ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/OpenCL/ecb.h
(Generate patch)

Comparing OpenCL/ecb.h (file contents):
Revision 1.2 by root, Thu Oct 11 05:02:54 2012 UTC vs.
Revision 1.4 by root, Mon Apr 7 22:36:54 2014 UTC

1/* 1/*
2 * libecb - http://software.schmorp.de/pkg/libecb 2 * libecb - http://software.schmorp.de/pkg/libecb
3 * 3 *
4 * Copyright (©) 2009-2012 Marc Alexander Lehmann <libecb@schmorp.de> 4 * Copyright (©) 2009-2014 Marc Alexander Lehmann <libecb@schmorp.de>
5 * Copyright (©) 2011 Emanuele Giaquinta 5 * Copyright (©) 2011 Emanuele Giaquinta
6 * All rights reserved. 6 * All rights reserved.
7 * 7 *
8 * Redistribution and use in source and binary forms, with or without modifica- 8 * Redistribution and use in source and binary forms, with or without modifica-
9 * tion, are permitted provided that the following conditions are met: 9 * tion, are permitted provided that the following conditions are met:
… …
29 29
30#ifndef ECB_H 30#ifndef ECB_H
31#define ECB_H 31#define ECB_H
32 32
33/* 16 bits major, 16 bits minor */ 33/* 16 bits major, 16 bits minor */
34#define ECB_VERSION 0x00010002 34#define ECB_VERSION 0x00010003
35 35
36#ifdef _WIN32 36#ifdef _WIN32
37 typedef signed char int8_t; 37 typedef signed char int8_t;
38 typedef unsigned char uint8_t; 38 typedef unsigned char uint8_t;
39 typedef signed short int16_t; 39 typedef signed short int16_t;
… …
63 #else 63 #else
64 #define ECB_PTRSIZE 4 64 #define ECB_PTRSIZE 4
65 #endif 65 #endif
66#endif 66#endif
67 67
68/* work around x32 idiocy by defining proper macros */
69#if __amd64 || __x86_64 || _M_AMD64 || _M_X64
70 #if _ILP32
71 #define ECB_AMD64_X32 1
72 #else
73 #define ECB_AMD64 1
74 #endif
75#endif
76
68/* many compilers define _GNUC_ to some versions but then only implement 77/* many compilers define _GNUC_ to some versions but then only implement
69 * what their idiot authors think are the "more important" extensions, 78 * what their idiot authors think are the "more important" extensions,
70 * causing enormous grief in return for some better fake benchmark numbers. 79 * causing enormous grief in return for some better fake benchmark numbers.
71 * or so. 80 * or so.
72 * we try to detect these and simply assume they are not gcc - if they have 81 * we try to detect these and simply assume they are not gcc - if they have
… …
78 #else 87 #else
79 #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) 88 #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor)))
80 #endif 89 #endif
81#endif 90#endif
82 91
83#define ECB_C (__STDC__+0) /* this assumes that __STDC__ is either empty or a number */
84#define ECB_C99 (__STDC_VERSION__ >= 199901L)
85#define ECB_C11 (__STDC_VERSION__ >= 201112L)
86#define ECB_CPP (__cplusplus+0) 92#define ECB_CPP (__cplusplus+0)
87#define ECB_CPP11 (__cplusplus >= 201103L) 93#define ECB_CPP11 (__cplusplus >= 201103L)
94
95#if ECB_CPP
96 #define ECB_C 0
97 #define ECB_STDC_VERSION 0
98#else
99 #define ECB_C 1
100 #define ECB_STDC_VERSION __STDC_VERSION__
101#endif
102
103#define ECB_C99 (ECB_STDC_VERSION >= 199901L)
104#define ECB_C11 (ECB_STDC_VERSION >= 201112L)
88 105
89#if ECB_CPP 106#if ECB_CPP
90 #define ECB_EXTERN_C extern "C" 107 #define ECB_EXTERN_C extern "C"
91 #define ECB_EXTERN_C_BEG ECB_EXTERN_C { 108 #define ECB_EXTERN_C_BEG ECB_EXTERN_C {
92 #define ECB_EXTERN_C_END } 109 #define ECB_EXTERN_C_END }
… …
125 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ 142 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__
126 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") 143 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory")
127 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ 144 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
128 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 145 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__
129 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 146 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
147 #elif __aarch64__
148 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
130 #elif __sparc || __sparc__ 149 #elif (__sparc || __sparc__) && !__sparcv8
131 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") 150 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
132 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") 151 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory")
133 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") 152 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore")
134 #elif defined __s390__ || defined __s390x__ 153 #elif defined __s390__ || defined __s390x__
