	x += 1;
    else if (B == C)
	y += 1;
    if (A == D)
	x += 1;
    else if (B == D)
	y += 1;
    if (x <= 1)
	r -= 1;
    if (y <= 1)
	r += 1;
    return r;
}

static inline int GetResult (uint32 A, uint32 B, uint32 C, uint32 D)
{
    int x = 0;
    int y = 0;
    int r = 0;

    if (A == C)
	x += 1;
    else if (B == C)
	y += 1;
    if (A == D)
	x += 1;
    else if (B == D)
	y += 1;
    if (x <= 1)
	r += 1;
    if (y <= 1)
	r -= 1;
    return r;
}

static inline uint32 INTERPOLATE (uint32 A, uint32 B)
{
    if (A != B)
    {
	return (((A & colorMask) >> 1) + ((B & colorMask) >> 1) +
		(A & B & lowPixelMask));
    }
    else
	return A;
}

static inline uint32 Q_INTERPOLATE (uint32 A, uint32 B, uint32 C, uint32 D)
{
    register uint32 x = ((A & qcolorMask) >> 2) +
	((B & qcolorMask) >> 2) +
	((C & qcolorMask) >> 2) + ((D & qcolorMask) >> 2);
    register uint32 y = (A & qlowpixelMask) +
	(B & qlowpixelMask) + (C & qlowpixelMask) + (D & qlowpixelMask);

    y = (y >> 2) & qlowpixelMask;
    return x + y;
}

#define BLUE_MASK565 0x001F001F
#define RED_MASK565 0xF800F800
#define GREEN_MASK565 0x07E007E0

#define BLUE_MASK555 0x001F001F
#define RED_MASK555 0x7C007C00
#define GREEN_MASK555 0x03E003E0

void Super2xSaI (uint8 *srcPtr, uint32 srcPitch,
		 uint8 *deltaPtr, uint8 *dstPtr, uint32 dstPitch,
		 int width, int height)
{
    uint16 *bP;
    uint8  *dP;
    uint32 inc_bP;
    uint32 Nextline = srcPitch >> 1;

#ifdef MMX
    if (cpu_mmx)
    {
	for (; height; height--)
	{
	    _2xSaISuper2xSaILine (srcPtr, deltaPtr, srcPitch, width,
				  dstPtr, dstPitch);
	    srcPtr += srcPitch;
	    dstPtr += dstPitch * 2;
	    deltaPtr += srcPitch;
	}
    }
    else
#endif

    {
	inc_bP = 1;

	for (; height; height--)
	{
	    bP = (uint16 *) srcPtr;
	    dP = (uint8 *) dstPtr;

	    for (uint32 finish = width; finish; finish -= inc_bP)
	    {
		uint32 color4, color5, color6;
		uint32 color1, color2, color3;
		uint32 colorA0, colorA1, colorA2, colorA3,
		    colorB0, colorB1, colorB2, colorB3, colorS1, colorS2;
		uint32 product1a, product1b, product2a, product2b;

//---------------------------------------    B1 B2
//                                         4  5  6 S2
//                                         1  2  3 S1
//                                           A1 A2

		colorB0 = *(bP - Nextline - 1);
		colorB1 = *(bP - Nextline);
		colorB2 = *(bP - Nextline + 1);
		colorB3 = *(bP - Nextline + 2);

		color4 = *(bP - 1);
		color5 = *(bP);
		color6 = *(bP + 1);
		colorS2 = *(bP + 2);

		color1 = *(bP + Nextline - 1);
		color2 = *(bP + Nextline);
		color3 = *(bP + Nextline + 1);
		colorS1 = *(bP + Nextline + 2);

		colorA0 = *(bP + Nextline + Nextline - 1);
		colorA1 = *(bP + Nextline + Nextline);
		colorA2 = *(bP + Nextline + Nextline + 1);
		colorA3 = *(bP + Nextline + Nextline + 2);

