src/video/SDL_blit.h
author Sam Lantinga
Sun, 06 Mar 2011 21:17:09 -0800
changeset 5426 7a3982c50af8
parent 5423 b69fa50e80d7
child 5439 3a778c6c0269
permissions -rw-r--r--
Fixed compiler warnings
slouken@0
     1
/*
slouken@0
     2
    SDL - Simple DirectMedia Layer
slouken@5262
     3
    Copyright (C) 1997-2011 Sam Lantinga
slouken@0
     4
slouken@0
     5
    This library is free software; you can redistribute it and/or
slouken@1312
     6
    modify it under the terms of the GNU Lesser General Public
slouken@0
     7
    License as published by the Free Software Foundation; either
slouken@1312
     8
    version 2.1 of the License, or (at your option) any later version.
slouken@0
     9
slouken@0
    10
    This library is distributed in the hope that it will be useful,
slouken@0
    11
    but WITHOUT ANY WARRANTY; without even the implied warranty of
slouken@0
    12
    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
slouken@1312
    13
    Lesser General Public License for more details.
slouken@0
    14
slouken@1312
    15
    You should have received a copy of the GNU Lesser General Public
slouken@1312
    16
    License along with this library; if not, write to the Free Software
slouken@1312
    17
    Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
slouken@0
    18
slouken@0
    19
    Sam Lantinga
slouken@252
    20
    slouken@libsdl.org
slouken@0
    21
*/
slouken@1402
    22
#include "SDL_config.h"
slouken@0
    23
slouken@0
    24
#ifndef _SDL_blit_h
slouken@0
    25
#define _SDL_blit_h
slouken@0
    26
slouken@2250
    27
#include "SDL_cpuinfo.h"
slouken@0
    28
#include "SDL_endian.h"
slouken@5163
    29
#include "SDL_surface.h"
slouken@0
    30
slouken@2262
    31
/* SDL blit copy flags */
slouken@2266
    32
#define SDL_COPY_MODULATE_COLOR     0x00000001
slouken@2266
    33
#define SDL_COPY_MODULATE_ALPHA     0x00000002
slouken@5140
    34
#define SDL_COPY_BLEND              0x00000010
slouken@5140
    35
#define SDL_COPY_ADD                0x00000020
slouken@5184
    36
#define SDL_COPY_MOD                0x00000040
slouken@2266
    37
#define SDL_COPY_COLORKEY           0x00000100
slouken@2266
    38
#define SDL_COPY_NEAREST            0x00000200
slouken@2266
    39
#define SDL_COPY_RLE_DESIRED        0x00001000
slouken@2266
    40
#define SDL_COPY_RLE_COLORKEY       0x00002000
slouken@2266
    41
#define SDL_COPY_RLE_ALPHAKEY       0x00004000
slouken@2853
    42
#define SDL_COPY_RLE_MASK           (SDL_COPY_RLE_DESIRED|SDL_COPY_RLE_COLORKEY|SDL_COPY_RLE_ALPHAKEY)
slouken@2262
    43
slouken@2262
    44
/* SDL blit CPU flags */
slouken@2266
    45
#define SDL_CPU_ANY                 0x00000000
slouken@2266
    46
#define SDL_CPU_MMX                 0x00000001
slouken@5389
    47
#define SDL_CPU_3DNOW               0x00000002
slouken@2266
    48
#define SDL_CPU_SSE                 0x00000004
slouken@2266
    49
#define SDL_CPU_SSE2                0x00000008
slouken@5389
    50
#define SDL_CPU_ALTIVEC_PREFETCH    0x00000010
