mplayer.hg: libvo/aclib.c annotate

annotate libvo/aclib.c @ 28335:31287e75b5d8

HAVE_3DNOW --> HAVE_AMD3DNOW

author	diego
date	Mon, 26 Jan 2009 09:29:53 +0000
parents	af0b0ae25b84
children	7681eab10aea

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"
27341 e7c989f7a7c9 Start unifying names of internal preprocessor directives. diego parents: 23523 diff changeset	2 #ifdef CONFIG_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>
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
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	21 #if ARCH_X86
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	22 #define CAN_COMPILE_X86_ASM
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	23 #endif
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	24
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	25 //Note: we have MMX, MMX2, 3DNOW version there is no 3DNOW+MMX2 one
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	26 //Plain C versions
28292 d6001126678f More #ifdef HAVE_MMX etc. missed by earlier search. reimar parents: 28290 diff changeset	27 //#if !HAVE_MMX \|\| defined (RUNTIME_CPUDETECT)
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	28 //#define COMPILE_C
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	29 //#endif
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	30
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	31 #ifdef CAN_COMPILE_X86_ASM
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	32
28335 31287e75b5d8 HAVE_3DNOW --> HAVE_AMD3DNOW diego parents: 28302 diff changeset	33 #if (HAVE_MMX && !HAVE_AMD3DNOW && !HAVE_MMX2) \|\| defined (RUNTIME_CPUDETECT)
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	34 #define COMPILE_MMX
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	35 #endif
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	36
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	37 #if (HAVE_MMX2 && !HAVE_SSE2) \|\| defined (RUNTIME_CPUDETECT)
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	38 #define COMPILE_MMX2
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	39 #endif
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	40
28335 31287e75b5d8 HAVE_3DNOW --> HAVE_AMD3DNOW diego parents: 28302 diff changeset	41 #if (HAVE_AMD3DNOW && !HAVE_MMX2) \|\| defined (RUNTIME_CPUDETECT)
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	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
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	45 #if HAVE_SSE2 \|\| defined (RUNTIME_CPUDETECT)
5208 b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	46 #define COMPILE_SSE
b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	47 #endif
b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	48
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	49 #undef HAVE_MMX
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	50 #undef HAVE_MMX2
28335 31287e75b5d8 HAVE_3DNOW --> HAVE_AMD3DNOW diego parents: 28302 diff changeset	51 #undef HAVE_AMD3DNOW
5208 b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	52 #undef HAVE_SSE
b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	53 #undef HAVE_SSE2
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	54 #define HAVE_MMX 0
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	55 #define HAVE_MMX2 0
28335 31287e75b5d8 HAVE_3DNOW --> HAVE_AMD3DNOW diego parents: 28302 diff changeset	56 #define HAVE_AMD3DNOW 0
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	57 #define HAVE_SSE 0
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	58 #define HAVE_SSE2 0
698 f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization) nickols_k parents: diff changeset	59 /*
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	60 #ifdef COMPILE_C
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	61 #undef HAVE_MMX
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	62 #undef HAVE_MMX2
28335 31287e75b5d8 HAVE_3DNOW --> HAVE_AMD3DNOW diego parents: 28302 diff changeset	63 #undef HAVE_AMD3DNOW
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	64 #undef HAVE_SSE
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	65 #undef HAVE_SSE2
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	66 #define HAVE_MMX 0
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	67 #define HAVE_MMX2 0
28335 31287e75b5d8 HAVE_3DNOW --> HAVE_AMD3DNOW diego parents: 28302 diff changeset	68 #define HAVE_AMD3DNOW 0
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	69 #define HAVE_SSE 0
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	70 #define HAVE_SSE2 0
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	71 #define RENAME(a) a ## _C
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	72 #include "aclib_template.c"
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	73 #endif
698 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 //MMX versions
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	76 #ifdef COMPILE_MMX
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	77 #undef RENAME
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	78 #undef HAVE_MMX
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	79 #undef HAVE_MMX2
28335 31287e75b5d8 HAVE_3DNOW --> HAVE_AMD3DNOW diego parents: 28302 diff changeset	80 #undef HAVE_AMD3DNOW
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
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	83 #define HAVE_MMX 1
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	84 #define HAVE_MMX2 0
28335 31287e75b5d8 HAVE_3DNOW --> HAVE_AMD3DNOW diego parents: 28302 diff changeset	85 #define HAVE_AMD3DNOW 0