//--------------------------------------
		if (color2 == color6 && color5 != color3)
		{
		    product2b = product1b = color2;
		}
		else if (color5 == color3 && color2 != color6)
		{
		    product2b = product1b = color5;
		}
		else if (color5 == color3 && color2 == color6)
		{
		    register int r = 0;

		    r += GetResult (color6, color5, color1, colorA1);
		    r += GetResult (color6, color5, color4, colorB1);
		    r += GetResult (color6, color5, colorA2, colorS1);
		    r += GetResult (color6, color5, colorB2, colorS2);

		    if (r > 0)
			product2b = product1b = color6;
		    else if (r < 0)
			product2b = product1b = color5;
		    else
		    {
			product2b = product1b = INTERPOLATE (color5, color6);
		    }
		}
		else
		{
		    if (color6 == color3 && color3 == colorA1
			    && color2 != colorA2 && color3 != colorA0)
			product2b =
			    Q_INTERPOLATE (color3, color3, color3, color2);
		    else if (color5 == color2 && color2 == colorA2
			     && colorA1 != color3 && color2 != colorA3)
			product2b =
			    Q_INTERPOLATE (color2, color2, color2, color3);
		    else
			product2b = INTERPOLATE (color2, color3);

		    if (color6 == color3 && color6 == colorB1
			    && color5 != colorB2 && color6 != colorB0)
			product1b =
			    Q_INTERPOLATE (color6, color6, color6, color5);
		    else if (color5 == color2 && color5 == colorB2
			     && colorB1 != color6 && color5 != colorB3)
			product1b =
			    Q_INTERPOLATE (color6, color5, color5, color5);
		    else
			product1b = INTERPOLATE (color5, color6);
		}

		if (color5 == color3 && color2 != color6 && color4 == color5
			&& color5 != colorA2)
		    product2a = INTERPOLATE (color2, color5);
		else
		    if (color5 == color1 && color6 == color5
			&& color4 != color2 && color5 != colorA0)
		    product2a = INTERPOLATE (color2, color5);
		else
		    product2a = color2;

		if (color2 == color6 && color5 != color3 && color1 == color2
			&& color2 != colorB2)
		    product1a = INTERPOLATE (color2, color5);
		else
		    if (color4 == color2 && color3 == color2
			&& color1 != color5 && color2 != colorB0)
		    product1a = INTERPOLATE (color2, color5);
		else
		    product1a = color5;

		product1a = product1a | (product1b << 16);
		product2a = product2a | (product2b << 16);

		*((uint32 *) dP) = product1a;
		*((uint32 *) (dP + dstPitch)) = product2a;

		bP += inc_bP;
		dP += sizeof (uint32);
	    }			// end of for ( finish= width etc..)

	    srcPtr   += srcPitch;
	    dstPtr   += dstPitch * 2;
	    deltaPtr += srcPitch;
	}			// endof: for (; height; height--)
    }
}

void SuperEagle (uint8 *srcPtr, uint32 srcPitch, uint8 *deltaPtr, 
		 uint8 *dstPtr, uint32 dstPitch, int width, int height)
{
    uint8  *dP;
    uint16 *bP;
    uint16 *xP;
    uint32 inc_bP;

#ifdef MMX
    if (cpu_mmx)
    {
	for (; height; height--)
	{
	    _2xSaISuperEagleLine (srcPtr, deltaPtr, srcPitch, width,
				  dstPtr, dstPitch);
	    srcPtr += srcPitch;
	    dstPtr += dstPitch * 2;
	    deltaPtr += srcPitch;
	}
    }
    else
#endif

    {
	inc_bP = 1;

	uint32 Nextline = srcPitch >> 1;

	for (; height; height--)
	{
	    bP = (uint16 *) srcPtr;
	    xP = (uint16 *) deltaPtr;
	    dP = dstPtr;
	    for (uint32 finish = width; finish; finish -= inc_bP)
	    {
		uint32 color4, color5, color6;
		uint32 color1, color2, color3;
		uint32 colorA1, colorA2, colorB1, colorB2, colorS1, colorS2;
		uint32 product1a, product1b, product2a, product2b;

		colorB1 = *(bP - Nextline);
		colorB2 = *(bP - Nextline + 1);

		color4 = *(bP - 1);
		color5 = *(bP);
		color6 = *(bP + 1);
		colorS2 = *(bP + 2);

		color1 = *(bP + Nextline - 1);
		color2 = *(bP + Nextline);
		color3 = *(bP + Nextline + 1);
		colorS1 = *(bP + Nextline + 2);

		colorA1 = *(bP + Nextline + Nextline);
		colorA2 = *(bP + Nextline + Nextline + 1);

