Mercurial > sdl-ios-xcode
annotate src/video/SDL_blit.h @ 2249:5a58b57b6724
Added SSE and MMX optimization for SDL_FillRect()
author | Sam Lantinga <slouken@libsdl.org> |
---|---|
date | Thu, 16 Aug 2007 05:56:24 +0000 |
parents | 93994f65c74c |
children | e1d228456537 |
rev | line source |
---|---|
0 | 1 /* |
2 SDL - Simple DirectMedia Layer | |
1312
c9b51268668f
Updated copyright information and removed rcs id lines (problematic in branch merges)
Sam Lantinga <slouken@libsdl.org>
parents:
1162
diff
changeset
|
3 Copyright (C) 1997-2006 Sam Lantinga |
0 | 4 |
5 This library is free software; you can redistribute it and/or | |
1312
c9b51268668f
Updated copyright information and removed rcs id lines (problematic in branch merges)
Sam Lantinga <slouken@libsdl.org>
parents:
1162
diff
changeset
|
6 modify it under the terms of the GNU Lesser General Public |
0 | 7 License as published by the Free Software Foundation; either |
1312
c9b51268668f
Updated copyright information and removed rcs id lines (problematic in branch merges)
Sam Lantinga <slouken@libsdl.org>
parents:
1162
diff
changeset
|
8 version 2.1 of the License, or (at your option) any later version. |
0 | 9 |
10 This library is distributed in the hope that it will be useful, | |
11 but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
1312
c9b51268668f
Updated copyright information and removed rcs id lines (problematic in branch merges)
Sam Lantinga <slouken@libsdl.org>
parents:
1162
diff
changeset
|
13 Lesser General Public License for more details. |
0 | 14 |
1312
c9b51268668f
Updated copyright information and removed rcs id lines (problematic in branch merges)
Sam Lantinga <slouken@libsdl.org>
parents:
1162
diff
changeset
|
15 You should have received a copy of the GNU Lesser General Public |
c9b51268668f
Updated copyright information and removed rcs id lines (problematic in branch merges)
Sam Lantinga <slouken@libsdl.org>
parents:
1162
diff
changeset
|
16 License along with this library; if not, write to the Free Software |
c9b51268668f
Updated copyright information and removed rcs id lines (problematic in branch merges)
Sam Lantinga <slouken@libsdl.org>
parents:
1162
diff
changeset
|
17 Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA |
0 | 18 |
19 Sam Lantinga | |
252
e8157fcb3114
Updated the source with the correct e-mail address
Sam Lantinga <slouken@libsdl.org>
parents:
163
diff
changeset
|
20 slouken@libsdl.org |
0 | 21 */ |
1402
d910939febfa
Use consistent identifiers for the various platforms we support.
Sam Lantinga <slouken@libsdl.org>
parents:
1312
diff
changeset
|
22 #include "SDL_config.h" |
0 | 23 |
24 #ifndef _SDL_blit_h | |
25 #define _SDL_blit_h | |
26 | |
2249
5a58b57b6724
Added SSE and MMX optimization for SDL_FillRect()
Sam Lantinga <slouken@libsdl.org>
parents:
2247
diff
changeset
|
27 #ifdef __MMX__ |
5a58b57b6724
Added SSE and MMX optimization for SDL_FillRect()
Sam Lantinga <slouken@libsdl.org>
parents:
2247
diff
changeset
|
28 #include <mmintrin.h> |
5a58b57b6724
Added SSE and MMX optimization for SDL_FillRect()
Sam Lantinga <slouken@libsdl.org>
parents:
2247
diff
changeset
|
29 #endif |
5a58b57b6724
Added SSE and MMX optimization for SDL_FillRect()
Sam Lantinga <slouken@libsdl.org>
parents:
2247
diff
changeset
|
30 #ifdef __SSE__ |
5a58b57b6724
Added SSE and MMX optimization for SDL_FillRect()
Sam Lantinga <slouken@libsdl.org>
parents:
2247
diff
changeset
|
31 #include <xmmintrin.h> |
5a58b57b6724
Added SSE and MMX optimization for SDL_FillRect()
Sam Lantinga <slouken@libsdl.org>
parents:
2247
diff
changeset
|
32 #endif |
5a58b57b6724
Added SSE and MMX optimization for SDL_FillRect()
Sam Lantinga <slouken@libsdl.org>
parents:
2247
diff
changeset
|
33 |
0 | 34 #include "SDL_endian.h" |
35 | |
36 /* The structure passed to the low level blit functions */ | |
1895
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
37 typedef struct |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
38 { |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
39 Uint8 *s_pixels; |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
40 int s_width; |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
