… | |
… | |
40 | |
40 | |
41 | #ifndef ECB_H |
41 | #ifndef ECB_H |
42 | #define ECB_H |
42 | #define ECB_H |
43 | |
43 | |
44 | /* 16 bits major, 16 bits minor */ |
44 | /* 16 bits major, 16 bits minor */ |
45 | #define ECB_VERSION 0x00010003 |
45 | #define ECB_VERSION 0x00010004 |
46 | |
46 | |
47 | #ifdef _WIN32 |
47 | #ifdef _WIN32 |
48 | typedef signed char int8_t; |
48 | typedef signed char int8_t; |
49 | typedef unsigned char uint8_t; |
49 | typedef unsigned char uint8_t; |
50 | typedef signed short int16_t; |
50 | typedef signed short int16_t; |
… | |
… | |
98 | #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) |
98 | #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) |
99 | #endif |
99 | #endif |
100 | |
100 | |
101 | #define ECB_CLANG_VERSION(major,minor) (__clang_major__ > (major) || (__clang_major__ == (major) && __clang_minor__ >= (minor))) |
101 | #define ECB_CLANG_VERSION(major,minor) (__clang_major__ > (major) || (__clang_major__ == (major) && __clang_minor__ >= (minor))) |
102 | |
102 | |
103 | #if __clang__ && defined(__has_builtin) |
103 | #if __clang__ && defined __has_builtin |
104 | #define ECB_CLANG_BUILTIN(x) __has_builtin(x) |
104 | #define ECB_CLANG_BUILTIN(x) __has_builtin (x) |
105 | #else |
105 | #else |
106 | #define ECB_CLANG_BUILTIN(x) 0 |
106 | #define ECB_CLANG_BUILTIN(x) 0 |
|
|
107 | #endif |
|
|
108 | |
|
|
109 | #if __clang__ && defined __has_extension |
|
|
110 | #define ECB_CLANG_EXTENSION(x) __has_extension (x) |
|
|
111 | #else |
|
|
112 | #define ECB_CLANG_EXTENSION(x) 0 |
107 | #endif |
113 | #endif |
108 | |
114 | |
109 | #define ECB_CPP (__cplusplus+0) |
115 | #define ECB_CPP (__cplusplus+0) |
110 | #define ECB_CPP11 (__cplusplus >= 201103L) |
116 | #define ECB_CPP11 (__cplusplus >= 201103L) |
111 | |
117 | |
… | |
… | |
195 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
201 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
196 | #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) |
202 | #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) |
197 | #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE) |
203 | #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE) |
198 | #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE) |
204 | #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE) |
199 | |
205 | |
200 | /* The __has_feature syntax from clang is so misdesigned that we cannot use it |
206 | #elif ECB_CLANG_EXTENSION(c_atomic) |
201 | * without risking compile time errors with other compilers. We *could* |
|
|
202 | * define our own ecb_clang_has_feature, but I just can't be bothered to work |
|
|
203 | * around this shit time and again. |
|
|
204 | * #elif defined __clang && __has_feature (cxx_atomic) |
|
|
205 | * // see comment below (stdatomic.h) about the C11 memory model. |
207 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
206 | * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) |
208 | #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) |
207 | * #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE) |
209 | #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE) |
208 | * #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE) |
210 | #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE) |
209 | */ |
|
|
210 | |
211 | |
211 | #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ |
212 | #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ |
212 | #define ECB_MEMORY_FENCE __sync_synchronize () |
213 | #define ECB_MEMORY_FENCE __sync_synchronize () |
213 | #elif _MSC_VER >= 1500 /* VC++ 2008 */ |
214 | #elif _MSC_VER >= 1500 /* VC++ 2008 */ |
214 | /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */ |
215 | /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */ |
… | |
… | |
277 | #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE |
278 | #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE |
278 | #endif |
279 | #endif |
279 | |
280 | |
