annotate libvo/aclib.c @ 6453:7f06ef0d58a6

8bpp bugfixes (brightness, contrast ...) 4bpp
author michael
date Mon, 17 Jun 2002 01:48:31 +0000
parents c75f75806af1
children 113d66d78967
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
698
f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff changeset
1 #include "../config.h"
3393
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
2 #ifdef USE_FASTMEMCPY
698
f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff changeset
3
3077
99f6db3255aa 10-20% faster fastmemcpy :) on my p3 at least but the algo is mostly from "amd athlon processor x86 code optimization guide" so it should be faster for amd chips too, but i fear it might be slower for mem->vram copies (someone should check that, i cant) ... there are 2 #defines to finetune it (BLOCK_SIZE & CONFUSION_FACTOR)
michael
parents: 1123
diff changeset
4 /*
698
f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff changeset
5 aclib - advanced C library ;)
f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff changeset
6 This file contains functions which improve and expand standard C-library
3393
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
7 see aclib_template.c ... this file only contains runtime cpu detection and config options stuff
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
8 runtime cpu detection by michael niedermayer (michaelni@gmx.at) is under GPL
698
f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff changeset
9 */
f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff changeset
10 #include <stddef.h>
3393
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
11 #include "../cpudetect.h"
698
f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff changeset
12
3077
99f6db3255aa 10-20% faster fastmemcpy :) on my p3 at least but the algo is mostly from "amd athlon processor x86 code optimization guide" so it should be faster for amd chips too, but i fear it might be slower for mem->vram copies (someone should check that, i cant) ... there are 2 #defines to finetune it (BLOCK_SIZE & CONFUSION_FACTOR)
michael
parents: 1123
diff changeset
13 #define BLOCK_SIZE 4096
99f6db3255aa 10-20% faster fastmemcpy :) on my p3 at least but the algo is mostly from "amd athlon processor x86 code optimization guide" so it should be faster for amd chips too, but i fear it might be slower for mem->vram copies (someone should check that, i cant) ... there are 2 #defines to finetune it (BLOCK_SIZE & CONFUSION_FACTOR)
michael
parents: 1123
diff changeset
14 #define CONFUSION_FACTOR 0
99f6db3255aa 10-20% faster fastmemcpy :) on my p3 at least but the algo is mostly from "amd athlon processor x86 code optimization guide" so it should be faster for amd chips too, but i fear it might be slower for mem->vram copies (someone should check that, i cant) ... there are 2 #defines to finetune it (BLOCK_SIZE & CONFUSION_FACTOR)
michael
parents: 1123
diff changeset
15 //Feel free to fine-tune the above 2, it might be possible to get some speedup with them :)
99f6db3255aa 10-20% faster fastmemcpy :) on my p3 at least but the algo is mostly from "amd athlon processor x86 code optimization guide" so it should be faster for amd chips too, but i fear it might be slower for mem->vram copies (someone should check that, i cant) ... there are 2 #defines to finetune it (BLOCK_SIZE & CONFUSION_FACTOR)
michael
parents: 1123
diff changeset
16
99f6db3255aa 10-20% faster fastmemcpy :) on my p3 at least but the algo is mostly from "amd athlon processor x86 code optimization guide" so it should be faster for amd chips too, but i fear it might be slower for mem->vram copies (someone should check that, i cant) ... there are 2 #defines to finetune it (BLOCK_SIZE & CONFUSION_FACTOR)
michael
parents: 1123
diff changeset
17 //#define STATISTICS
3393
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
18 #ifdef ARCH_X86
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
19 #define CAN_COMPILE_X86_ASM
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
20 #endif
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
21
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
22 //Note: we have MMX, MMX2, 3DNOW version there is no 3DNOW+MMX2 one
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
23 //Plain C versions
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
24 //#if !defined (HAVE_MMX) || defined (RUNTIME_CPUDETECT)
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
25 //#define COMPILE_C
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
26 //#endif
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
27
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
28 #ifdef CAN_COMPILE_X86_ASM
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
29
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
30 #if (defined (HAVE_MMX) && !defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
31 #define COMPILE_MMX
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
32 #endif
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
33
5208
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
34 #if (defined (HAVE_MMX2) && !defined (HAVE_SSE2)) || defined (RUNTIME_CPUDETECT)
3393
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
35 #define COMPILE_MMX2
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
36 #endif
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
37
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
38 #if (defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
39 #define COMPILE_3DNOW
698
f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff changeset
40 #endif
f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff changeset
41
5208
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
42 #if defined (HAVE_SSE2) || defined (RUNTIME_CPUDETECT)
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
43 #define COMPILE_SSE
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
44 #endif
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
45
3393
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
46 #undef HAVE_MMX
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
47 #undef HAVE_MMX2
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
48 #undef HAVE_3DNOW
5208
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
