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

Comparing libecb/ecb.h (file contents):
Revision 1.161 by sf-exg, Thu Mar 19 20:55:22 2015 UTC vs.
Revision 1.173 by root, Tue Nov 24 22:47:12 2015 UTC

… …
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 0x00010004 45#define ECB_VERSION 0x00010005
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;
… …
67 typedef uint32_t uintptr_t; 67 typedef uint32_t uintptr_t;
68 typedef int32_t intptr_t; 68 typedef int32_t intptr_t;
69 #endif 69 #endif
70#else 70#else
71 #include <inttypes.h> 71 #include <inttypes.h>
72 #if UINTMAX_MAX > 0xffffffffU 72 #if (defined INTPTR_MAX ? INTPTR_MAX : ULONG_MAX) > 0xffffffffU
73 #define ECB_PTRSIZE 8 73 #define ECB_PTRSIZE 8
74 #else 74 #else
75 #define ECB_PTRSIZE 4 75 #define ECB_PTRSIZE 4
76 #endif 76 #endif
77#endif 77#endif
… …
148 #define ECB_NO_SMP 1 148 #define ECB_NO_SMP 1
149#endif 149#endif
150 150
151#if ECB_NO_SMP 151#if ECB_NO_SMP
152 #define ECB_MEMORY_FENCE do { } while (0) 152 #define ECB_MEMORY_FENCE do { } while (0)
153#endif
154
155/* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/compiler_ref/compiler_builtins.html */
156#if __xlC__ && ECB_CPP
157 #include <builtins.h>
158#endif
159
160#if 1400 <= _MSC_VER
161 #include <intrin.h> /* fence functions _ReadBarrier, also bit search functions _BitScanReverse */
153#endif 162#endif
154 163
155#ifndef ECB_MEMORY_FENCE 164#ifndef ECB_MEMORY_FENCE
156 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 165 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
157 #if __i386 || __i386__ 166 #if __i386 || __i386__
… …
170 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ 179 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
171 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 180 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__
172 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 181 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
173 #elif __aarch64__ 182 #elif __aarch64__
174 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory") 183 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
175 #elif (__sparc || __sparc__) && !__sparcv8 184 #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8)
176 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") 185 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
177 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") 186 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory")
178 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") 187 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore")
179 #elif defined __s390__ || defined __s390x__ 188 #elif defined __s390__ || defined __s390x__
180 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") 189 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory")
… …
352 #define ecb_deprecated __declspec (deprecated) 361 #define ecb_deprecated __declspec (deprecated)
353#else 362#else
354 #define ecb_deprecated ecb_attribute ((__deprecated__)) 363 #define ecb_deprecated ecb_attribute ((__deprecated__))
355#endif 364#endif
356 365
357#if __MSC_VER >= 1500 366#if _MSC_VER >= 1500
358 #define ecb_deprecated_message(msg) __declspec (deprecated (msg)) 367 #define ecb_deprecated_message(msg) __declspec (deprecated (msg))
359#elif ECB_GCC_VERSION(4,5) 368#elif ECB_GCC_VERSION(4,5)
360 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg)) 369 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg))
361#else 370#else
362 #define ecb_deprecated_message(msg) ecb_deprecated 371 #define ecb_deprecated_message(msg) ecb_deprecated
… …
371#define ecb_unused ecb_attribute ((__unused__)) 380#define ecb_unused ecb_attribute ((__unused__))
372#define ecb_const ecb_attribute ((__const__)) 381#define ecb_const ecb_attribute ((__const__))
373#define ecb_pure ecb_attribute ((__pure__)) 382#define ecb_pure ecb_attribute ((__pure__))
374 383
375#if ECB_C11 || __IBMC_NORETURN 384#if ECB_C11 || __IBMC_NORETURN
376 /* http://pic.dhe.ibm.com/infocenter/compbg/v121v141/topic/com.ibm.xlcpp121.bg.doc/language_ref/noreturn.html */ 385 /* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/language_ref/noreturn.html */
377 #define ecb_noreturn _Noreturn 386 #define ecb_noreturn _Noreturn
378#elif ECB_CPP11 387#elif ECB_CPP11
379 #define ecb_noreturn [[noreturn]] 388 #define ecb_noreturn [[noreturn]]
380#elif _MSC_VER >= 1200 389#elif _MSC_VER >= 1200
381 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */ 390 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */
… …
418#else 427#else
419 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x); 428 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x);
420 ecb_function_ ecb_const int 429 ecb_function_ ecb_const int
421 ecb_ctz32 (uint32_t x) 430 ecb_ctz32 (uint32_t x)
422 { 431 {
432#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
433 unsigned long r;
434 _BitScanForward (&r, x);
435 return (int)r;
436#else
423 int r = 0; 437 int r = 0;
424 438
425 x &= ~x + 1; /* this isolates the lowest bit */ 439 x &= ~x + 1; /* this isolates the lowest bit */
426 440