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	86 #define HAVE_SSE 0
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	87 #define HAVE_SSE2 0
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	88 #define RENAME(a) a ## _MMX
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	89 #include "aclib_template.c"
698 f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization) nickols_k parents: diff changeset	90 #endif
f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization) nickols_k parents: diff changeset	91
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	92 //MMX2 versions
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	93 #ifdef COMPILE_MMX2
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	94 #undef RENAME
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	95 #undef HAVE_MMX
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	96 #undef HAVE_MMX2
28335 31287e75b5d8 HAVE_3DNOW --> HAVE_AMD3DNOW diego parents: 28302 diff changeset	97 #undef HAVE_AMD3DNOW
5208 b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	98 #undef HAVE_SSE
b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	99 #undef HAVE_SSE2
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	100 #define HAVE_MMX 1
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	101 #define HAVE_MMX2 1
28335 31287e75b5d8 HAVE_3DNOW --> HAVE_AMD3DNOW diego parents: 28302 diff changeset	102 #define HAVE_AMD3DNOW 0
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	103 #define HAVE_SSE 0
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	104 #define HAVE_SSE2 0
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	105 #define RENAME(a) a ## _MMX2
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	106 #include "aclib_template.c"
698 f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization) nickols_k parents: diff changeset	107 #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	108
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	109 //3DNOW versions
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	110 #ifdef COMPILE_3DNOW
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	111 #undef RENAME
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	112 #undef HAVE_MMX
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	113 #undef HAVE_MMX2
28335 31287e75b5d8 HAVE_3DNOW --> HAVE_AMD3DNOW diego parents: 28302 diff changeset	114 #undef HAVE_AMD3DNOW
5208 b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	115 #undef HAVE_SSE
b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	116 #undef HAVE_SSE2
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	117 #define HAVE_MMX 1
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	118 #define HAVE_MMX2 0
28335 31287e75b5d8 HAVE_3DNOW --> HAVE_AMD3DNOW diego parents: 28302 diff changeset	119 #define HAVE_AMD3DNOW 1
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	120 #define HAVE_SSE 0
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	121 #define HAVE_SSE2 0
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	122 #define RENAME(a) a ## _3DNow
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	123 #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	124 #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	125
5208 b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	126 //SSE versions (only used on SSE2 cpus)
b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	127 #ifdef COMPILE_SSE
b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	128 #undef RENAME
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	129 #undef HAVE_MMX
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	130 #undef HAVE_MMX2
28335 31287e75b5d8 HAVE_3DNOW --> HAVE_AMD3DNOW diego parents: 28302 diff changeset	131 #undef HAVE_AMD3DNOW
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	132 #undef HAVE_SSE
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	133 #undef HAVE_SSE2
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	134 #define HAVE_MMX 1
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	135 #define HAVE_MMX2 1
28335 31287e75b5d8 HAVE_3DNOW --> HAVE_AMD3DNOW diego parents: 28302 diff changeset	136 #define HAVE_AMD3DNOW 0
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	137 #define HAVE_SSE 1
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	138 #define HAVE_SSE2 1
5208 b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	139 #define RENAME(a) a ## _SSE
b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	140 #include "aclib_template.c"
b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	141 #endif
b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	142
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	143 #endif // CAN_COMPILE_X86_ASM
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	144
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
23523 273aa6124f66 avoid utter breakage on non-x86, patch from Chris Roccati <roccati@at@pobox.dot.com> lu_zero parents: 21982 diff changeset	146 #undef fast_memcpy
7072 113d66d78967 removed nonsense 'inline' arpi parents: 5543 diff changeset	147 void * fast_memcpy(void * to, const void * from, size_t len)
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	148 {
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	149 #ifdef RUNTIME_CPUDETECT
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	150 #ifdef CAN_COMPILE_X86_ASM
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	151 // ordered per speed fasterst first
5208 b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	152 if(gCpuCaps.hasSSE2)
b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	153 fast_memcpy_SSE(to, from, len);