41 int s_height; |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
42 int s_skip; |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
43 Uint8 *d_pixels; |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
44 int d_width; |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
45 int d_height; |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
46 int d_skip; |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
47 void *aux_data; |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
48 SDL_PixelFormat *src; |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
49 Uint8 *table; |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
50 SDL_PixelFormat *dst; |
0 | 51 } SDL_BlitInfo; |
52 | |
53 /* The type definition for the low level blit functions */ | |
1895
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
54 typedef void (*SDL_loblit) (SDL_BlitInfo * info); |
0 | 55 |
56 /* This is the private info structure for software accelerated blits */ | |
1895
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
57 struct private_swaccel |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
58 { |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
59 SDL_loblit blit; |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
60 void *aux_data; |
0 | 61 }; |
62 | |
63 /* Blit mapping definition */ | |
1895
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
64 typedef struct SDL_BlitMap |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
65 { |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
66 SDL_Surface *dst; |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
67 int identity; |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
68 Uint8 *table; |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
69 SDL_blit sw_blit; |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
70 struct private_swaccel *sw_data; |
0 | 71 |
1895
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
72 /* the version count matches the destination; mismatch indicates |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
73 an invalid mapping */ |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
74 unsigned int format_version; |
0 | 75 } SDL_BlitMap; |
76 | |
2247
93994f65c74c
Removed hermes since it's LGPL and not compatible with a commercial license.
Sam Lantinga <slouken@libsdl.org>
parents:
1895
diff
changeset
|
77 #define SDL_BLIT_ANY 0x00000000 |
93994f65c74c
Removed hermes since it's LGPL and not compatible with a commercial license.
Sam Lantinga <slouken@libsdl.org>
parents:
1895
diff
changeset
|
78 #define SDL_BLIT_MMX 0x00000001 |
93994f65c74c
Removed hermes since it's LGPL and not compatible with a commercial license.
Sam Lantinga <slouken@libsdl.org>
parents:
1895
diff
changeset
|
79 #define SDL_BLIT_SSE 0x00000002 |
93994f65c74c
Removed hermes since it's LGPL and not compatible with a commercial license.
Sam Lantinga <slouken@libsdl.org>
parents:
1895
diff
changeset
|
80 #define SDL_BLIT_ALTIVEC_PREFETCH 0x00000004 |
93994f65c74c
Removed hermes since it's LGPL and not compatible with a commercial license.
Sam Lantinga <slouken@libsdl.org>
parents:
1895
diff
changeset
|
81 #define SDL_BLIT_ALTIVEC_NOPREFETCH 0x00000008 |
93994f65c74c
Removed hermes since it's LGPL and not compatible with a commercial license.
Sam Lantinga <slouken@libsdl.org>
parents:
1895
diff
changeset
|
82 |
93994f65c74c
Removed hermes since it's LGPL and not compatible with a commercial license.
Sam Lantinga <slouken@libsdl.org>
parents:
1895
diff
changeset
|
83 typedef struct SDL_BlitEntry |
93994f65c74c
Removed hermes since it's LGPL and not compatible with a commercial license.
Sam Lantinga <slouken@libsdl.org>
parents:
1895
diff
changeset
|
84 { |
93994f65c74c
Removed hermes since it's LGPL and not compatible with a commercial license.
Sam Lantinga <slouken@libsdl.org>
parents:
1895
diff
changeset
|
85 Uint32 features; |
93994f65c74c
Removed hermes since it's LGPL and not compatible with a commercial license.
Sam Lantinga <slouken@libsdl.org>
parents:
1895
diff
changeset
|
86 SDL_loblit blit; |
93994f65c74c
Removed hermes since it's LGPL and not compatible with a commercial license.