280 | /*****************************************************************************/ |
281 | /*****************************************************************************/ |
281 | |
282 | |
282 | #if __cplusplus |
283 | #if ECB_CPP |
283 | #define ecb_inline static inline |
284 | #define ecb_inline static inline |
284 | #elif ECB_GCC_VERSION(2,5) |
285 | #elif ECB_GCC_VERSION(2,5) |
285 | #define ecb_inline static __inline__ |
286 | #define ecb_inline static __inline__ |
286 | #elif ECB_C99 |
287 | #elif ECB_C99 |
287 | #define ecb_inline static inline |
288 | #define ecb_inline static inline |
… | |
… | |
305 | #define ECB_STRINGIFY(a) ECB_STRINGIFY_(a) |
306 | #define ECB_STRINGIFY(a) ECB_STRINGIFY_(a) |
306 | |
307 | |
307 | #define ecb_function_ ecb_inline |
308 | #define ecb_function_ ecb_inline |
308 | |
309 | |
309 | #if ECB_GCC_VERSION(3,1) || ECB_CLANG_VERSION(2,8) |
310 | #if ECB_GCC_VERSION(3,1) || ECB_CLANG_VERSION(2,8) |
310 | #define ecb_attribute(attrlist) __attribute__(attrlist) |
311 | #define ecb_attribute(attrlist) __attribute__ (attrlist) |
311 | #else |
312 | #else |
312 | #define ecb_attribute(attrlist) |
313 | #define ecb_attribute(attrlist) |
313 | #endif |
314 | #endif |
314 | |
315 | |
315 | #if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_constant_p) |
316 | #if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_constant_p) |
… | |
… | |
333 | #else |
334 | #else |
334 | #define ecb_prefetch(addr,rw,locality) |
335 | #define ecb_prefetch(addr,rw,locality) |
335 | #endif |
336 | #endif |
336 | |
337 | |
337 | /* no emulation for ecb_decltype */ |
338 | /* no emulation for ecb_decltype */ |
338 | #if ECB_GCC_VERSION(4,5) |
339 | #if ECB_CPP11 |
|
|
340 | // older implementations might have problems with decltype(x)::type, work around it |
|
|
341 | template<class T> struct ecb_decltype_t { typedef T type; }; |
339 | #define ecb_decltype(x) __decltype(x) |
342 | #define ecb_decltype(x) ecb_decltype_t<decltype (x)>::type |
340 | #elif ECB_GCC_VERSION(3,0) |
343 | #elif ECB_GCC_VERSION(3,0) || ECB_CLANG_VERSION(2,8) |
341 | #define ecb_decltype(x) __typeof(x) |
344 | #define ecb_decltype(x) __typeof__ (x) |
342 | #endif |
345 | #endif |
343 | |
346 | |
344 | #if _MSC_VER >= 1300 |
347 | #if _MSC_VER >= 1300 |
345 | #define ecb_deprecated __declspec(deprecated) |
348 | #define ecb_deprecated __declspec (deprecated) |
346 | #else |
349 | #else |
347 | #define ecb_deprecated ecb_attribute ((__deprecated__)) |
350 | #define ecb_deprecated ecb_attribute ((__deprecated__)) |
348 | #endif |
351 | #endif |
349 | |
352 | |
350 | #define ecb_noinline ecb_attribute ((__noinline__)) |
353 | #define ecb_noinline ecb_attribute ((__noinline__)) |
351 | #define ecb_unused ecb_attribute ((__unused__)) |
354 | #define ecb_unused ecb_attribute ((__unused__)) |
352 | #define ecb_const ecb_attribute ((__const__)) |
355 | #define ecb_const ecb_attribute ((__const__)) |
353 | #define ecb_pure ecb_attribute ((__pure__)) |
356 | #define ecb_pure ecb_attribute ((__pure__)) |
354 | |
357 | |
355 | /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx __declspec(noreturn) */ |
358 | /* TODO http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx __declspec(noreturn) */ |
356 | #if ECB_C11 |
359 | #if ECB_C11 || __IBMC_NORETURN |
|
|
360 | /* http://pic.dhe.ibm.com/infocenter/compbg/v121v141/topic/com.ibm.xlcpp121.bg.doc/language_ref/noreturn.html */ |
357 | #define ecb_noreturn _Noreturn |
361 | #define ecb_noreturn _Noreturn |
358 | #else |
362 | #else |
359 | #define ecb_noreturn ecb_attribute ((__noreturn__)) |
363 | #define ecb_noreturn ecb_attribute ((__noreturn__)) |
360 | #endif |
364 | #endif |
361 | |
365 | |
… | |
… | |
377 | /* for compatibility to the rest of the world */ |
381 | /* for compatibility to the rest of the world */ |
378 | #define ecb_likely(expr) ecb_expect_true (expr) |
382 | #define ecb_likely(expr) ecb_expect_true (expr) |
379 | #define ecb_unlikely(expr) ecb_expect_false (expr) |