slouken@5389
    51
#define SDL_CPU_ALTIVEC_NOPREFETCH  0x00000020
slouken@2262
    52
slouken@2267
    53
typedef struct
slouken@2267
    54
{
slouken@2262
    55
    Uint8 *src;
slouken@2262
    56
    int src_w, src_h;
slouken@2262
    57
    int src_pitch;
slouken@2267
    58
    int src_skip;
slouken@2262
    59
    Uint8 *dst;
slouken@2262
    60
    int dst_w, dst_h;
slouken@2262
    61
    int dst_pitch;
slouken@2267
    62
    int dst_skip;
slouken@2262
    63
    SDL_PixelFormat *src_fmt;
slouken@2262
    64
    SDL_PixelFormat *dst_fmt;
slouken@1895
    65
    Uint8 *table;
slouken@2262
    66
    int flags;
slouken@2262
    67
    Uint32 colorkey;
slouken@2262
    68
    Uint8 r, g, b, a;
slouken@0
    69
} SDL_BlitInfo;
slouken@0
    70
slouken@2267
    71
typedef void (SDLCALL * SDL_BlitFunc) (SDL_BlitInfo * info);
slouken@2262
    72
slouken@2267
    73
typedef struct
slouken@2267
    74
{
slouken@2262
    75
    Uint32 src_format;
slouken@2262
    76
    Uint32 dst_format;
slouken@2262
    77
    int flags;
slouken@2262
    78
    int cpu;
slouken@2262
    79
    SDL_BlitFunc func;
slouken@2262
    80
} SDL_BlitFuncEntry;
slouken@0
    81
slouken@0
    82
/* Blit mapping definition */
slouken@1895
    83
typedef struct SDL_BlitMap
slouken@1895
    84
{
slouken@1895
    85
    SDL_Surface *dst;
slouken@1895
    86
    int identity;
slouken@2257
    87
    SDL_blit blit;
slouken@2257
    88
    void *data;
slouken@2262
    89
    SDL_BlitInfo info;
slouken@0
    90
slouken@1895
    91
    /* the version count matches the destination; mismatch indicates
slouken@1895
    92
       an invalid mapping */
slouken@5288
    93
    Uint32 palette_version;
slouken@0
    94
} SDL_BlitMap;
slouken@0
    95
slouken@0
    96
/* Functions found in SDL_blit.c */
slouken@1895
    97
extern int SDL_CalculateBlit(SDL_Surface * surface);
slouken@0
    98
slouken@2267
    99
/* Functions found in SDL_blit_*.c */
slouken@2267
   100
extern SDL_BlitFunc SDL_CalculateBlit0(SDL_Surface * surface);
slouken@2267
   101
extern SDL_BlitFunc SDL_CalculateBlit1(SDL_Surface * surface);
slouken@2267
   102
extern SDL_BlitFunc SDL_CalculateBlitN(SDL_Surface * surface);
slouken@2267
   103
extern SDL_BlitFunc SDL_CalculateBlitA(SDL_Surface * surface);
slouken@2263
   104
slouken@0
   105
/*
slouken@0
   106
 * Useful macros for blitting routines
slouken@0
   107
 */
slouken@0
   108
slouken@2249
   109
#if defined(__GNUC__)
slouken@2249
   110
#define DECLARE_ALIGNED(t,v,a)  t __attribute__((aligned(a))) v
slouken@2249
   111
#elif defined(_MSC_VER)
slouken@2251
   112
#define DECLARE_ALIGNED(t,v,a)  __declspec(align(a)) t v
slouken@2249
   113
#else
slouken@2249
   114
#define DECLARE_ALIGNED(t,v,a)  t v
slouken@2249
   115
#endif
slouken@2249
   116
slouken@5423
   117
/* The Nintendo surfaces are special. Bit 15 is the transparency
slouken@5423
   118
 * bit. It must be set for the pixel to be displayed. By setting that
slouken@5423
   119