Sam Lantinga <slouken@libsdl.org>
parents:
1895
diff
changeset
|
87 } SDL_BlitEntry; |
0 | 88 |
89 /* Functions found in SDL_blit.c */ | |
1895
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
90 extern int SDL_CalculateBlit(SDL_Surface * surface); |
0 | 91 |
92 /* Functions found in SDL_blit_{0,1,N,A}.c */ | |
1895
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
93 extern SDL_loblit SDL_CalculateBlit0(SDL_Surface * surface, int complex); |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
94 extern SDL_loblit SDL_CalculateBlit1(SDL_Surface * surface, int complex); |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
95 extern SDL_loblit SDL_CalculateBlitN(SDL_Surface * surface, int complex); |
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
96 extern SDL_loblit SDL_CalculateAlphaBlit(SDL_Surface * surface, int complex); |
0 | 97 |
98 /* | |
99 * Useful macros for blitting routines | |
100 */ | |
101 | |
2249
5a58b57b6724
Added SSE and MMX optimization for SDL_FillRect()
Sam Lantinga <slouken@libsdl.org>
parents:
2247
diff
changeset
|
102 #if defined(__GNUC__) |
5a58b57b6724
Added SSE and MMX optimization for SDL_FillRect()
Sam Lantinga <slouken@libsdl.org>
parents:
2247
diff
changeset
|
103 #define DECLARE_ALIGNED(t,v,a) t __attribute__((aligned(a))) v |
5a58b57b6724
Added SSE and MMX optimization for SDL_FillRect()
Sam Lantinga <slouken@libsdl.org>
parents:
2247
diff
changeset
|
104 #elif defined(_MSC_VER) |
5a58b57b6724
Added SSE and MMX optimization for SDL_FillRect()
Sam Lantinga <slouken@libsdl.org>
parents:
2247
diff
changeset
|
105 #define DECLARE_ALIGNED(t,v,a) t __declspec(align(a)) v |
5a58b57b6724
Added SSE and MMX optimization for SDL_FillRect()
Sam Lantinga <slouken@libsdl.org>
parents:
2247
diff
changeset
|
106 #else |
5a58b57b6724
Added SSE and MMX optimization for SDL_FillRect()
Sam Lantinga <slouken@libsdl.org>
parents:
2247
diff
changeset
|
107 #define DECLARE_ALIGNED(t,v,a) t v |
5a58b57b6724
Added SSE and MMX optimization for SDL_FillRect()
Sam Lantinga <slouken@libsdl.org>
parents:
2247
diff
changeset
|
108 #endif |
5a58b57b6724
Added SSE and MMX optimization for SDL_FillRect()
Sam Lantinga <slouken@libsdl.org>
parents:
2247
diff
changeset
|
109 |
0 | 110 #define FORMAT_EQUAL(A, B) \ |
111 ((A)->BitsPerPixel == (B)->BitsPerPixel \ | |
112 && ((A)->Rmask == (B)->Rmask) && ((A)->Amask == (B)->Amask)) | |
113 | |
114 /* Load pixel of the specified format from a buffer and get its R-G-B values */ | |
115 /* FIXME: rescale values to 0..255 here? */ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
116 #define RGB_FROM_PIXEL(Pixel, fmt, r, g, b) \ |
0 | 117 { \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
118 r = (((Pixel&fmt->Rmask)>>fmt->Rshift)<<fmt->Rloss); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
119 g = (((Pixel&fmt->Gmask)>>fmt->Gshift)<<fmt->Gloss); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
120 b = (((Pixel&fmt->Bmask)>>fmt->Bshift)<<fmt->Bloss); \ |
0 | 121 } |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
122 #define RGB_FROM_RGB565(Pixel, r, g, b) \ |
0 | 123 { \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
124 r = (((Pixel&0xF800)>>11)<<3); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
125 g = (((Pixel&0x07E0)>>5)<<2); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
126 b = ((Pixel&0x001F)<<3); \ |
0 | 127 } |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
128 #define RGB_FROM_RGB555(Pixel, r, g, b) \ |
0 | 129 { \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
130 r = (((Pixel&0x7C00)>>10)<<3); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
131 g = (((Pixel&0x03E0)>>5)<<3); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
132 b = ((Pixel&0x001F)<<3); \ |
0 | 133 } |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
134 #define RGB_FROM_RGB888(Pixel, r, g, b) \ |
0 | 135 { \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
136 r = ((Pixel&0xFF0000)>>16); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
137 g = ((Pixel&0xFF00)>>8); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
138 b = (Pixel&0xFF); \ |
0 | 139 } |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
140 #define RETRIEVE_RGB_PIXEL(buf, bpp, Pixel) \ |
0 | 141 do { \ |
142 switch (bpp) { \ | |
143 case 2: \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
144 Pixel = *((Uint16 *)(buf)); \ |
0 | 145 break; \ |
146 \ | |
147 case 3: { \ | |
148 Uint8 *B = (Uint8 *)(buf); \ | |
149 if(SDL_BYTEORDER == SDL_LIL_ENDIAN) { \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