		// --------------------------------------
		if (color2 == color6 && color5 != color3)
		{
		    product1b = product2a = color2;
		    if ((color1 == color2) || (color6 == colorB2))
		    {
			product1a = INTERPOLATE (color2, color5);
			product1a = INTERPOLATE (color2, product1a);
//                       product1a = color2;
		    }
		    else
		    {
			product1a = INTERPOLATE (color5, color6);
		    }

		    if ((color6 == colorS2) || (color2 == colorA1))
		    {
			product2b = INTERPOLATE (color2, color3);
			product2b = INTERPOLATE (color2, product2b);
//                       product2b = color2;
		    }
		    else
		    {
			product2b = INTERPOLATE (color2, color3);
		    }
		}
		else if (color5 == color3 && color2 != color6)
		{
		    product2b = product1a = color5;

		    if ((colorB1 == color5) || (color3 == colorS1))
		    {
			product1b = INTERPOLATE (color5, color6);
			product1b = INTERPOLATE (color5, product1b);
//                       product1b = color5;
		    }
		    else
		    {
			product1b = INTERPOLATE (color5, color6);
		    }

		    if ((color3 == colorA2) || (color4 == color5))
		    {
			product2a = INTERPOLATE (color5, color2);
			product2a = INTERPOLATE (color5, product2a);
//                       product2a = color5;
		    }
		    else
		    {
			product2a = INTERPOLATE (color2, color3);
		    }

		}
		else if (color5 == color3 && color2 == color6)
		{
		    register int r = 0;

		    r += GetResult (color6, color5, color1, colorA1);
		    r += GetResult (color6, color5, color4, colorB1);
		    r += GetResult (color6, color5, colorA2, colorS1);
		    r += GetResult (color6, color5, colorB2, colorS2);

		    if (r > 0)
		    {
			product1b = product2a = color2;
			product1a = product2b = INTERPOLATE (color5, color6);
		    }
		    else if (r < 0)
		    {
			product2b = product1a = color5;
			product1b = product2a = INTERPOLATE (color5, color6);
		    }
		    else
		    {
			product2b = product1a = color5;
			product1b = product2a = color2;
		    }
		}
		else
		{
		    product2b = product1a = INTERPOLATE (color2, color6);
		    product2b =
			Q_INTERPOLATE (color3, color3, color3, product2b);
		    product1a =
			Q_INTERPOLATE (color5, color5, color5, product1a);

		    product2a = product1b = INTERPOLATE (color5, color3);
		    product2a =
			Q_INTERPOLATE (color2, color2, color2, product2a);
		    product1b =
			Q_INTERPOLATE (color6, color6, color6, product1b);

//                    product1a = color5;
//                    product1b = color6;
//                    product2a = color2;
//                    product2b = color3;
		}
		product1a = product1a | (product1b << 16);
		product2a = product2a | (product2b << 16);

		*((uint32 *) dP) = product1a;
		*((uint32 *) (dP + dstPitch)) = product2a;
		*xP = color5;

		bP += inc_bP;
		xP += inc_bP;
		dP += sizeof (uint32);
	    }			// end of for ( finish= width etc..)

	    srcPtr += srcPitch;
	    dstPtr += dstPitch * 2;
	    deltaPtr += srcPitch;
	}			// endof: for (height; height; height--)
    }
}

void _2xSaI (uint8 *srcPtr, uint32 srcPitch, uint8 *deltaPtr,
	     uint8 *dstPtr, uint32 dstPitch, int width, int height)
{
    uint8  *dP;
    uint16 *bP;
    uint32 inc_bP;

#ifdef MMX
    if (cpu_mmx)
    {
	for (; height; height -= 1)
	{
	    _2xSaILine (srcPtr, deltaPtr, srcPitch, width, dstPtr, dstPitch);
	    srcPtr += srcPitch;
	    dstPtr += dstPitch * 2;
	    deltaPtr += srcPitch;
	}
    }
    else
#endif

    {
	inc_bP = 1;

	uint32 Nextline = srcPitch >> 1;

	for (; height; height--)
	{
	    bP = (uint16 *) srcPtr;
	    dP = dstPtr;

	    for (uint32 finish = width; finish; finish -= inc_bP)
	    {

		register uint32 colorA, colorB;
		uint32 colorC, colorD,
		    colorE, colorF, colorG, colorH,
		    colorI, colorJ, colorK, colorL,

		    colorM, colorN, colorO, colorP;
		uint32 product, product1, product2;

//---------------------------------------
// Map of the pixels:                    I|E F|J
//                                       G|A B|K
//                                       H|C D|L