49 #undef HAVE_SSE
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
50 #undef HAVE_SSE2
3393
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
51 #undef ARCH_X86
698
f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff changeset
52 /*
3393
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
53 #ifdef COMPILE_C
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
54 #undef HAVE_MMX
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
55 #undef HAVE_MMX2
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
56 #undef HAVE_3DNOW
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
57 #undef ARCH_X86
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
58 #define RENAME(a) a ## _C
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
59 #include "aclib_template.c"
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
60 #endif
698
f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff changeset
61 */
3393
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
62 //MMX versions
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
63 #ifdef COMPILE_MMX
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
64 #undef RENAME
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
65 #define HAVE_MMX
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
66 #undef HAVE_MMX2
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
67 #undef HAVE_3DNOW
5208
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
68 #undef HAVE_SSE
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
69 #undef HAVE_SSE2
3393
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
70 #define ARCH_X86
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
71 #define RENAME(a) a ## _MMX
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
72 #include "aclib_template.c"
698
f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff changeset
73 #endif
f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff changeset
74
3393
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
75 //MMX2 versions
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
76 #ifdef COMPILE_MMX2
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
77 #undef RENAME
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
78 #define HAVE_MMX
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
79 #define HAVE_MMX2
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
80 #undef HAVE_3DNOW
5208
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
81 #undef HAVE_SSE
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
82 #undef HAVE_SSE2
3393
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
83 #define ARCH_X86
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
84 #define RENAME(a) a ## _MMX2
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
85 #include "aclib_template.c"
698
f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff changeset
86 #endif
3077
99f6db3255aa 10-20% faster fastmemcpy :) on my p3 at least but the algo is mostly from "amd athlon processor x86 code optimization guide" so it should be faster for amd chips too, but i fear it might be slower for mem->vram copies (someone should check that, i cant) ... there are 2 #defines to finetune it (BLOCK_SIZE & CONFUSION_FACTOR)
michael
parents: 1123
diff changeset
87
3393
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
88 //3DNOW versions
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
89 #ifdef COMPILE_3DNOW
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
90 #undef RENAME
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
91 #define HAVE_MMX
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
92 #undef HAVE_MMX2
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
93 #define HAVE_3DNOW
5208
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
94 #undef HAVE_SSE
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
95 #undef HAVE_SSE2
3393
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
96 #define ARCH_X86
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
97 #define RENAME(a) a ## _3DNow
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
98 #include "aclib_template.c"
3077
99f6db3255aa 10-20% faster fastmemcpy :) on my p3 at least but the algo is mostly from "amd athlon processor x86 code optimization guide" so it should be faster for amd chips too, but i fear it might be slower for mem->vram copies (someone should check that, i cant) ... there are 2 #defines to finetune it (BLOCK_SIZE & CONFUSION_FACTOR)
michael
parents: 1123
diff changeset
99 #endif
99f6db3255aa 10-20% faster fastmemcpy :) on my p3 at least but the algo is mostly from "amd athlon processor x86 code optimization guide" so it should be faster for amd chips too, but i fear it might be slower for mem->vram copies (someone should check that, i cant) ... there are 2 #defines to finetune it (BLOCK_SIZE & CONFUSION_FACTOR)
michael
parents: 1123
diff changeset
100
5208
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
101 //SSE versions (only used on SSE2 cpus)
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
102 #ifdef COMPILE_SSE
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
103 #undef RENAME
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
104 #define HAVE_MMX
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
105 #define HAVE_MMX2
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
106 #undef HAVE_3DNOW
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
107 #define HAVE_SSE
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
108 #define HAVE_SSE2
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
109 #define ARCH_X86
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
110 #define RENAME(a) a ## _SSE
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
111 #include "aclib_template.c"
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
112 #endif
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
113
3393
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
114 #endif // CAN_COMPILE_X86_ASM
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
115
3077
99f6db3255aa 10-20% faster fastmemcpy :) on my p3 at least but the algo is mostly from "amd athlon processor x86 code optimization guide" so it should be faster for amd chips too, but i fear it might be slower for mem->vram copies (someone should check that, i cant) ... there are 2 #defines to finetune it (BLOCK_SIZE & CONFUSION_FACTOR)
