The Tor Browser / file comparison
comparison: media/libtheora/lib/ocintrin.h
media/libtheora/lib/ocintrin.h
- changeset 0
- 6474c204b198
equal
deleted
inserted
replaced
| -1:000000000000 | 0:63f4bc6ab9ea |
|---|---|
| 1 /******************************************************************** | |
| 2 * * | |
| 3 * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE. * | |
| 4 * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS * | |
| 5 * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE * | |
| 6 * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING. * | |
| 7 * * | |
| 8 * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009 * | |
| 9 * by the Xiph.Org Foundation and contributors http://www.xiph.org/ * | |
| 10 * * | |
| 11 ******************************************************************** | |
| 12 | |
| 13 function: | |
| 14 last mod: $Id: ocintrin.h 16503 2009-08-22 18:14:02Z giles $ | |
| 15 | |
| 16 ********************************************************************/ | |
| 17 | |
| 18 /*Some common macros for potential platform-specific optimization.*/ | |
| 19 #include <math.h> | |
| 20 #if !defined(_ocintrin_H) | |
| 21 # define _ocintrin_H (1) | |
| 22 | |
| 23 /*Some specific platforms may have optimized intrinsic or inline assembly | |
| 24 versions of these functions which can substantially improve performance. | |
| 25 We define macros for them to allow easy incorporation of these non-ANSI | |
| 26 features.*/ | |
| 27 | |
| 28 /*Note that we do not provide a macro for abs(), because it is provided as a | |
| 29 library function, which we assume is translated into an intrinsic to avoid | |
| 30 the function call overhead and then implemented in the smartest way for the | |
| 31 target platform. | |
| 32 With modern gcc (4.x), this is true: it uses cmov instructions if the | |
| 33 architecture supports it and branchless bit-twiddling if it does not (the | |
| 34 speed difference between the two approaches is not measurable). | |
| 35 Interestingly, the bit-twiddling method was patented in 2000 (US 6,073,150) | |
| 36 by Sun Microsystems, despite prior art dating back to at least 1996: | |
| 37 http://web.archive.org/web/19961201174141/www.x86.org/ftp/articles/pentopt/PENTOPT.TXT | |
| 38 On gcc 3.x, however, our assumption is not true, as abs() is translated to a | |
| 39 conditional jump, which is horrible on deeply piplined architectures (e.g., | |
| 40 all consumer architectures for the past decade or more). | |
| 41 Also be warned that -C*abs(x) where C is a constant is mis-optimized as | |
| 42 abs(C*x) on every gcc release before 4.2.3. | |
| 43 See bug http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34130 */ | |
| 44 | |
| 45 /*Modern gcc (4.x) can compile the naive versions of min and max with cmov if | |
| 46 given an appropriate architecture, but the branchless bit-twiddling versions | |
| 47 are just as fast, and do not require any special target architecture. | |
| 48 Earlier gcc versions (3.x) compiled both code to the same assembly | |
| 49 instructions, because of the way they represented ((_b)>(_a)) internally.*/ | |
| 50 #define OC_MAXI(_a,_b) ((_a)-((_a)-(_b)&-((_b)>(_a)))) | |
| 51 #define OC_MINI(_a,_b) ((_a)+((_b)-(_a)&-((_b)<(_a)))) | |
| 52 /*Clamps an integer into the given range. | |
| 53 If _a>_c, then the lower bound _a is respected over the upper bound _c (this | |
| 54 behavior is required to meet our documented API behavior). | |
| 55 _a: The lower bound. | |
| 56 _b: The value to clamp. | |
| 57 _c: The upper boud.*/ | |
| 58 #define OC_CLAMPI(_a,_b,_c) (OC_MAXI(_a,OC_MINI(_b,_c))) | |