150 Pixel = B[0] + (B[1] << 8) + (B[2] << 16); \ |
0 | 151 } else { \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
152 Pixel = (B[0] << 16) + (B[1] << 8) + B[2]; \ |
0 | 153 } \ |
154 } \ | |
155 break; \ | |
156 \ | |
157 case 4: \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
158 Pixel = *((Uint32 *)(buf)); \ |
0 | 159 break; \ |
160 \ | |
161 default: \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
162 Pixel = 0; /* appease gcc */ \ |
0 | 163 break; \ |
164 } \ | |
165 } while(0) | |
166 | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
167 #define DISEMBLE_RGB(buf, bpp, fmt, Pixel, r, g, b) \ |
0 | 168 do { \ |
169 switch (bpp) { \ | |
170 case 2: \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
171 Pixel = *((Uint16 *)(buf)); \ |
0 | 172 break; \ |
173 \ | |
174 case 3: { \ | |
175 Uint8 *B = (Uint8 *)buf; \ | |
176 if(SDL_BYTEORDER == SDL_LIL_ENDIAN) { \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
177 Pixel = B[0] + (B[1] << 8) + (B[2] << 16); \ |
0 | 178 } else { \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
179 Pixel = (B[0] << 16) + (B[1] << 8) + B[2]; \ |
0 | 180 } \ |
181 } \ | |
182 break; \ | |
183 \ | |
184 case 4: \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
185 Pixel = *((Uint32 *)(buf)); \ |
0 | 186 break; \ |
187 \ | |
188 default: \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
189 Pixel = 0; /* prevent gcc from complaining */ \ |
0 | 190 break; \ |
191 } \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
192 RGB_FROM_PIXEL(Pixel, fmt, r, g, b); \ |
0 | 193 } while(0) |
194 | |
195 /* Assemble R-G-B values into a specified pixel format and store them */ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
196 #define PIXEL_FROM_RGB(Pixel, fmt, r, g, b) \ |
0 | 197 { \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
198 Pixel = ((r>>fmt->Rloss)<<fmt->Rshift)| \ |
0 | 199 ((g>>fmt->Gloss)<<fmt->Gshift)| \ |
200 ((b>>fmt->Bloss)<<fmt->Bshift); \ | |
201 } | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
202 #define RGB565_FROM_RGB(Pixel, r, g, b) \ |
0 | 203 { \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
204 Pixel = ((r>>3)<<11)|((g>>2)<<5)|(b>>3); \ |
0 | 205 } |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
206 #define RGB555_FROM_RGB(Pixel, r, g, b) \ |
0 | 207 { \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
208 Pixel = ((r>>3)<<10)|((g>>3)<<5)|(b>>3); \ |
0 | 209 } |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
210 #define RGB888_FROM_RGB(Pixel, r, g, b) \ |
0 | 211 { \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
212 Pixel = (r<<16)|(g<<8)|b; \ |
0 | 213 } |
214 #define ASSEMBLE_RGB(buf, bpp, fmt, r, g, b) \ | |
215 { \ | |
216 switch (bpp) { \ | |
217 case 2: { \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
218 Uint16 Pixel; \ |
0 | 219 \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
220 PIXEL_FROM_RGB(Pixel, fmt, r, g, b); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
221 *((Uint16 *)(buf)) = Pixel; \ |
0 | 222 } \ |
223 break; \ | |
224 \ | |
225 case 3: { \ | |
226 if(SDL_BYTEORDER == SDL_LIL_ENDIAN) { \ | |
227 *((buf)+fmt->Rshift/8) = r; \ | |
228 *((buf)+fmt->Gshift/8) = g; \ | |
229 *((buf)+fmt->Bshift/8) = b; \ | |
230 } else { \ | |
231 *((buf)+2-fmt->Rshift/8) = r; \ | |
232 *((buf)+2-fmt->Gshift/8) = g; \ | |
233 *((buf)+2-fmt->Bshift/8) = b; \ | |
234 } \ | |
235 } \ | |
236 break; \ | |
237 \ | |
238 case 4: { \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
239 Uint32 Pixel; \ |
0 | 240 \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
241 PIXEL_FROM_RGB(Pixel, fmt, r, g, b); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
242 *((Uint32 *)(buf)) = Pixel; \ |
0 | 243 } \ |
244 break; \ | |
245 } \ | |
246 } | |
247 #define ASSEMBLE_RGB_AMASK(buf, bpp, fmt, r, g, b, Amask) \ | |
248 { \ | |
249 switch (bpp) { \ | |
250 case 2: { \ | |
251 Uint16 *bufp; \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
252 Uint16 Pixel; \ |
0 | 253 \ |
254 bufp = (Uint16 *)buf; \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
255 PIXEL_FROM_RGB(Pixel, fmt, r, g, b); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
256 *bufp = Pixel | (*bufp & Amask); \ |
0 | 257 } \ |
258 break; \ | |
259 \ | |
260 case 3: { \ | |
261 if(SDL_BYTEORDER == SDL_LIL_ENDIAN) { \ | |
262 *((buf)+fmt->Rshift/8) = r; \ | |
263 *((buf)+fmt->Gshift/8) = g; \ | |
264 *((buf)+fmt->Bshift/8) = b; \ | |
265 } else { \ | |
266 *((buf)+2-fmt->Rshift/8) = r; \ | |
267 *((buf)+2-fmt->Gshift/8) = g; \ | |
268 *((buf)+2-fmt->Bshift/8) = b; \ | |