 * value to 0 for other platforms, their compiler should optimize it
slouken@5423
   120
 * out. */
slouken@5423
   121
#ifdef __NDS__
slouken@5423
   122
#define NDS_BIT15 0x8000
slouken@5423
   123
#else
slouken@5423
   124
#define NDS_BIT15 0
slouken@5423
   125
#endif
slouken@5423
   126
slouken@0
   127
/* Load pixel of the specified format from a buffer and get its R-G-B values */
slouken@0
   128
/* FIXME: rescale values to 0..255 here? */
icculus@1162
   129
#define RGB_FROM_PIXEL(Pixel, fmt, r, g, b)				\
slouken@0
   130
{									\
icculus@1162
   131
	r = (((Pixel&fmt->Rmask)>>fmt->Rshift)<<fmt->Rloss); 		\
icculus@1162
   132
	g = (((Pixel&fmt->Gmask)>>fmt->Gshift)<<fmt->Gloss); 		\
icculus@1162
   133
	b = (((Pixel&fmt->Bmask)>>fmt->Bshift)<<fmt->Bloss); 		\
slouken@0
   134
}
icculus@1162
   135
#define RGB_FROM_RGB565(Pixel, r, g, b)					\
slouken@0
   136
{									\
icculus@1162
   137
	r = (((Pixel&0xF800)>>11)<<3);		 			\
icculus@1162
   138
	g = (((Pixel&0x07E0)>>5)<<2); 					\
icculus@1162
   139
	b = ((Pixel&0x001F)<<3); 					\
slouken@0
   140
}
icculus@1162
   141
#define RGB_FROM_RGB555(Pixel, r, g, b)					\
slouken@0
   142
{									\
icculus@1162
   143
	r = (((Pixel&0x7C00)>>10)<<3);		 			\
icculus@1162
   144
	g = (((Pixel&0x03E0)>>5)<<3); 					\
icculus@1162
   145
	b = ((Pixel&0x001F)<<3); 					\
slouken@0
   146
}
icculus@1162
   147
#define RGB_FROM_RGB888(Pixel, r, g, b)					\
slouken@0
   148
{									\
icculus@1162
   149
	r = ((Pixel&0xFF0000)>>16);		 			\
icculus@1162
   150
	g = ((Pixel&0xFF00)>>8);		 			\
icculus@1162
   151
	b = (Pixel&0xFF);			 			\
slouken@0
   152
}
icculus@1162
   153
#define RETRIEVE_RGB_PIXEL(buf, bpp, Pixel)				   \
slouken@0
   154
do {									   \
slouken@0
   155
	switch (bpp) {							   \
slouken@0
   156
		case 2:							   \
icculus@1162
   157
			Pixel = *((Uint16 *)(buf));			   \
slouken@0
   158
		break;							   \
slouken@0
   159
									   \
slouken@0
   160
		case 3: {						   \
slouken@0
   161
		        Uint8 *B = (Uint8 *)(buf);			   \
slouken@2824
   162
			if (SDL_BYTEORDER == SDL_LIL_ENDIAN) {		   \
icculus@1162
   163
			        Pixel = B[0] + (B[1] << 8) + (B[2] << 16); \
slouken@0
   164
			} else {					   \
icculus@1162
   165
			        Pixel = (B[0] << 16) + (B[1] << 8) + B[2]; \
slouken@0
   166
			}						   \
slouken@0
   167
		}							   \
slouken@0
   168
		break;							   \
slouken@0
   169
									   \
slouken@0
   170
		case 4:							   \
icculus@1162
   171
			Pixel = *((Uint32 *)(buf));			   \
slouken@0
   172
		break;							   \
slouken@0
   173
									   \
slouken@0
   174
		default:						   \
slouken@5426
   175
		        Pixel = 0; /* stop gcc complaints */		   \
slouken@0
   176
		break;							   \
slouken@0
   177
	}								   \
slouken@2824
   178
} while (0)
slouken@0
   179
icculus@1162
   180
#define DISEMBLE_RGB(buf, bpp, fmt, Pixel, r, g, b)			   \
slouken@0
   181
do {									   \
slouken@0
   182
	switch (bpp) {							   \
slouken@0