135 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") 154 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory")
136 #elif defined __mips__ 155 #elif defined __mips__
156 /* GNU/Linux emulates sync on mips1 architectures, so we force its use */
157 /* anybody else who still uses mips1 is supposed to send in their version, with detection code. */
137 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 158 #define ECB_MEMORY_FENCE __asm__ __volatile__ (".set mips2; sync; .set mips0" : : : "memory")
138 #elif defined __alpha__ 159 #elif defined __alpha__
139 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory") 160 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory")
140 #elif defined __hppa__ 161 #elif defined __hppa__
141 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") 162 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory")
142 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 163 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("")
143 #elif defined __ia64__ 164 #elif defined __ia64__
144 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mf" : : : "memory") 165 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mf" : : : "memory")
166 #elif defined __m68k__
167 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory")
168 #elif defined __m88k__
169 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("tb1 0,%%r0,128" : : : "memory")
170 #elif defined __sh__
171 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory")
145 #endif 172 #endif
146 #endif 173 #endif
147#endif 174#endif
148 175
149#ifndef ECB_MEMORY_FENCE 176#ifndef ECB_MEMORY_FENCE
150 #if ECB_GCC_VERSION(4,7) 177 #if ECB_GCC_VERSION(4,7)
151 /* see comment below (stdatomic.h) about the C11 memory model. */ 178 /* see comment below (stdatomic.h) about the C11 memory model. */
152 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) 179 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST)
180 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE)
181 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE)
153 182
154 /* The __has_feature syntax from clang is so misdesigned that we cannot use it 183 /* The __has_feature syntax from clang is so misdesigned that we cannot use it
155 * without risking compile time errors with other compilers. We *could* 184 * without risking compile time errors with other compilers. We *could*
156 * define our own ecb_clang_has_feature, but I just can't be bothered to work 185 * define our own ecb_clang_has_feature, but I just can't be bothered to work
157 * around this shit time and again. 186 * around this shit time and again.
158 * #elif defined __clang && __has_feature (cxx_atomic) 187 * #elif defined __clang && __has_feature (cxx_atomic)
159 * // see comment below (stdatomic.h) about the C11 memory model. 188 * // see comment below (stdatomic.h) about the C11 memory model.
160 * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) 189 * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)
190 * #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE)
191 * #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE)
161 */ 192 */
162 193
163 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ 194 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__
164 #define ECB_MEMORY_FENCE __sync_synchronize () 195 #define ECB_MEMORY_FENCE __sync_synchronize ()
196 #elif _MSC_VER >= 1500 /* VC++ 2008 */
197 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */
198 #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier)
199 #define ECB_MEMORY_FENCE _ReadWriteBarrier (); MemoryBarrier()
200 #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier (); MemoryBarrier() /* according to msdn, _ReadBarrier is not a load fence */
201 #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier (); MemoryBarrier()
165 #elif _MSC_VER >= 1400 /* VC++ 2005 */ 202 #elif _MSC_VER >= 1400 /* VC++ 2005 */
166 #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) 203 #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier)
167 #define ECB_MEMORY_FENCE _ReadWriteBarrier () 204 #define ECB_MEMORY_FENCE _ReadWriteBarrier ()
168 #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */ 205 #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */
169 #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier () 206 #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier ()
… …
189 /* any fence other than seq_cst, which isn't very efficient for us. */ 226 /* any fence other than seq_cst, which isn't very efficient for us. */