269 } \ | |
270 } \ | |
271 break; \ | |
272 \ | |
273 case 4: { \ | |
274 Uint32 *bufp; \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
275 Uint32 Pixel; \ |
0 | 276 \ |
277 bufp = (Uint32 *)buf; \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
278 PIXEL_FROM_RGB(Pixel, fmt, r, g, b); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
279 *bufp = Pixel | (*bufp & Amask); \ |
0 | 280 } \ |
281 break; \ | |
282 } \ | |
283 } | |
284 | |
285 /* FIXME: Should we rescale alpha into 0..255 here? */ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
286 #define RGBA_FROM_PIXEL(Pixel, fmt, r, g, b, a) \ |
0 | 287 { \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
288 r = ((Pixel&fmt->Rmask)>>fmt->Rshift)<<fmt->Rloss; \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
289 g = ((Pixel&fmt->Gmask)>>fmt->Gshift)<<fmt->Gloss; \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
290 b = ((Pixel&fmt->Bmask)>>fmt->Bshift)<<fmt->Bloss; \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
291 a = ((Pixel&fmt->Amask)>>fmt->Ashift)<<fmt->Aloss; \ |
0 | 292 } |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
293 #define RGBA_FROM_8888(Pixel, fmt, r, g, b, a) \ |
0 | 294 { \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
295 r = (Pixel&fmt->Rmask)>>fmt->Rshift; \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
296 g = (Pixel&fmt->Gmask)>>fmt->Gshift; \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
297 b = (Pixel&fmt->Bmask)>>fmt->Bshift; \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
298 a = (Pixel&fmt->Amask)>>fmt->Ashift; \ |
0 | 299 } |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
300 #define RGBA_FROM_RGBA8888(Pixel, r, g, b, a) \ |
0 | 301 { \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
302 r = (Pixel>>24); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
303 g = ((Pixel>>16)&0xFF); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
304 b = ((Pixel>>8)&0xFF); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
305 a = (Pixel&0xFF); \ |
0 | 306 } |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
307 #define RGBA_FROM_ARGB8888(Pixel, r, g, b, a) \ |
0 | 308 { \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
309 r = ((Pixel>>16)&0xFF); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
310 g = ((Pixel>>8)&0xFF); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
311 b = (Pixel&0xFF); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
312 a = (Pixel>>24); \ |
0 | 313 } |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
314 #define RGBA_FROM_ABGR8888(Pixel, r, g, b, a) \ |
0 | 315 { \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
316 r = (Pixel&0xFF); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
317 g = ((Pixel>>8)&0xFF); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
318 b = ((Pixel>>16)&0xFF); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
319 a = (Pixel>>24); \ |
0 | 320 } |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
321 #define DISEMBLE_RGBA(buf, bpp, fmt, Pixel, r, g, b, a) \ |
0 | 322 do { \ |
323 switch (bpp) { \ | |
324 case 2: \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
325 Pixel = *((Uint16 *)(buf)); \ |
0 | 326 break; \ |
327 \ | |
328 case 3: {/* FIXME: broken code (no alpha) */ \ | |
329 Uint8 *b = (Uint8 *)buf; \ | |
330 if(SDL_BYTEORDER == SDL_LIL_ENDIAN) { \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
331 Pixel = b[0] + (b[1] << 8) + (b[2] << 16); \ |
0 | 332 } else { \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
333 Pixel = (b[0] << 16) + (b[1] << 8) + b[2]; \ |
0 | 334 } \ |
335 } \ | |
336 break; \ | |
337 \ | |
338 case 4: \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
339 Pixel = *((Uint32 *)(buf)); \ |
0 | 340 break; \ |
341 \ | |
342 default: \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
343 Pixel = 0; /* stop gcc complaints */ \ |
0 | 344 break; \ |
345 } \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
346 RGBA_FROM_PIXEL(Pixel, fmt, r, g, b, a); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
347 Pixel &= ~fmt->Amask; \ |
0 | 348 } while(0) |
349 | |
350 /* FIXME: this isn't correct, especially for Alpha (maximum != 255) */ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
351 #define PIXEL_FROM_RGBA(Pixel, fmt, r, g, b, a) \ |
0 | 352 { \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
353 Pixel = ((r>>fmt->Rloss)<<fmt->Rshift)| \ |
0 | 354 ((g>>fmt->Gloss)<<fmt->Gshift)| \ |
355 ((b>>fmt->Bloss)<<fmt->Bshift)| \ | |
535
917cc5c56176
Fixed alpha blending bug (thanks Glenn!)