b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	154 else if(gCpuCaps.hasMMX2)
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	155 fast_memcpy_MMX2(to, from, len);
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	156 else if(gCpuCaps.has3DNow)
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	157 fast_memcpy_3DNow(to, from, len);
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	158 else if(gCpuCaps.hasMMX)
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	159 fast_memcpy_MMX(to, from, len);
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	160 else
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	161 #endif //CAN_COMPILE_X86_ASM
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	162 memcpy(to, from, len); // prior to mmx we use the standart memcpy
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	163 #else
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	164 #if HAVE_SSE2
5208 b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	165 fast_memcpy_SSE(to, from, len);
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	166 #elif HAVE_MMX2
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	167 fast_memcpy_MMX2(to, from, len);
28335 31287e75b5d8 HAVE_3DNOW --> HAVE_AMD3DNOW diego parents: 28302 diff changeset	168 #elif HAVE_AMD3DNOW
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	169 fast_memcpy_3DNow(to, from, len);
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	170 #elif HAVE_MMX
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	171 fast_memcpy_MMX(to, from, len);
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	172 #else
3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	173 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	174 #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	175
3393 3624cd351618 runtime cpu detection michael parents: 3077 diff changeset	176 #endif //!RUNTIME_CPUDETECT
5543 c75f75806af1 memcpy must return destination ptr patch by Adam <adam@cfar.umd.edu> michael parents: 5208 diff changeset	177 return to;
698 f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization) nickols_k parents: diff changeset	178 }
f0fbf1a9bf31 Moving fast_memcpy to separate file (Size optimization) nickols_k parents: diff changeset	179
8127 e7153e62a7f4 On non-x86 platforms, memcpy was re-implemented in mplayer and was called jkeil parents: 8123 diff changeset	180 #undef mem2agpcpy
7072 113d66d78967 removed nonsense 'inline' arpi parents: 5543 diff changeset	181 void * mem2agpcpy(void * to, const void * from, size_t len)
4681 8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	182 {
8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	183 #ifdef RUNTIME_CPUDETECT
8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	184 #ifdef CAN_COMPILE_X86_ASM
8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	185 // ordered per speed fasterst first
5208 b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	186 if(gCpuCaps.hasSSE2)
b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	187 mem2agpcpy_SSE(to, from, len);
b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	188 else if(gCpuCaps.hasMMX2)
4681 8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	189 mem2agpcpy_MMX2(to, from, len);
8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	190 else if(gCpuCaps.has3DNow)
8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	191 mem2agpcpy_3DNow(to, from, len);
8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	192 else if(gCpuCaps.hasMMX)
8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	193 mem2agpcpy_MMX(to, from, len);
8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	194 else
8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	195 #endif //CAN_COMPILE_X86_ASM
8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	196 memcpy(to, from, len); // prior to mmx we use the standart memcpy
8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	197 #else
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	198 #if HAVE_SSE2
5208 b08228af4098 fixing runtime cpudetect with pre SSE cpus michael parents: 4681 diff changeset	199 mem2agpcpy_SSE(to, from, len);
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	200 #elif HAVE_MMX2
4681 8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	201 mem2agpcpy_MMX2(to, from, len);
28335 31287e75b5d8 HAVE_3DNOW --> HAVE_AMD3DNOW diego parents: 28302 diff changeset	202 #elif HAVE_AMD3DNOW
4681 8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	203 mem2agpcpy_3DNow(to, from, len);
28290 25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_... reimar parents: 27341 diff changeset	204 #elif HAVE_MMX
4681 8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	205 mem2agpcpy_MMX(to, from, len);
8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	206 #else
8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	207 memcpy(to, from, len); // prior to mmx we use the standart memcpy
8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	208 #endif
8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	209
8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	210 #endif //!RUNTIME_CPUDETECT
8123 9fc45fe0d444 HUGE set of compiler warning fixes, unused variables removal arpi parents: 7072 diff changeset	211 return to;
4681 8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	212 }
8db59073127e mem2agpcpy() michael parents: 3393 diff changeset	213
27341 e7c989f7a7c9 Start unifying names of internal preprocessor directives. diego parents: 23523 diff changeset	214 #endif /* CONFIG_FASTMEMCPY */

Mercurial > mplayer.hg

annotate libvo/aclib.c @ 28335:31287e75b5d8