383 | #define ecb_unlikely(expr) ecb_expect_false (expr) |
380 | |
384 | |
381 | /* count trailing zero bits and count # of one bits */ |
385 | /* count trailing zero bits and count # of one bits */ |
|
|
386 | #if ECB_GCC_VERSION(3,4) \ |
382 | #if ECB_GCC_VERSION(3,4) || (ECB_CLANG_BUILTIN(__builtin_clz) && ECB_CLANG_BUILTIN(__builtin_clzll) \ |
387 | || (ECB_CLANG_BUILTIN(__builtin_clz) && ECB_CLANG_BUILTIN(__builtin_clzll) \ |
383 | && ECB_CLANG_BUILTIN(__builtin_ctz) && ECB_CLANG_BUILTIN(__builtin_ctzll) \ |
388 | && ECB_CLANG_BUILTIN(__builtin_ctz) && ECB_CLANG_BUILTIN(__builtin_ctzll) \ |
384 | && ECB_CLANG_BUILTIN(__builtin_popcount)) |
389 | && ECB_CLANG_BUILTIN(__builtin_popcount)) |
385 | /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */ |
390 | /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */ |
386 | #define ecb_ld32(x) (__builtin_clz (x) ^ 31) |
391 | #define ecb_ld32(x) (__builtin_clz (x) ^ 31) |
387 | #define ecb_ld64(x) (__builtin_clzll (x) ^ 63) |
392 | #define ecb_ld64(x) (__builtin_clzll (x) ^ 63) |
388 | #define ecb_ctz32(x) __builtin_ctz (x) |
393 | #define ecb_ctz32(x) __builtin_ctz (x) |
389 | #define ecb_ctz64(x) __builtin_ctzll (x) |
394 | #define ecb_ctz64(x) __builtin_ctzll (x) |
… | |
… | |
595 | #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) |
600 | #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) |
596 | #else |
601 | #else |
597 | #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) |
602 | #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) |
598 | #endif |
603 | #endif |
599 | |
604 | |
600 | #if __cplusplus |
605 | #if ECB_CPP |
601 | template<typename T> |
606 | template<typename T> |
602 | static inline T ecb_div_rd (T val, T div) |
607 | static inline T ecb_div_rd (T val, T div) |
603 | { |
608 | { |
604 | return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; |
609 | return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; |
605 | } |
610 | } |
… | |
… | |
665 | #define ECB_NAN NAN |
670 | #define ECB_NAN NAN |
666 | #else |
671 | #else |
667 | #define ECB_NAN ECB_INFINITY |
672 | #define ECB_NAN ECB_INFINITY |
668 | #endif |
673 | #endif |
669 | |
674 | |
|
|
675 | #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L |
|
|
676 | #define ecb_ldexpf(x,e) ldexpf (x, e) |
|
|
677 | #else |
|
|
678 | #define ecb_ldexpf(x,e) (float) ldexp ((double)(x), e) |
|
|
679 | #endif |
|
|
680 | |
670 | /* converts an ieee half/binary16 to a float */ |
681 | /* converts an ieee half/binary16 to a float */ |
671 | ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const; |
682 | ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const; |
672 | ecb_function_ float |
683 | ecb_function_ float |
673 | ecb_binary16_to_float (uint16_t x) |
684 | ecb_binary16_to_float (uint16_t x) |
674 | { |
685 | { |
675 | int e = (x >> 10) & 0x1f; |
686 | int e = (x >> 10) & 0x1f; |
676 | int m = x & 0x3ff; |
687 | int m = x & 0x3ff; |
677 | float r; |
688 | float r; |
678 | |
689 | |
679 | if (!e ) r = ldexpf (m , -24); |
690 | if (!e ) r = ecb_ldexpf (m , -24); |
680 | else if (e != 31) r = ldexpf (m + 0x400, e - 25); |
691 | else if (e != 31) r = ecb_ldexpf (m + 0x400, e - 25); |
681 | else if (m ) r = ECB_NAN; |
692 | else if (m ) r = ECB_NAN; |
682 | else r = ECB_INFINITY; |
693 | else r = ECB_INFINITY; |
683 | |
694 | |
684 | return x & 0x8000 ? -r : r; |
695 | return x & 0x8000 ? -r : r; |
685 | } |
696 | } |
… | |
… | |
744 | x |= 0x800000U; |
755 | x |= 0x800000U; |
745 | else |
756 | else |
746 | e = 1; |
757 | e = 1; |
747 | |
758 | |
748 | /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */ |
759 | /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */ |
749 | r = ldexpf (x * (0.5f / 0x800000U), e - 126); |
760 | r = ecb_ldexpf (x * (0.5f / 0x800000U), e - 126); |
750 | |
761 | |
751 | r = neg ? -r : r; |
762 | r = neg ? -r : r; |
752 | #endif |
763 | #endif |
753 | |
764 | |
754 | return r; |
765 | return r; |