427#if ECB_branchless_on_i386 441#if ECB_branchless_on_i386
… …
437 if (x & 0xff00ff00) r += 8; 451 if (x & 0xff00ff00) r += 8;
438 if (x & 0xffff0000) r += 16; 452 if (x & 0xffff0000) r += 16;
439#endif 453#endif
440 454
441 return r; 455 return r;
456#endif
442 } 457 }
443 458
444 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x); 459 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
445 ecb_function_ ecb_const int 460 ecb_function_ ecb_const int
446 ecb_ctz64 (uint64_t x) 461 ecb_ctz64 (uint64_t x)
447 { 462 {
463#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
464 unsigned long r;
465 _BitScanForward64 (&r, x);
466 return (int)r;
467#else
448 int shift = x & 0xffffffffU ? 0 : 32; 468 int shift = x & 0xffffffff ? 0 : 32;
449 return ecb_ctz32 (x >> shift) + shift; 469 return ecb_ctz32 (x >> shift) + shift;
470#endif
450 } 471 }
451 472
452 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x); 473 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
453 ecb_function_ ecb_const int 474 ecb_function_ ecb_const int
454 ecb_popcount32 (uint32_t x) 475 ecb_popcount32 (uint32_t x)
… …
462 } 483 }
463 484
464 ecb_function_ ecb_const int ecb_ld32 (uint32_t x); 485 ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
465 ecb_function_ ecb_const int ecb_ld32 (uint32_t x) 486 ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
466 { 487 {
488#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
489 unsigned long r;
490 _BitScanReverse (&r, x);
491 return (int)r;
492#else
467 int r = 0; 493 int r = 0;
468 494
469 if (x >> 16) { x >>= 16; r += 16; } 495 if (x >> 16) { x >>= 16; r += 16; }
470 if (x >> 8) { x >>= 8; r += 8; } 496 if (x >> 8) { x >>= 8; r += 8; }
471 if (x >> 4) { x >>= 4; r += 4; } 497 if (x >> 4) { x >>= 4; r += 4; }
472 if (x >> 2) { x >>= 2; r += 2; } 498 if (x >> 2) { x >>= 2; r += 2; }
473 if (x >> 1) { r += 1; } 499 if (x >> 1) { r += 1; }
474 500
475 return r; 501 return r;
502#endif
476 } 503 }
477 504
478 ecb_function_ ecb_const int ecb_ld64 (uint64_t x); 505 ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
479 ecb_function_ ecb_const int ecb_ld64 (uint64_t x) 506 ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
480 { 507 {
508#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
509 unsigned long r;
510 _BitScanReverse64 (&r, x);
511 return (int)r;
512#else
481 int r = 0; 513 int r = 0;
482 514
483 if (x >> 32) { x >>= 32; r += 32; } 515 if (x >> 32) { x >>= 32; r += 32; }
484 516
485 return r + ecb_ld32 (x); 517 return r + ecb_ld32 (x);
518#endif
486 } 519 }
487#endif 520#endif
488 521
489ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x); 522ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x);
490ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } 523ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); }
… …
547ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } 580ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); }
548ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } 581ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); }
549ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } 582ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); }
550 583
551#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64)) 584#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64))
585 #if ECB_GCC_VERSION(4,8) || ECB_CLANG_BUILTIN(__builtin_bswap16)
586 #define ecb_bswap16(x) __builtin_bswap16 (x)
587 #else
552 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) 588 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16)
589 #endif
553 #define ecb_bswap32(x) __builtin_bswap32 (x) 590 #define ecb_bswap32(x) __builtin_bswap32 (x)
554 #define ecb_bswap64(x) __builtin_bswap64 (x) 591 #define ecb_bswap64(x) __builtin_bswap64 (x)
592#elif _MSC_VER
593 #include <stdlib.h>
594 #define ecb_bswap16(x) ((uint16_t)_byteswap_ushort ((uint16_t)(x)))
595 #define ecb_bswap32(x) ((uint32_t)_byteswap_ulong ((uint32_t)(x)))
596 #define ecb_bswap64(x) ((uint64_t)_byteswap_uint64 ((uint64_t)(x)))
555#else 597#else
556 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x); 598 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x);
557 ecb_function_ ecb_const uint16_t 599 ecb_function_ ecb_const uint16_t
558 ecb_bswap16 (uint16_t x) 600 ecb_bswap16 (uint16_t x)
559 { 601 {
… …
647 return N; 689 return N;
648 } 690 }
649#else 691#else
650 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) 692 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
651#endif 693#endif
694
695ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
696ecb_function_ ecb_const uint32_t
697ecb_binary16_to_binary32 (uint32_t x)
698{
699 unsigned int s = (x & 0x8000) << (31 - 15);
700 int e = (x >> 10) & 0x001f;
701 unsigned int m = x & 0x03ff;
702
703 if (ecb_expect_false (e == 31))
704 /* infinity or NaN */
705 e = 255 - (127 - 15);
706 else if (ecb_expect_false (!e))
707 {
708 if (ecb_expect_true (!m))
709 /* zero, handled by code below by forcing e to 0 */
710 e = 0 - (127 - 15);
711 else
712 {
713 /* subnormal, renormalise */
714 unsigned int s = 10 - ecb_ld32 (m);
715