| 59 #define OC_CLAMP255(_x) ((unsigned char)((((_x)<0)-1)&((_x)|-((_x)>255)))) | |
| 60 /*This has a chance of compiling branchless, and is just as fast as the | |
| 61 bit-twiddling method, which is slightly less portable, since it relies on a | |
| 62 sign-extended rightshift, which is not guaranteed by ANSI (but present on | |
| 63 every relevant platform).*/ | |
| 64 #define OC_SIGNI(_a) (((_a)>0)-((_a)<0)) | |
| 65 /*Slightly more portable than relying on a sign-extended right-shift (which is | |
| 66 not guaranteed by ANSI), and just as fast, since gcc (3.x and 4.x both) | |
| 67 compile it into the right-shift anyway.*/ | |
| 68 #define OC_SIGNMASK(_a) (-((_a)<0)) | |
| 69 /*Divides an integer by a power of two, truncating towards 0. | |
| 70 _dividend: The integer to divide. | |
| 71 _shift: The non-negative power of two to divide by. | |
| 72 _rmask: (1<<_shift)-1*/ | |
| 73 #define OC_DIV_POW2(_dividend,_shift,_rmask)\ | |
| 74 ((_dividend)+(OC_SIGNMASK(_dividend)&(_rmask))>>(_shift)) | |
| 75 /*Divides _x by 65536, truncating towards 0.*/ | |
| 76 #define OC_DIV2_16(_x) OC_DIV_POW2(_x,16,0xFFFF) | |
| 77 /*Divides _x by 2, truncating towards 0.*/ | |
| 78 #define OC_DIV2(_x) OC_DIV_POW2(_x,1,0x1) | |
| 79 /*Divides _x by 8, truncating towards 0.*/ | |
| 80 #define OC_DIV8(_x) OC_DIV_POW2(_x,3,0x7) | |
| 81 /*Divides _x by 16, truncating towards 0.*/ | |
| 82 #define OC_DIV16(_x) OC_DIV_POW2(_x,4,0xF) | |
| 83 /*Right shifts _dividend by _shift, adding _rval, and subtracting one for | |
| 84 negative dividends first. | |
| 85 When _rval is (1<<_shift-1), this is equivalent to division with rounding | |
| 86 ties away from zero.*/ | |
| 87 #define OC_DIV_ROUND_POW2(_dividend,_shift,_rval)\ | |
| 88 ((_dividend)+OC_SIGNMASK(_dividend)+(_rval)>>(_shift)) | |
| 89 /*Divides a _x by 2, rounding towards even numbers.*/ | |
| 90 #define OC_DIV2_RE(_x) ((_x)+((_x)>>1&1)>>1) | |
| 91 /*Divides a _x by (1<<(_shift)), rounding towards even numbers.*/ | |
| 92 #define OC_DIV_POW2_RE(_x,_shift) \ | |
| 93 ((_x)+((_x)>>(_shift)&1)+((1<<(_shift))-1>>1)>>(_shift)) | |
| 94 /*Swaps two integers _a and _b if _a>_b.*/ | |
| 95 #define OC_SORT2I(_a,_b) \ | |
| 96 do{ \ | |
| 97 int t__; \ | |
| 98 t__=((_a)^(_b))&-((_b)<(_a)); \ | |
| 99 (_a)^=t__; \ | |
| 100 (_b)^=t__; \ | |
| 101 } \ | |
| 102 while(0) | |
| 103 | |
| 104 /*Accesses one of four (signed) bytes given an index. | |
| 105 This can be used to avoid small lookup tables.*/ | |
| 106 #define OC_BYTE_TABLE32(_a,_b,_c,_d,_i) \ | |
| 107 ((signed char) \ | |
| 108 (((_a)&0xFF|((_b)&0xFF)<<8|((_c)&0xFF)<<16|((_d)&0xFF)<<24)>>(_i)*8)) | |
| 109 /*Accesses one of eight (unsigned) nibbles given an index. | |
| 110 This can be used to avoid small lookup tables.*/ | |
| 111 #define OC_UNIBBLE_TABLE32(_a,_b,_c,_d,_e,_f,_g,_h,_i) \ | |
| 112 ((((_a)&0xF|((_b)&0xF)<<4|((_c)&0xF)<<8|((_d)&0xF)<<12| \ | |
| 113 ((_e)&0xF)<<16|((_f)&0xF)<<20|((_g)&0xF)<<24|((_h)&0xF)<<28)>>(_i)*4)&0xF) | |
| 114 | |
| 115 | |
| 116 | |
| 117 /*All of these macros should expect floats as arguments.*/ | |
| 118 #define OC_MAXF(_a,_b) ((_a)<(_b)?(_b):(_a)) | |
| 119 #define OC_MINF(_a,_b) ((_a)>(_b)?(_b):(_a)) | |
| 120 #define OC_CLAMPF(_a,_b,_c) (OC_MINF(_a,OC_MAXF(_b,_c))) | |
| 121 #define OC_FABSF(_f) ((float)fabs(_f)) | |
| 122 #define OC_SQRTF(_f) ((float)sqrt(_f)) | |
| 123 #define OC_POWF(_b,_e) ((float)pow(_b,_e)) | |
| 124 #define OC_LOGF(_f) ((float)log(_f)) | |
| 125 #define OC_IFLOORF(_f) ((int)floor(_f)) | |
| 126 #define OC_ICEILF(_f) ((int)ceil(_f)) | |
| 127 | |
| 128 #endif |