   183
		case 2:							   \
icculus@1162
   184
			Pixel = *((Uint16 *)(buf));			   \
slouken@2824
   185
			RGB_FROM_PIXEL(Pixel, fmt, r, g, b);		   \
slouken@0
   186
		break;							   \
slouken@0
   187
									   \
slouken@2824
   188
		case 3:	{						   \
slouken@5426
   189
            Pixel = 0;                  \
slouken@2824
   190
                        if (SDL_BYTEORDER == SDL_LIL_ENDIAN) {		   \
slouken@2824
   191
			        r = *((buf)+fmt->Rshift/8);		   \
slouken@2824
   192
				g = *((buf)+fmt->Gshift/8);		   \
slouken@2824
   193
				b = *((buf)+fmt->Bshift/8);		   \
slouken@0
   194
			} else {					   \
slouken@2824
   195
			        r = *((buf)+2-fmt->Rshift/8);		   \
slouken@2824
   196
				g = *((buf)+2-fmt->Gshift/8);		   \
slouken@2824
   197
				b = *((buf)+2-fmt->Bshift/8);		   \
slouken@0
   198
			}						   \
slouken@0
   199
		}							   \
slouken@0
   200
		break;							   \
slouken@0
   201
									   \
slouken@0
   202
		case 4:							   \
icculus@1162
   203
			Pixel = *((Uint32 *)(buf));			   \
slouken@2824
   204
			RGB_FROM_PIXEL(Pixel, fmt, r, g, b);		   \
slouken@0
   205
		break;							   \
slouken@0
   206
									   \
slouken@2824
   207
		default:						   \
slouken@5426
   208
                /* stop gcc complaints */		   \
slouken@5426
   209
		        Pixel = 0;              \
slouken@5426
   210
                r = g = b = 0;          \
slouken@0
   211
		break;							   \
slouken@0
   212
	}								   \
slouken@2824
   213
} while (0)
slouken@0
   214
slouken@0
   215
/* Assemble R-G-B values into a specified pixel format and store them */
icculus@1162
   216
#define PIXEL_FROM_RGB(Pixel, fmt, r, g, b)				\
slouken@0
   217
{									\
icculus@1162
   218
	Pixel = ((r>>fmt->Rloss)<<fmt->Rshift)|				\
slouken@0
   219
		((g>>fmt->Gloss)<<fmt->Gshift)|				\
slouken@0
   220
		((b>>fmt->Bloss)<<fmt->Bshift);				\
slouken@0
   221
}
icculus@1162
   222
#define RGB565_FROM_RGB(Pixel, r, g, b)					\
slouken@0
   223
{									\
icculus@1162
   224
	Pixel = ((r>>3)<<11)|((g>>2)<<5)|(b>>3);			\
slouken@0
   225
}
icculus@1162
   226
#define RGB555_FROM_RGB(Pixel, r, g, b)					\
slouken@0
   227
{									\
icculus@1162
   228
	Pixel = ((r>>3)<<10)|((g>>3)<<5)|(b>>3);			\
slouken@0
   229
}
icculus@1162
   230
#define RGB888_FROM_RGB(Pixel, r, g, b)					\
slouken@0
   231
{									\
icculus@1162
   232
	Pixel = (r<<16)|(g<<8)|b;					\
slouken@0
   233
}
slouken@2899
   234
#define ARGB8888_FROM_RGBA(Pixel, r, g, b, a)				\
slouken@2899
   235
{									\
slouken@2899
   236
	Pixel = (a<<24)|(r<<16)|(g<<8)|b;				\
slouken@2899
   237
}
slouken@3054
   238
#define RGBA8888_FROM_RGBA(Pixel, r, g, b, a)				\
slouken@3054
   239
{									\
slouken@3054
   240
	Pixel = (r<<24)|(g<<16)|(b<<8)|a;				\
slouken@3054
   241
}
slouken@3054
   242
#define ABGR8888_FROM_RGBA(Pixel, r, g, b, a)				\
slouken@3054
   243
{									\
slouken@3054
   244
	Pixel = (a<<24)|(b<<16)|(g<<8)|r;				\
slouken@3054
   245
}
slouken@3054
