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

Comparing libecb/ecb.h (file contents):
Revision 1.164 by root, Mon Apr 20 20:06:30 2015 UTC vs.
Revision 1.174 by root, Wed Nov 25 02:36:53 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")
… …
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)); }
… …
593#endif 626#endif
594 627
595/* try to tell the compiler that some condition is definitely true */ 628/* try to tell the compiler that some condition is definitely true */
596#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 629#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
597 630
598ecb_inline ecb_const unsigned char ecb_byteorder_helper (void); 631ecb_inline ecb_const uint32_t ecb_byteorder_helper (void);
599ecb_inline ecb_const unsigned char 632ecb_inline ecb_const uint32_t
600ecb_byteorder_helper (void) 633ecb_byteorder_helper (void)
601{ 634{
602 /* the union code still generates code under pressure in gcc, */ 635 /* the union code still generates code under pressure in gcc, */
603 /* but less than using pointers, and always seems to */ 636 /* but less than using pointers, and always seems to */
604 /* successfully return a constant. */ 637 /* successfully return a constant. */
605 /* the reason why we have this horrible preprocessor mess */ 638 /* the reason why we have this horrible preprocessor mess */
606 /* is to avoid it in all cases, at least on common architectures */ 639 /* is to avoid it in all cases, at least on common architectures */
607 /* or when using a recent enough gcc version (>= 4.6) */ 640 /* or when using a recent enough gcc version (>= 4.6) */
608#if ((__i386 || __i386__) && !__VOS__) || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64
609 return 0x44;
610#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ 641#if (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
642 || ((__i386 || __i386__ || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64) && !__VOS__)
643 #define ECB_LITTLE_ENDIAN 1
611 return 0x44; 644 return 0x44332211;
612#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 645#elif (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) \
646 || ((__AARCH64EB__ || __MIPSEB__ || __ARMEB__) && !__VOS__)
647 #define ECB_BIG_ENDIAN 1
613 return 0x11; 648 return 0x11223344;
614#else 649#else
615 union 650 union
616 { 651 {
652 uint8_t c[4];
617 uint32_t i; 653 uint32_t u;
618 uint8_t c;
619 } u = { 0x11223344 }; 654 } u = { 0x11, 0x22, 0x33, 0x44 };
620 return u.c; 655 return u.u;
621#endif 656#endif
622} 657}
623 658
624ecb_inline ecb_const ecb_bool ecb_big_endian (void); 659ecb_inline ecb_const ecb_bool ecb_big_endian (void);
625ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } 660ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11223344; }
626ecb_inline ecb_const ecb_bool ecb_little_endian (void); 661ecb_inline ecb_const ecb_bool ecb_little_endian (void);
627ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } 662ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44332211; }
628 663
629#if ECB_GCC_VERSION(3,0) || ECB_C99 664#if ECB_GCC_VERSION(3,0) || ECB_C99
630 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) 665 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
631#else 666#else
632 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) 667 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n)))
… …
656 return N; 691 return N;
657 } 692 }
658#else 693#else
659 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) 694 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
660#endif 695#endif
696
697ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
698ecb_function_ ecb_const uint32_t
699ecb_binary16_to_binary32 (uint32_t x)
700{
701 unsigned int s = (x & 0x8000) << (31 - 15);
702 int e = (x >> 10) & 0x001f;
703 unsigned int m = x & 0x03ff;
704
705 if (ecb_expect_false (e == 31))
706 /* infinity or NaN */
707 e = 255 - (127 - 15);
708 else if (ecb_expect_false (!e))
709 {
710 if (ecb_expect_true (!m))
711 /* zero, handled by code below by forcing e to 0 */
712 e = 0 - (127 - 15);
713 else
714 {
715 /* subnormal, renormalise */
716 unsigned int s = 10 - ecb_ld32 (m);
717
718 m = (m << s) & 0x3ff; /* mask implicit bit */