716 m = (m << s) & 0x3ff; /* mask implicit bit */
717 e -= s - 1;
718 }
719 }
720
721 /* e and m now are normalised, or zero, (or inf or nan) */
722 e += 127 - 15;
723
724 return s | (e << 23) | (m << (23 - 10));
725}
726
727ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
728ecb_function_ ecb_const uint16_t
729ecb_binary32_to_binary16 (uint32_t x)
730{
731 unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
732 unsigned int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
733 unsigned int m = x & 0x007fffff;
734
735 x &= 0x7fffffff;
736
737 /* if it's within range of binary16 normals, use fast path */
738 if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
739 {
740 /* mantissa round-to-even */
741 m += 0x00000fff + ((m >> (23 - 10)) & 1);
742
743 /* handle overflow */
744 if (ecb_expect_false (m >= 0x00800000))
745 {
746 m >>= 1;
747 e += 1;
748 }
749
750 return s | (e << 10) | (m >> (23 - 10));
751 }
752
753 /* handle large numbers and infinity */
754 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
755 return s | 0x7c00;
756
757 /* handle zero, subnormals and small numbers */
758 if (ecb_expect_true (x < 0x38800000))
759 {
760 /* zero */
761 if (ecb_expect_true (!x))
762 return s;
763
764 /* handle subnormals */
765
766 /* too small, will be zero */
767 if (e < (14 - 24)) /* might not be sharp, but is good enough */
768 return s;
769
770 m |= 0x00800000; /* make implicit bit explicit */
771
772 /* very tricky - we need to round to the nearest e (+10) bit value */
773 {
774 unsigned int bits = 14 - e;
775 unsigned int half = (1 << (bits - 1)) - 1;
776 unsigned int even = (m >> bits) & 1;
777
778 /* if this overflows, we will end up with a normalised number */
779 m = (m + half + even) >> bits;
780 }
781
782 return s | m;
783 }
784
785 /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
786 m >>= 13;
787
788 return s | 0x7c00 | m | !m;
789}
652 790
653/*******************************************************************************/ 791/*******************************************************************************/
654/* floating point stuff, can be disabled by defining ECB_NO_LIBM */ 792/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
655 793
656/* basically, everything uses "ieee pure-endian" floating point numbers */ 794/* basically, everything uses "ieee pure-endian" floating point numbers */
… …
693 #define ECB_NAN ECB_INFINITY 831 #define ECB_NAN ECB_INFINITY
694 #endif 832 #endif
695 833
696 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L 834 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L
697 #define ecb_ldexpf(x,e) ldexpf ((x), (e)) 835 #define ecb_ldexpf(x,e) ldexpf ((x), (e))
836 #define ecb_frexpf(x,e) frexpf ((x), (e))
698 #else 837 #else
699 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e)) 838 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
839 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
700 #endif 840 #endif
701
702 /* converts an ieee half/binary16 to a float */
703 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
704 ecb_function_ ecb_const float
705 ecb_binary16_to_float (uint16_t x)
706 {
707 int e = (x >> 10) & 0x1f;
708 int m = x & 0x3ff;
709 float r;
710
711 if (!e ) r = ecb_ldexpf (m , -24);
712 else if (e != 31) r = ecb_ldexpf (m + 0x400, e - 25);
713 else if (m ) r = ECB_NAN;
714 else r = ECB_INFINITY;
715
716 return x & 0x8000 ? -r : r;
717 }
718 841
719 /* convert a float to ieee single/binary32 */ 842 /* convert a float to ieee single/binary32 */
720 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x); 843 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
721 ecb_function_ ecb_const uint32_t 844 ecb_function_ ecb_const uint32_t
722 ecb_float_to_binary32 (float x) 845 ecb_float_to_binary32 (float x)
… …
733 if (x == 0e0f ) return 0x00000000U; 856 if (x == 0e0f ) return 0x00000000U;
734 if (x > +3.40282346638528860e+38f) return 0x7f800000U; 857 if (x > +3.40282346638528860e+38f) return 0x7f800000U;
735 if (x < -3.40282346638528860e+38f) return 0xff800000U; 858 if (x < -3.40282346638528860e+38f) return 0xff800000U;
736 if (x != x ) return 0x7fbfffffU; 859 if (x != x ) return 0x7fbfffffU;
737 860
738 m = frexpf (x, &e) * 0x1000000U; 861 m = ecb_frexpf (x, &e) * 0x1000000U;
739 862
740 r = m & 0x80000000U; 863 r = m & 0x80000000U;
741 864
742 if (r) 865 if (r)
743 m = -m; 866 m = -m;
… …
854 #endif 977 #endif
855 978
856 return r; 979 return r;
857 } 980 }
858 981
859#endif 982 /* convert a float to ieee half/binary16 */
983 ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
984 ecb_function_ ecb_const uint16_t
985 ecb_float_to_binary16 (float x)
986 {
987 return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
988 }
860 989
861#endif 990 /* convert an ieee half/binary16 to float */
991 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
992 ecb_function_ ecb_const float
993 ecb_binary16_to_float (uint16_t x)
994 {
995 return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
996 }
862 997
998#endif
999
1000#endif
1001

Diff Legend

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