   246
#define BGRA8888_FROM_RGBA(Pixel, r, g, b, a)				\
slouken@3054
   247
{									\
slouken@3054
   248
	Pixel = (b<<24)|(g<<16)|(r<<8)|a;				\
slouken@3054
   249
}
slouken@0
   250
#define ASSEMBLE_RGB(buf, bpp, fmt, r, g, b) 				\
slouken@0
   251
{									\
slouken@0
   252
	switch (bpp) {							\
slouken@0
   253
		case 2: {						\
icculus@1162
   254
			Uint16 Pixel;					\
slouken@0
   255
									\
icculus@1162
   256
			PIXEL_FROM_RGB(Pixel, fmt, r, g, b);		\
slouken@5423
   257
			*((Uint16 *)(buf)) = Pixel | NDS_BIT15;		\
slouken@0
   258
		}							\
slouken@0
   259
		break;							\
slouken@0
   260
									\
slouken@0
   261
		case 3: {						\
slouken@2824
   262
                        if (SDL_BYTEORDER == SDL_LIL_ENDIAN) {		\
slouken@0
   263
			        *((buf)+fmt->Rshift/8) = r;		\
slouken@0
   264
				*((buf)+fmt->Gshift/8) = g;		\
slouken@0
   265
				*((buf)+fmt->Bshift/8) = b;		\
slouken@0
   266
			} else {					\
slouken@0
   267
			        *((buf)+2-fmt->Rshift/8) = r;		\
slouken@0
   268
				*((buf)+2-fmt->Gshift/8) = g;		\
slouken@0
   269
				*((buf)+2-fmt->Bshift/8) = b;		\
slouken@0
   270
			}						\
slouken@0
   271
		}							\
slouken@0
   272
		break;							\
slouken@0
   273
									\
slouken@0
   274
		case 4: {						\
icculus@1162
   275
			Uint32 Pixel;					\
slouken@0
   276
									\
icculus@1162
   277
			PIXEL_FROM_RGB(Pixel, fmt, r, g, b);		\
icculus@1162
   278
			*((Uint32 *)(buf)) = Pixel;			\
slouken@0
   279
		}							\
slouken@0
   280
		break;							\
slouken@0
   281
	}								\
slouken@0
   282
}
slouken@0
   283
slouken@0
   284
/* FIXME: Should we rescale alpha into 0..255 here? */
icculus@1162
   285
#define RGBA_FROM_PIXEL(Pixel, fmt, r, g, b, a)				\
slouken@0
   286
{									\
icculus@1162
   287
	r = ((Pixel&fmt->Rmask)>>fmt->Rshift)<<fmt->Rloss; 		\
icculus@1162
   288
	g = ((Pixel&fmt->Gmask)>>fmt->Gshift)<<fmt->Gloss; 		\
icculus@1162
   289
	b = ((Pixel&fmt->Bmask)>>fmt->Bshift)<<fmt->Bloss; 		\
icculus@1162
   290
	a = ((Pixel&fmt->Amask)>>fmt->Ashift)<<fmt->Aloss;	 	\
slouken@0
   291
}
icculus@1162
   292
#define RGBA_FROM_8888(Pixel, fmt, r, g, b, a)	\
slouken@0
   293
{						\
icculus@1162
   294
	r = (Pixel&fmt->Rmask)>>fmt->Rshift;	\
icculus@1162
   295
	g = (Pixel&fmt->Gmask)>>fmt->Gshift;	\
icculus@1162
   296
	b = (Pixel&fmt->Bmask)>>fmt->Bshift;	\
icculus@1162
   297
	a = (Pixel&fmt->Amask)>>fmt->Ashift;	\
slouken@0
   298
}
icculus@1162
   299
#define RGBA_FROM_RGBA8888(Pixel, r, g, b, a)				\
slouken@0
   300
{									\
icculus@1162
   301
	r = (Pixel>>24);						\
icculus@1162
   302
	g = ((Pixel>>16)&0xFF);						\
icculus@1162
   303
	b = ((Pixel>>8)&0xFF);						\
icculus@1162
   304
	a = (Pixel&0xFF);						\
slouken@0
   305
}
icculus@1162
   306
#define RGBA_FROM_ARGB8888(Pixel, r, g, b, a)				\
slouken@0
   307
{									\
icculus@1162
   308
	r = ((Pixel>>16)&0xFF);						\
icculus@1162
   309
	g = ((Pixel>>8)&0xFF);						\
icculus@1162
   310