719 e -= s - 1;
720 }
721 }
722
723 /* e and m now are normalised, or zero, (or inf or nan) */
724 e += 127 - 15;
725
726 return s | (e << 23) | (m << (23 - 10));
727}
728
729ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
730ecb_function_ ecb_const uint16_t
731ecb_binary32_to_binary16 (uint32_t x)
732{
733 unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
734 unsigned int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
735 unsigned int m = x & 0x007fffff;
736
737 x &= 0x7fffffff;
738
739 /* if it's within range of binary16 normals, use fast path */
740 if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
741 {
742 /* mantissa round-to-even */
743 m += 0x00000fff + ((m >> (23 - 10)) & 1);
744
745 /* handle overflow */
746 if (ecb_expect_false (m >= 0x00800000))
747 {
748 m >>= 1;
749 e += 1;
750 }
751
752 return s | (e << 10) | (m >> (23 - 10));
753 }
754
755 /* handle large numbers and infinity */
756 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
757 return s | 0x7c00;
758
759 /* handle zero, subnormals and small numbers */
760 if (ecb_expect_true (x < 0x38800000))
761 {
762 /* zero */
763 if (ecb_expect_true (!x))
764 return s;
765
766 /* handle subnormals */
767
768 /* too small, will be zero */
769 if (e < (14 - 24)) /* might not be sharp, but is good enough */
770 return s;
771
772 m |= 0x00800000; /* make implicit bit explicit */
773
774 /* very tricky - we need to round to the nearest e (+10) bit value */
775 {
776 unsigned int bits = 14 - e;
777 unsigned int half = (1 << (bits - 1)) - 1;
778 unsigned int even = (m >> bits) & 1;
779
780 /* if this overflows, we will end up with a normalised number */
781 m = (m + half + even) >> bits;
782 }
783
784 return s | m;
785 }
786
787 /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
788 m >>= 13;
789
790 return s | 0x7c00 | m | !m;
791}
661 792
662/*******************************************************************************/ 793/*******************************************************************************/
663/* floating point stuff, can be disabled by defining ECB_NO_LIBM */ 794/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
664 795
665/* basically, everything uses "ieee pure-endian" floating point numbers */ 796/* basically, everything uses "ieee pure-endian" floating point numbers */
… …
708 #else 839 #else
709 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e)) 840 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
710 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e)) 841 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
711 #endif 842 #endif
712 843
713 /* converts an ieee half/binary16 to a float */
714 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
715 ecb_function_ ecb_const float
716 ecb_binary16_to_float (uint16_t x)
717 {
718 int e = (x >> 10) & 0x1f;
719 int m = x & 0x3ff;
720 float r;
721
722 if (!e ) r = ecb_ldexpf (m , -24);
723 else if (e != 31) r = ecb_ldexpf (m + 0x400, e - 25);
724 else if (m ) r = ECB_NAN;
725 else r = ECB_INFINITY;
726
727 return x & 0x8000 ? -r : r;
728 }
729
730 /* convert a float to ieee single/binary32 */ 844 /* convert a float to ieee single/binary32 */
731 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x); 845 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
732 ecb_function_ ecb_const uint32_t 846 ecb_function_ ecb_const uint32_t
733 ecb_float_to_binary32 (float x) 847 ecb_float_to_binary32 (float x)
734 { 848 {
… …
865 #endif 979 #endif
866 980
867 return r; 981 return r;
868 } 982 }
869 983
870#endif 984 /* convert a float to ieee half/binary16 */
985 ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
986 ecb_function_ ecb_const uint16_t
987 ecb_float_to_binary16 (float x)
988 {
989 return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
990 }
871 991
872#endif 992 /* convert an ieee half/binary16 to float */
993 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
994 ecb_function_ ecb_const float
995 ecb_binary16_to_float (uint16_t x)
996 {
997 return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
998 }
873 999
1000#endif
1001
1002#endif
1003

Diff Legend

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