190 /* Why that is, we don't know - either the C11 memory model is quite useless */ 227 /* Why that is, we don't know - either the C11 memory model is quite useless */
191 /* for most usages, or gcc and clang have a bug */ 228 /* for most usages, or gcc and clang have a bug */
192 /* I *currently* lean towards the latter, and inefficiently implement */ 229 /* I *currently* lean towards the latter, and inefficiently implement */
193 /* all three of ecb's fences as a seq_cst fence */ 230 /* all three of ecb's fences as a seq_cst fence */
231 /* Update, gcc-4.8 generates mfence for all c++ fences, but nothing */
232 /* for all __atomic_thread_fence's except seq_cst */
194 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst) 233 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst)
195 #endif 234 #endif
196#endif 235#endif
197 236
198#ifndef ECB_MEMORY_FENCE 237#ifndef ECB_MEMORY_FENCE
… …
255 #define ecb_is_constant(expr) __builtin_constant_p (expr) 294 #define ecb_is_constant(expr) __builtin_constant_p (expr)
256 #define ecb_expect(expr,value) __builtin_expect ((expr),(value)) 295 #define ecb_expect(expr,value) __builtin_expect ((expr),(value))
257 #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality) 296 #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality)
258#else 297#else
259 #define ecb_attribute(attrlist) 298 #define ecb_attribute(attrlist)
299
300 /* possible C11 impl for integral types
301 typedef struct ecb_is_constant_struct ecb_is_constant_struct;
302 #define ecb_is_constant(expr) _Generic ((1 ? (struct ecb_is_constant_struct *)0 : (void *)((expr) - (expr)), ecb_is_constant_struct *: 0, default: 1)) */
303
260 #define ecb_is_constant(expr) 0 304 #define ecb_is_constant(expr) 0
261 #define ecb_expect(expr,value) (expr) 305 #define ecb_expect(expr,value) (expr)
262 #define ecb_prefetch(addr,rw,locality) 306 #define ecb_prefetch(addr,rw,locality)
263#endif 307#endif
264 308
… …
550/* the only noteworthy exception is ancient armle, which uses order 43218765 */ 594/* the only noteworthy exception is ancient armle, which uses order 43218765 */
551#if 0 \ 595#if 0 \
552 || __i386 || __i386__ \ 596 || __i386 || __i386__ \
553 || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ 597 || __amd64 || __amd64__ || __x86_64 || __x86_64__ \
554 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ 598 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \
555 || defined __arm__ && defined __ARM_EABI__ \
556 || defined __s390__ || defined __s390x__ \ 599 || defined __s390__ || defined __s390x__ \
557 || defined __mips__ \ 600 || defined __mips__ \
558 || defined __alpha__ \ 601 || defined __alpha__ \
559 || defined __hppa__ \ 602 || defined __hppa__ \
560 || defined __ia64__ \ 603 || defined __ia64__ \
604 || defined __m68k__ \
605 || defined __m88k__ \
606 || defined __sh__ \
561 || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 607 || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 \
608 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \
609 || defined __aarch64__
562 #define ECB_STDFP 1 610 #define ECB_STDFP 1
563 #include <string.h> /* for memcpy */ 611 #include <string.h> /* for memcpy */
564#else 612#else
565 #define ECB_STDFP 0 613 #define ECB_STDFP 0
566 #include <math.h> /* for frexp*, ldexp* */
567#endif 614#endif
568 615
569#ifndef ECB_NO_LIBM 616#ifndef ECB_NO_LIBM
617
618 #include <math.h> /* for frexp*, ldexp*, INFINITY, NAN */
619
620 /* only the oldest of old doesn't have this one. solaris. */
621 #ifdef INFINITY
622 #define ECB_INFINITY INFINITY
623 #else
624 #define ECB_INFINITY HUGE_VAL
625 #endif
626
627 #ifdef NAN
628 #define ECB_NAN NAN
629 #else
630 #define ECB_NAN ECB_INFINITY
631 #endif
632
633 /* converts an ieee half/binary16 to a float */
634 ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const;
635 ecb_function_ float
636 ecb_binary16_to_float (uint16_t x)
637 {
638 int e = (x >> 10) & 0x1f;
639 int m = x & 0x3ff;
640 float r;
641
642 if (!e ) r = ldexpf (m , -24);
643 else if (e != 31) r = ldexpf (m + 0x400, e - 25);
644 else if (m ) r = ECB_NAN;
645 else r = ECB_INFINITY;
646
647 return x & 0x8000 ? -r : r;
648 }
570 649
571 /* convert a float to ieee single/binary32 */ 650 /* convert a float to ieee single/binary32 */
572 ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; 651 ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const;
573 ecb_function_ uint32_t 652 ecb_function_ uint32_t
574 ecb_float_to_binary32 (float x) 653 ecb_float_to_binary32 (float x)

Diff Legend

– Removed lines
+ Added lines
< Changed lines
> Changed lines