	b = (Pixel&0xFF);						\
icculus@1162
   311
	a = (Pixel>>24);						\
slouken@0
   312
}
icculus@1162
   313
#define RGBA_FROM_ABGR8888(Pixel, r, g, b, a)				\
slouken@0
   314
{									\
icculus@1162
   315
	r = (Pixel&0xFF);						\
icculus@1162
   316
	g = ((Pixel>>8)&0xFF);						\
icculus@1162
   317
	b = ((Pixel>>16)&0xFF);						\
icculus@1162
   318
	a = (Pixel>>24);						\
slouken@0
   319
}
slouken@3054
   320
#define RGBA_FROM_BGRA8888(Pixel, r, g, b, a)				\
slouken@3054
   321
{									\
slouken@3054
   322
	r = ((Pixel>>8)&0xFF);						\
slouken@3054
   323
	g = ((Pixel>>16)&0xFF);						\
slouken@3054
   324
	b = (Pixel>>24);						\
slouken@3054
   325
	a = (Pixel&0xFF);						\
slouken@3054
   326
}
icculus@1162
   327
#define DISEMBLE_RGBA(buf, bpp, fmt, Pixel, r, g, b, a)			   \
slouken@0
   328
do {									   \
slouken@0
   329
	switch (bpp) {							   \
slouken@0
   330
		case 2:							   \
icculus@1162
   331
			Pixel = *((Uint16 *)(buf));			   \
slouken@2824
   332
			RGBA_FROM_PIXEL(Pixel, fmt, r, g, b, a);	   \
slouken@0
   333
		break;							   \
slouken@0
   334
									   \
slouken@2824
   335
		case 3:	{						   \
slouken@5426
   336
            Pixel = 0; \
slouken@2824
   337
                        if (SDL_BYTEORDER == SDL_LIL_ENDIAN) {		   \
slouken@2824
   338
			        r = *((buf)+fmt->Rshift/8);		   \
slouken@2824
   339
				g = *((buf)+fmt->Gshift/8);		   \
slouken@2824
   340
				b = *((buf)+fmt->Bshift/8);		   \
slouken@0
   341
			} else {					   \
slouken@2824
   342
			        r = *((buf)+2-fmt->Rshift/8);		   \
slouken@2824
   343
				g = *((buf)+2-fmt->Gshift/8);		   \
slouken@2824
   344
				b = *((buf)+2-fmt->Bshift/8);		   \
slouken@0
   345
			}						   \
slouken@2824
   346
			a = 0xFF;					   \
slouken@0
   347
		}							   \
slouken@0
   348
		break;							   \
slouken@0
   349
									   \
slouken@0
   350
		case 4:							   \
icculus@1162
   351
			Pixel = *((Uint32 *)(buf));			   \
slouken@2824
   352
			RGBA_FROM_PIXEL(Pixel, fmt, r, g, b, a);	   \
slouken@0
   353
		break;							   \
slouken@0
   354
									   \
slouken@0
   355
		default:						   \
slouken@5426
   356
                /* stop gcc complaints */		   \
slouken@5426
   357
		        Pixel = 0;              \
slouken@5426
   358
                r = g = b = a = 0;      \
slouken@0
   359
		break;							   \
slouken@0
   360
	}								   \
slouken@2824
   361
} while (0)
slouken@0
   362
slouken@0
   363
/* FIXME: this isn't correct, especially for Alpha (maximum != 255) */
icculus@1162
   364
#define PIXEL_FROM_RGBA(Pixel, fmt, r, g, b, a)				\
slouken@0
   365
{									\
icculus@1162
   366
	Pixel = ((r>>fmt->Rloss)<<fmt->Rshift)|				\
slouken@0
   367
		((g>>fmt->Gloss)<<fmt->Gshift)|				\
slouken@0
   368
		((b>>fmt->Bloss)<<fmt->Bshift)|				\
slouken@535
   369
		((a>>fmt->Aloss)<<fmt->Ashift);				\
slouken@0
   370
}
slouken@0
   371
#define ASSEMBLE_RGBA(buf, bpp, fmt, r, g, b, a)			\
slouken@0
   372
{									\
slouken@0
   373
