Merge svn changes up to r27441
[mplayer.git] / libvo / aclib.c
blob68548f579e820ed09c2f012693881ef880f690f0
1 #include "config.h"
2 #ifdef CONFIG_FASTMEMCPY
4 /*
5 aclib - advanced C library ;)
6 This file contains functions which improve and expand standard C-library
7 see aclib_template.c ... this file only contains runtime cpu detection and config options stuff
8 runtime cpu detection by michael niedermayer (michaelni@gmx.at) is under GPL
9 */
10 #include <stddef.h>
11 #include <string.h>
12 #include "cpudetect.h"
13 #include "fastmemcpy.h"
14 #undef memcpy
16 #define BLOCK_SIZE 4096
17 #define CONFUSION_FACTOR 0
18 //Feel free to fine-tune the above 2, it might be possible to get some speedup with them :)
20 //#define STATISTICS
21 #ifdef ARCH_X86
22 #define CAN_COMPILE_X86_ASM
23 #endif
25 //Note: we have MMX, MMX2, 3DNOW version there is no 3DNOW+MMX2 one
26 //Plain C versions
27 //#if !defined (HAVE_MMX) || defined (RUNTIME_CPUDETECT)
28 //#define COMPILE_C
29 //#endif
31 #ifdef CAN_COMPILE_X86_ASM
33 #if (defined (HAVE_MMX) && !defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)
34 #define COMPILE_MMX
35 #endif
37 #if (defined (HAVE_MMX2) && !defined (HAVE_SSE2)) || defined (RUNTIME_CPUDETECT)
38 #define COMPILE_MMX2
39 #endif
41 #if (defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)
42 #define COMPILE_3DNOW
43 #endif
45 #if defined (HAVE_SSE2) || defined (RUNTIME_CPUDETECT)
46 #define COMPILE_SSE
47 #endif
49 #undef HAVE_MMX
50 #undef HAVE_MMX2
51 #undef HAVE_3DNOW
52 #undef HAVE_SSE
53 #undef HAVE_SSE2
55 #ifdef COMPILE_C
56 #undef HAVE_MMX
57 #undef HAVE_MMX2
58 #undef HAVE_3DNOW
59 #undef ARCH_X86
60 #define RENAME(a) a ## _C
61 #include "aclib_template.c"
62 #endif
64 //MMX versions
65 #ifdef COMPILE_MMX
66 #undef RENAME
67 #define HAVE_MMX
68 #undef HAVE_MMX2
69 #undef HAVE_3DNOW
70 #undef HAVE_SSE
71 #undef HAVE_SSE2
72 #define RENAME(a) a ## _MMX
73 #include "aclib_template.c"
74 #endif
76 //MMX2 versions
77 #ifdef COMPILE_MMX2
78 #undef RENAME
79 #define HAVE_MMX
80 #define HAVE_MMX2
81 #undef HAVE_3DNOW
82 #undef HAVE_SSE
83 #undef HAVE_SSE2
84 #define RENAME(a) a ## _MMX2
85 #include "aclib_template.c"
86 #endif
88 //3DNOW versions
89 #ifdef COMPILE_3DNOW
90 #undef RENAME
91 #define HAVE_MMX
92 #undef HAVE_MMX2
93 #define HAVE_3DNOW
94 #undef HAVE_SSE
95 #undef HAVE_SSE2
96 #define RENAME(a) a ## _3DNow
97 #include "aclib_template.c"
98 #endif
100 //SSE versions (only used on SSE2 cpus)
101 #ifdef COMPILE_SSE
102 #undef RENAME
103 #define HAVE_MMX
104 #define HAVE_MMX2
105 #undef HAVE_3DNOW
106 #define HAVE_SSE
107 #define HAVE_SSE2
108 #define RENAME(a) a ## _SSE
109 #include "aclib_template.c"
110 #endif
112 #endif // CAN_COMPILE_X86_ASM
115 #undef fast_memcpy
116 void * fast_memcpy(void * to, const void * from, size_t len)
118 #ifdef RUNTIME_CPUDETECT
119 #ifdef CAN_COMPILE_X86_ASM
120 // ordered per speed fasterst first
121 if(gCpuCaps.hasSSE2)
122 fast_memcpy_SSE(to, from, len);
123 else if(gCpuCaps.hasMMX2)
124 fast_memcpy_MMX2(to, from, len);
125 else if(gCpuCaps.has3DNow)
126 fast_memcpy_3DNow(to, from, len);
127 else if(gCpuCaps.hasMMX)
128 fast_memcpy_MMX(to, from, len);
129 else
130 #endif //CAN_COMPILE_X86_ASM
131 memcpy(to, from, len); // prior to mmx we use the standart memcpy
132 #else
133 #ifdef HAVE_SSE2
134 fast_memcpy_SSE(to, from, len);
135 #elif defined (HAVE_MMX2)
136 fast_memcpy_MMX2(to, from, len);
137 #elif defined (HAVE_3DNOW)
138 fast_memcpy_3DNow(to, from, len);
139 #elif defined (HAVE_MMX)
140 fast_memcpy_MMX(to, from, len);
141 #else
142 memcpy(to, from, len); // prior to mmx we use the standart memcpy
143 #endif
145 #endif //!RUNTIME_CPUDETECT
146 return to;
149 #undef mem2agpcpy
150 void * mem2agpcpy(void * to, const void * from, size_t len)
152 #ifdef RUNTIME_CPUDETECT
153 #ifdef CAN_COMPILE_X86_ASM
154 // ordered per speed fasterst first
155 if(gCpuCaps.hasSSE2)
156 mem2agpcpy_SSE(to, from, len);
157 else if(gCpuCaps.hasMMX2)
158 mem2agpcpy_MMX2(to, from, len);
159 else if(gCpuCaps.has3DNow)
160 mem2agpcpy_3DNow(to, from, len);
161 else if(gCpuCaps.hasMMX)
162 mem2agpcpy_MMX(to, from, len);
163 else
164 #endif //CAN_COMPILE_X86_ASM
165 memcpy(to, from, len); // prior to mmx we use the standart memcpy
166 #else
167 #ifdef HAVE_SSE2
168 mem2agpcpy_SSE(to, from, len);
169 #elif defined (HAVE_MMX2)
170 mem2agpcpy_MMX2(to, from, len);
171 #elif defined (HAVE_3DNOW)
172 mem2agpcpy_3DNow(to, from, len);
173 #elif defined (HAVE_MMX)
174 mem2agpcpy_MMX(to, from, len);
175 #else
176 memcpy(to, from, len); // prior to mmx we use the standart memcpy
177 #endif
179 #endif //!RUNTIME_CPUDETECT
180 return to;
183 #endif /* CONFIG_FASTMEMCPY */