Sam Lantinga <slouken@libsdl.org>
parents:
297
diff
changeset
|
356 ((a>>fmt->Aloss)<<fmt->Ashift); \ |
0 | 357 } |
358 #define ASSEMBLE_RGBA(buf, bpp, fmt, r, g, b, a) \ | |
359 { \ | |
360 switch (bpp) { \ | |
361 case 2: { \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
362 Uint16 Pixel; \ |
0 | 363 \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
364 PIXEL_FROM_RGBA(Pixel, fmt, r, g, b, a); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
365 *((Uint16 *)(buf)) = Pixel; \ |
0 | 366 } \ |
367 break; \ | |
368 \ | |
369 case 3: { /* FIXME: broken code (no alpha) */ \ | |
370 if(SDL_BYTEORDER == SDL_LIL_ENDIAN) { \ | |
371 *((buf)+fmt->Rshift/8) = r; \ | |
372 *((buf)+fmt->Gshift/8) = g; \ | |
373 *((buf)+fmt->Bshift/8) = b; \ | |
374 } else { \ | |
375 *((buf)+2-fmt->Rshift/8) = r; \ | |
376 *((buf)+2-fmt->Gshift/8) = g; \ | |
377 *((buf)+2-fmt->Bshift/8) = b; \ | |
378 } \ | |
379 } \ | |
380 break; \ | |
381 \ | |
382 case 4: { \ | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
383 Uint32 Pixel; \ |
0 | 384 \ |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
385 PIXEL_FROM_RGBA(Pixel, fmt, r, g, b, a); \ |
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
386 *((Uint32 *)(buf)) = Pixel; \ |
0 | 387 } \ |
388 break; \ | |
389 } \ | |
390 } | |
391 | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
392 /* Blend the RGB values of two Pixels based on a source alpha value */ |
0 | 393 #define ALPHA_BLEND(sR, sG, sB, A, dR, dG, dB) \ |
394 do { \ | |
395 dR = (((sR-dR)*(A))>>8)+dR; \ | |
396 dG = (((sG-dG)*(A))>>8)+dG; \ | |
397 dB = (((sB-dB)*(A))>>8)+dB; \ | |
398 } while(0) | |
399 | |
1162
2651158f59b8
Enable altivec blitters on PowerPC Linux, and some fixes for recent
Ryan C. Gordon <icculus@icculus.org>
parents:
1047
diff
changeset
|
400 /* Blend the RGB values of two Pixels based on a source alpha value */ |
1047
ffaaf7ecf685
Altivec-optimized blitters!
Ryan C. Gordon <icculus@icculus.org>
parents:
769
diff
changeset
|
401 #define ACCURATE_ALPHA_BLEND(sR, sG, sB, sA, dR, dG, dB) \ |
ffaaf7ecf685
Altivec-optimized blitters!
Ryan C. Gordon <icculus@icculus.org>
parents:
769
diff
changeset
|
402 do { \ |
ffaaf7ecf685
Altivec-optimized blitters!
Ryan C. Gordon <icculus@icculus.org>
parents:
769
diff
changeset
|
403 unsigned tR, tG, tB, tA; \ |
ffaaf7ecf685
Altivec-optimized blitters!