	switch (bpp) {							\
slouken@0
   374
		case 2: {						\
icculus@1162
   375
			Uint16 Pixel;					\
slouken@0
   376
									\
icculus@1162
   377
			PIXEL_FROM_RGBA(Pixel, fmt, r, g, b, a);	\
slouken@5423
   378
			*((Uint16 *)(buf)) = Pixel | NDS_BIT15;		\
slouken@0
   379
		}							\
slouken@0
   380
		break;							\
slouken@0
   381
									\
slouken@2824
   382
		case 3: {						\
slouken@2824
   383
                        if (SDL_BYTEORDER == SDL_LIL_ENDIAN) {		\
slouken@0
   384
			        *((buf)+fmt->Rshift/8) = r;		\
slouken@0
   385
				*((buf)+fmt->Gshift/8) = g;		\
slouken@0
   386
				*((buf)+fmt->Bshift/8) = b;		\
slouken@0
   387
			} else {					\
slouken@0
   388
			        *((buf)+2-fmt->Rshift/8) = r;		\
slouken@0
   389
				*((buf)+2-fmt->Gshift/8) = g;		\
slouken@0
   390
				*((buf)+2-fmt->Bshift/8) = b;		\
slouken@0
   391
			}						\
slouken@0
   392
		}							\
slouken@0
   393
		break;							\
slouken@0
   394
									\
slouken@0
   395
		case 4: {						\
icculus@1162
   396
			Uint32 Pixel;					\
slouken@0
   397
									\
icculus@1162
   398
			PIXEL_FROM_RGBA(Pixel, fmt, r, g, b, a);	\
icculus@1162
   399
			*((Uint32 *)(buf)) = Pixel;			\
slouken@0
   400
		}							\
slouken@0
   401
		break;							\
slouken@0
   402
	}								\
slouken@0
   403
}
slouken@0
   404
icculus@1162
   405
/* Blend the RGB values of two Pixels based on a source alpha value */
slouken@0
   406
#define ALPHA_BLEND(sR, sG, sB, A, dR, dG, dB)	\
slouken@0
   407
do {						\
icculus@3631
   408
	dR = ((((int)(sR-dR)*(int)A)/255)+dR);	\
icculus@3631
   409
	dG = ((((int)(sG-dG)*(int)A)/255)+dG);	\
icculus@3631
   410
	dB = ((((int)(sB-dB)*(int)A)/255)+dB);	\
icculus@1047
   411
} while(0)
icculus@1047
   412
icculus@1047
   413
slouken@0
   414
/* This is a very useful loop for optimizing blitters */
slouken@553
   415
#if defined(_MSC_VER) && (_MSC_VER == 1300)
slouken@553
   416
/* There's a bug in the Visual C++ 7 optimizer when compiling this code */
slouken@553
   417
#else
slouken@0
   418
#define USE_DUFFS_LOOP
slouken@553
   419
#endif
slouken@0
   420
#ifdef USE_DUFFS_LOOP
slouken@0
   421
slouken@0
   422
/* 8-times unrolled loop */
slouken@0
   423
#define DUFFS_LOOP8(pixel_copy_increment, width)			\
slouken@0
   424
{ int n = (width+7)/8;							\
slouken@91
   425
	switch (width & 7) {						\
slouken@0
   426
	case 0: do {	pixel_copy_increment;				\
slouken@0
   427
	case 7:		pixel_copy_increment;				\
slouken@0
   428
	case 6:		pixel_copy_increment;				\
slouken@0
   429
	case 5:		pixel_copy_increment;				\
slouken@0
   430
	case 4:		pixel_copy_increment;				\
slouken@0
   431
	case 3:		pixel_copy_increment;				\
slouken@0
   432
	case 2:		pixel_copy_increment;				\
slouken@0
   433
	case 1:		pixel_copy_increment;				\
slouken@0
   434
		} while ( --n > 0 );					\
slouken@0
   435
	}								\
slouken@0
   436
}
slouken@0
   437
slouken@0
   438
/* 4-times unrolled loop */
slouken@0
   439
#define DUFFS_LOOP4(pixel_copy_increment, width)			\
