Mercurial > mplayer.hg
annotate libvo/aclib.c @ 25555:273c1eaae7c8
Happy New Year!
| author | zuxy |
|---|---|
| date | Wed, 02 Jan 2008 07:38:40 +0000 |
| parents | 273aa6124f66 |
| children | e7c989f7a7c9 |
| rev | line source |
|---|---|
|
12650
ac3fd2ff2561
Unify the config.h #include, use "config.h" instead of "../config.h"
diego
parents:
12492
diff
changeset
|
1 #include "config.h" |
| 3393 | 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 | 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 | |
|
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> |
|
21982
fa66a03e8920
Include string.h to make sure memcpy is not used without prototype
reimar
parents:
20577
diff
changeset
|
11 #include <string.h> |
|
13787
e047e70a9767
Handle "xxx.h" vs "../xxx.h" include paths in a consistent way.
diego
parents:
13720
diff
changeset
|
12 #include "cpudetect.h" |
|
8123
9fc45fe0d444
*HUGE* set of compiler warning fixes, unused variables removal
arpi
parents:
7072
diff
changeset
|
13 #include "fastmemcpy.h" |
|
12492
4b8417674f1c
fix crash due to fast_memcpy calling itself instead of libc memcpy
reimar
parents:
8127
diff
changeset
|
14 #undef memcpy |
|
698
f0fbf1a9bf31
Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff
changeset
|
15 |
|
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
|
16 #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
|
17 #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
|
18 //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
|
19 |
|
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
|
20 //#define STATISTICS |
| 20577 | 21 #ifdef ARCH_X86 |
| 3393 | 22 #define CAN_COMPILE_X86_ASM |
| 23 #endif | |
| 24 | |
| 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 | |
| 30 | |
| 31 #ifdef CAN_COMPILE_X86_ASM | |
| 32 | |
| 33 #if (defined (HAVE_MMX) && !defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT) | |
| 34 #define COMPILE_MMX | |
| 35 #endif | |
| 36 | |
| 5208 | 37 #if (defined (HAVE_MMX2) && !defined (HAVE_SSE2)) || defined (RUNTIME_CPUDETECT) |
| 3393 | 38 #define COMPILE_MMX2 |
| 39 #endif | |
| 40 | |
| 41 #if (defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT) | |
| 42 #define COMPILE_3DNOW | |
|
698
f0fbf1a9bf31
Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff
changeset
|
43 #endif |
|
f0fbf1a9bf31
Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff
changeset
|
44 |
| 5208 | 45 #if defined (HAVE_SSE2) || defined (RUNTIME_CPUDETECT) |
| 46 #define COMPILE_SSE | |
| 47 #endif | |
| 48 | |
| 3393 | 49 #undef HAVE_MMX |
| 50 #undef HAVE_MMX2 | |
| 51 #undef HAVE_3DNOW | |
| 5208 | 52 #undef HAVE_SSE |
| 53 #undef HAVE_SSE2 | |
|
698
f0fbf1a9bf31
Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff
changeset
|
54 /* |
| 3393 | 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 | |
|
698
f0fbf1a9bf31
Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff
changeset
|
63 */ |
| 3393 | 64 //MMX versions |
| 65 #ifdef COMPILE_MMX | |
| 66 #undef RENAME | |
| 67 #define HAVE_MMX | |
| 68 #undef HAVE_MMX2 | |
| 69 #undef HAVE_3DNOW | |
| 5208 | 70 #undef HAVE_SSE |
| 71 #undef HAVE_SSE2 | |
| 3393 | 72 #define RENAME(a) a ## _MMX |
| 73 #include "aclib_template.c" | |
|
698
f0fbf1a9bf31
Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff
changeset
|
74 #endif |
|
f0fbf1a9bf31
Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff
changeset
|
75 |
| 3393 | 76 //MMX2 versions |
| 77 #ifdef COMPILE_MMX2 | |
| 78 #undef RENAME | |
| 79 #define HAVE_MMX | |
| 80 #define HAVE_MMX2 | |
| 81 #undef HAVE_3DNOW | |
| 5208 | 82 #undef HAVE_SSE |
| 83 #undef HAVE_SSE2 | |
| 3393 | 84 #define RENAME(a) a ## _MMX2 |
| 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 | 88 //3DNOW versions |
| 89 #ifdef COMPILE_3DNOW | |
| 90 #undef RENAME | |
| 91 #define HAVE_MMX | |
| 92 #undef HAVE_MMX2 | |
| 93 #define HAVE_3DNOW | |
| 5208 | 94 #undef HAVE_SSE |
| 95 #undef HAVE_SSE2 | |
| 3393 | 96 #define RENAME(a) a ## _3DNow |
| 97 #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
|
98 #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
|
99 |
| 5208 | 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 | |
| 111 | |
| 3393 | 112 #endif // CAN_COMPILE_X86_ASM |
| 113 | |
|
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
|
114 |
|
23523
273aa6124f66
avoid utter breakage on non-x86, patch from Chris Roccati <roccati@at@pobox.dot.com>
lu_zero
parents:
21982
diff
changeset
|
115 #undef fast_memcpy |
| 7072 | 116 void * fast_memcpy(void * to, const void * from, size_t len) |
| 3393 | 117 { |
| 118 #ifdef RUNTIME_CPUDETECT | |
| 119 #ifdef CAN_COMPILE_X86_ASM | |
| 120 // ordered per speed fasterst first | |
| 5208 | 121 if(gCpuCaps.hasSSE2) |
| 122 fast_memcpy_SSE(to, from, len); | |
| 123 else if(gCpuCaps.hasMMX2) | |
| 3393 | 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 | |
| 5208 | 133 #ifdef HAVE_SSE2 |
| 134 fast_memcpy_SSE(to, from, len); | |
| 135 #elif defined (HAVE_MMX2) | |
| 3393 | 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 | |
|
698
f0fbf1a9bf31
Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff
changeset
|
143 #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
|
144 |
| 3393 | 145 #endif //!RUNTIME_CPUDETECT |
|
5543
c75f75806af1
memcpy must return destination ptr patch by Adam <adam@cfar.umd.edu>
michael
parents:
5208
diff
changeset
|
146 return to; |
|
698
f0fbf1a9bf31
Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff
changeset
|
147 } |
|
f0fbf1a9bf31
Moving fast_memcpy to separate file (Size optimization)
nickols_k
parents:
diff
changeset
|
148 |
|
8127
e7153e62a7f4
On non-x86 platforms, memcpy was re-implemented in mplayer and was called
jkeil
parents:
8123
diff
changeset
|
149 #undef mem2agpcpy |
| 7072 | 150 void * mem2agpcpy(void * to, const void * from, size_t len) |
| 4681 | 151 { |
| 152 #ifdef RUNTIME_CPUDETECT | |
| 153 #ifdef CAN_COMPILE_X86_ASM | |
| 154 // ordered per speed fasterst first | |
| 5208 | 155 if(gCpuCaps.hasSSE2) |
| 156 mem2agpcpy_SSE(to, from, len); | |
| 157 else if(gCpuCaps.hasMMX2) | |
| 4681 | 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 | |
| 5208 | 167 #ifdef HAVE_SSE2 |
| 168 mem2agpcpy_SSE(to, from, len); | |
| 169 #elif defined (HAVE_MMX2) | |
| 4681 | 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 | |
| 178 | |
| 179 #endif //!RUNTIME_CPUDETECT | |
|
8123
9fc45fe0d444
*HUGE* set of compiler warning fixes, unused variables removal
arpi
parents:
7072
diff
changeset
|
180 return to; |
| 4681 | 181 } |
| 182 | |
| 183 #endif /* use fastmemcpy */ | |
| 184 |