Ryan C. Gordon <icculus@icculus.org>
parents:
769
diff
changeset
|
404 tA = 255 - sA; \ |
ffaaf7ecf685
Altivec-optimized blitters!
Ryan C. Gordon <icculus@icculus.org>
parents:
769
diff
changeset
|
405 tR = 1 + (sR * sA) + (dR * tA); \ |
ffaaf7ecf685
Altivec-optimized blitters!
Ryan C. Gordon <icculus@icculus.org>
parents:
769
diff
changeset
|
406 dR = (tR + (tR >> 8)) >> 8; \ |
ffaaf7ecf685
Altivec-optimized blitters!
Ryan C. Gordon <icculus@icculus.org>
parents:
769
diff
changeset
|
407 tG = 1 + (sG * sA) + (dG * tA); \ |
ffaaf7ecf685
Altivec-optimized blitters!
Ryan C. Gordon <icculus@icculus.org>
parents:
769
diff
changeset
|
408 dG = (tG + (tG >> 8)) >> 8; \ |
ffaaf7ecf685
Altivec-optimized blitters!
Ryan C. Gordon <icculus@icculus.org>
parents:
769
diff
changeset
|
409 tB = 1 + (sB * sA) + (dB * tA); \ |
ffaaf7ecf685
Altivec-optimized blitters!
Ryan C. Gordon <icculus@icculus.org>
parents:
769
diff
changeset
|
410 dB = (tB + (tB >> 8)) >> 8; \ |
ffaaf7ecf685
Altivec-optimized blitters!
Ryan C. Gordon <icculus@icculus.org>
parents:
769
diff
changeset
|
411 } while(0) |
ffaaf7ecf685
Altivec-optimized blitters!
Ryan C. Gordon <icculus@icculus.org>
parents:
769
diff
changeset
|
412 |
ffaaf7ecf685
Altivec-optimized blitters!
Ryan C. Gordon <icculus@icculus.org>
parents:
769
diff
changeset
|
413 |
0 | 414 /* This is a very useful loop for optimizing blitters */ |
553
417f8709e648
There's a bug in the VC7 optimizer relating to the duff loop optimization
Sam Lantinga <slouken@libsdl.org>
parents:
535
diff
changeset
|
415 #if defined(_MSC_VER) && (_MSC_VER == 1300) |
417f8709e648
There's a bug in the VC7 optimizer relating to the duff loop optimization
Sam Lantinga <slouken@libsdl.org>
parents:
535
diff
changeset
|
416 /* There's a bug in the Visual C++ 7 optimizer when compiling this code */ |
417f8709e648
There's a bug in the VC7 optimizer relating to the duff loop optimization
Sam Lantinga <slouken@libsdl.org>
parents:
535
diff
changeset
|
417 #else |
0 | 418 #define USE_DUFFS_LOOP |
553
417f8709e648
There's a bug in the VC7 optimizer relating to the duff loop optimization
Sam Lantinga <slouken@libsdl.org>
parents:
535
diff
changeset
|
419 #endif |
0 | 420 #ifdef USE_DUFFS_LOOP |
421 | |
422 /* 8-times unrolled loop */ | |
423 #define DUFFS_LOOP8(pixel_copy_increment, width) \ | |
424 { int n = (width+7)/8; \ | |
91
e85e03f195b4
From: "Markus F.X.J. Oberhumer"
Sam Lantinga <slouken@lokigames.com>
parents:
0
diff
changeset
|
425 switch (width & 7) { \ |
0 | 426 case 0: do { pixel_copy_increment; \ |
427 case 7: pixel_copy_increment; \ | |
428 case 6: pixel_copy_increment; \ | |
429 case 5: pixel_copy_increment; \ | |
430 case 4: pixel_copy_increment; \ | |
431 case 3: pixel_copy_increment; \ | |
432 case 2: pixel_copy_increment; \ | |
433 case 1: pixel_copy_increment; \ | |
434 } while ( --n > 0 ); \ | |
435 } \ | |
436 } | |
437 | |
438 /* 4-times unrolled loop */ | |
439 #define DUFFS_LOOP4(pixel_copy_increment, width) \ | |
440 { int n = (width+3)/4; \ | |
91
e85e03f195b4
From: "Markus F.X.J. Oberhumer"
Sam Lantinga <slouken@lokigames.com>
parents:
0
diff
changeset
|
441 switch (width & 3) { \ |
0 | 442 case 0: do { pixel_copy_increment; \ |
443 case 3: pixel_copy_increment; \ | |
444 case 2: pixel_copy_increment; \ | |
445 case 1: pixel_copy_increment; \ | |
446 } while ( --n > 0 ); \ | |
447 } \ | |
448 } | |
449 | |
689
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
450 /* 2 - times unrolled loop */ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
451 #define DUFFS_LOOP_DOUBLE2(pixel_copy_increment, \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
452 double_pixel_copy_increment, width) \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
453 { int n, w = width; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
454 if( w & 1 ) { \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
455 pixel_copy_increment; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
456 w--; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
457 } \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
458 if ( w > 0 ) { \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
459 n = ( w + 2) / 4; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
460 switch( w & 2 ) { \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
461 case 0: do { double_pixel_copy_increment; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
462 case 2: double_pixel_copy_increment; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
463 } while ( --n > 0 ); \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
464 } \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
465 } \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
466 } |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
467 |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
468 /* 2 - times unrolled loop 4 pixels */ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
469 #define DUFFS_LOOP_QUATRO2(pixel_copy_increment, \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
470 double_pixel_copy_increment, \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