michael
parents: 1123
diff changeset
116
3393
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
117 inline void * fast_memcpy(void * to, const void * from, size_t len)
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
118 {
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
119 #ifdef RUNTIME_CPUDETECT
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
120 #ifdef CAN_COMPILE_X86_ASM
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
121 // ordered per speed fasterst first
5208
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
122 if(gCpuCaps.hasSSE2)
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
123 fast_memcpy_SSE(to, from, len);
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
124 else if(gCpuCaps.hasMMX2)
3393
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
125 fast_memcpy_MMX2(to, from, len);
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
126 else if(gCpuCaps.has3DNow)
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
127 fast_memcpy_3DNow(to, from, len);
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
128 else if(gCpuCaps.hasMMX)
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
129 fast_memcpy_MMX(to, from, len);
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
130 else
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
131 #endif //CAN_COMPILE_X86_ASM
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
132 memcpy(to, from, len); // prior to mmx we use the standart memcpy
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
133 #else
5208
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
134 #ifdef HAVE_SSE2
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
135 fast_memcpy_SSE(to, from, len);
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
136 #elif defined (HAVE_MMX2)
3393
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
137 fast_memcpy_MMX2(to, from, len);
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
138 #elif defined (HAVE_3DNOW)
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
139 fast_memcpy_3DNow(to, from, len);
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
140 #elif defined (HAVE_MMX)
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
141 fast_memcpy_MMX(to, from, len);
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
142 #else
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
143 memcpy(to, from, len); // prior to mmx we use the standart memcpy
698
f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff changeset
144 #endif
3077
99f6db3255aa 10-20% faster fastmemcpy :) on my p3 at least but the algo is mostly from "amd athlon processor x86 code optimization guide" so it should be faster for amd chips too, but i fear it might be slower for mem->vram copies (someone should check that, i cant) ... there are 2 #defines to finetune it (BLOCK_SIZE & CONFUSION_FACTOR)
michael
parents: 1123
diff changeset
145
3393
3624cd351618 runtime cpu detection
michael
parents: 3077
diff changeset
146 #endif //!RUNTIME_CPUDETECT
5543
c75f75806af1 memcpy must return destination ptr patch by Adam <adam@cfar.umd.edu>
michael
parents: 5208
diff changeset
147 return to;
698
f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff changeset
148 }
f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff changeset
149
4681
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
150 inline void * mem2agpcpy(void * to, const void * from, size_t len)
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
151 {
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
152 #ifdef RUNTIME_CPUDETECT
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
153 #ifdef CAN_COMPILE_X86_ASM
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
154 // ordered per speed fasterst first
5208
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
155 if(gCpuCaps.hasSSE2)
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
156 mem2agpcpy_SSE(to, from, len);
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
157 else if(gCpuCaps.hasMMX2)
4681
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
158 mem2agpcpy_MMX2(to, from, len);
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
159 else if(gCpuCaps.has3DNow)
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
160 mem2agpcpy_3DNow(to, from, len);
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
161 else if(gCpuCaps.hasMMX)
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
162 mem2agpcpy_MMX(to, from, len);
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
163 else
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
164 #endif //CAN_COMPILE_X86_ASM
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
165 memcpy(to, from, len); // prior to mmx we use the standart memcpy
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
166 #else
5208
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
167 #ifdef HAVE_SSE2
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
168 mem2agpcpy_SSE(to, from, len);
b08228af4098 fixing runtime cpudetect with pre SSE cpus
michael
parents: 4681
diff changeset
169 #elif defined (HAVE_MMX2)
4681
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
170 mem2agpcpy_MMX2(to, from, len);
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
171 #elif defined (HAVE_3DNOW)
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
172 mem2agpcpy_3DNow(to, from, len);
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
173 #elif defined (HAVE_MMX)
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
174 mem2agpcpy_MMX(to, from, len);
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
175 #else
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
176 memcpy(to, from, len); // prior to mmx we use the standart memcpy
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
177 #endif
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
178
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
179 #endif //!RUNTIME_CPUDETECT
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
180 }
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
181
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
182 #endif /* use fastmemcpy */
8db59073127e mem2agpcpy()
michael
parents: 3393
diff changeset
183