slouken@0
   440
{ int n = (width+3)/4;							\
slouken@91
   441
	switch (width & 3) {						\
slouken@0
   442
	case 0: do {	pixel_copy_increment;				\
slouken@0
   443
	case 3:		pixel_copy_increment;				\
slouken@0
   444
	case 2:		pixel_copy_increment;				\
slouken@0
   445
	case 1:		pixel_copy_increment;				\
slouken@3035
   446
		} while (--n > 0);					\
slouken@689
   447
	}								\
slouken@689
   448
}
slouken@689
   449
slouken@0
   450
/* Use the 8-times version of the loop by default */
slouken@0
   451
#define DUFFS_LOOP(pixel_copy_increment, width)				\
slouken@0
   452
	DUFFS_LOOP8(pixel_copy_increment, width)
slouken@0
   453
slouken@3035
   454
/* Special version of Duff's device for even more optimization */
slouken@3035
   455
#define DUFFS_LOOP_124(pixel_copy_increment1,				\
slouken@3035
   456
                       pixel_copy_increment2,				\
slouken@3035
   457
                       pixel_copy_increment4, width)			\
slouken@3035
   458
{ int n = width;							\
slouken@3035
   459
	if (n & 1) {							\
slouken@3035
   460
		pixel_copy_increment1; n -= 1;				\
slouken@3035
   461
	}								\
slouken@3035
   462
	if (n & 2) {							\
slouken@3035
   463
		pixel_copy_increment2; n -= 2;				\
slouken@3035
   464
	}								\
slouken@3035
   465
	if (n) {							\
slouken@3035
   466
		n = (n+7)/ 8;						\
slouken@3035
   467
		switch (n & 4) {					\
slouken@3035
   468
		case 0: do {	pixel_copy_increment4;			\
slouken@3035
   469
		case 4:		pixel_copy_increment4;			\
slouken@3035
   470
			} while (--n > 0);				\
slouken@3035
   471
		}							\
slouken@3035
   472
	}								\
slouken@689
   473
}
slouken@689
   474
slouken@3035
   475
#else
slouken@689
   476
slouken@689
   477
/* Don't use Duff's device to unroll loops */
slouken@0
   478
#define DUFFS_LOOP(pixel_copy_increment, width)				\
slouken@0
   479
{ int n;								\
slouken@0
   480
	for ( n=width; n > 0; --n ) {					\
slouken@0
   481
		pixel_copy_increment;					\
slouken@0
   482
	}								\
slouken@0
   483
}
slouken@0
   484
#define DUFFS_LOOP8(pixel_copy_increment, width)			\
slouken@0
   485
	DUFFS_LOOP(pixel_copy_increment, width)
slouken@0
   486
#define DUFFS_LOOP4(pixel_copy_increment, width)			\
slouken@0
   487
	DUFFS_LOOP(pixel_copy_increment, width)
slouken@3035
   488
#define DUFFS_LOOP_124(pixel_copy_increment1,				\
slouken@3035
   489
                       pixel_copy_increment2,				\
slouken@3035
   490
                       pixel_copy_increment4, width)			\
slouken@3035
   491
	DUFFS_LOOP(pixel_copy_increment1, width)
slouken@0
   492
slouken@0
   493
#endif /* USE_DUFFS_LOOP */
slouken@0
   494
slouken@0
   495
/* Prevent Visual C++ 6.0 from printing out stupid warnings */
slouken@0
   496
#if defined(_MSC_VER) && (_MSC_VER >= 600)
slouken@0
   497
#pragma warning(disable: 4550)
slouken@0
   498
#endif
slouken@0
   499
slouken@0
   500
#endif /* _SDL_blit_h */
slouken@2898
   501
slouken@1895
   502
/* vi: set ts=4 sw=4 expandtab: */