471 quatro_pixel_copy_increment, width) \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
472 { int n, w = width; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
473 if(w & 1) { \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
474 pixel_copy_increment; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
475 w--; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
476 } \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
477 if(w & 2) { \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
478 double_pixel_copy_increment; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
479 w -= 2; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
480 } \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
481 if ( w > 0 ) { \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
482 n = ( w + 7 ) / 8; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
483 switch( w & 4 ) { \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
484 case 0: do { quatro_pixel_copy_increment; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
485 case 4: quatro_pixel_copy_increment; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
486 } while ( --n > 0 ); \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
487 } \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
488 } \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
489 } |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
490 |
0 | 491 /* Use the 8-times version of the loop by default */ |
492 #define DUFFS_LOOP(pixel_copy_increment, width) \ | |
493 DUFFS_LOOP8(pixel_copy_increment, width) | |
494 | |
495 #else | |
496 | |
497 /* Don't use Duff's device to unroll loops */ | |
689
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
498 #define DUFFS_LOOP_DOUBLE2(pixel_copy_increment, \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
499 double_pixel_copy_increment, width) \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
500 { int n = width; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
501 if( n & 1 ) { \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
502 pixel_copy_increment; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
503 n--; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
504 } \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
505 n=n>>1; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
506 for(; n > 0; --n) { \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
507 double_pixel_copy_increment; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
508 } \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
509 } |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
510 |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
511 /* Don't use Duff's device to unroll loops */ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
512 #define DUFFS_LOOP_QUATRO2(pixel_copy_increment, \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
513 double_pixel_copy_increment, \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
514 quatro_pixel_copy_increment, width) \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
515 { int n = width; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
516 if(n & 1) { \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
517 pixel_copy_increment; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
518 n--; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
519 } \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
520 if(n & 2) { \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
521 double_pixel_copy_increment; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
522 n -= 2; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
523 } \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
524 n=n>>2; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
525 for(; n > 0; --n) { \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
526 quatro_pixel_copy_increment; \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
527 } \ |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
528 } |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
529 |
5bb080d35049
Date: Tue, 19 Aug 2003 17:57:00 +0200
Sam Lantinga <slouken@libsdl.org>
parents:
553
diff
changeset
|
530 /* Don't use Duff's device to unroll loops */ |
0 | 531 #define DUFFS_LOOP(pixel_copy_increment, width) \ |
532 { int n; \ | |
533 for ( n=width; n > 0; --n ) { \ | |
534 pixel_copy_increment; \ | |
535 } \ | |
536 } | |
537 #define DUFFS_LOOP8(pixel_copy_increment, width) \ | |
538 DUFFS_LOOP(pixel_copy_increment, width) | |
539 #define DUFFS_LOOP4(pixel_copy_increment, width) \ | |
540 DUFFS_LOOP(pixel_copy_increment, width) | |
541 | |
542 #endif /* USE_DUFFS_LOOP */ | |
543 | |
544 /* Prevent Visual C++ 6.0 from printing out stupid warnings */ | |
545 #if defined(_MSC_VER) && (_MSC_VER >= 600) | |
546 #pragma warning(disable: 4550) | |
547 #endif | |
548 | |
549 #endif /* _SDL_blit_h */ | |
1895
c121d94672cb
SDL 1.2 is moving to a branch, and SDL 1.3 is becoming the head.
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
550 /* vi: set ts=4 sw=4 expandtab: */ |