… | |
… | |
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; |
… | |
… | |
74 | #else |
74 | #else |
75 | #define ECB_PTRSIZE 4 |
75 | #define ECB_PTRSIZE 4 |
76 | #endif |
76 | #endif |
77 | #endif |
77 | #endif |
78 | |
78 | |
|
|
79 | #define ECB_GCC_AMD64 (__amd64 || __amd64__ || __x86_64 || __x86_64__) |
|
|
80 | #define ECB_MSVC_AMD64 (_M_AMD64 || _M_X64) |
|
|
81 | |
79 | /* work around x32 idiocy by defining proper macros */ |
82 | /* work around x32 idiocy by defining proper macros */ |
80 | #if __amd64 || __x86_64 || _M_AMD64 || _M_X64 |
83 | #if ECB_GCC_AMD64 || ECB_MSVC_AMD64 |
81 | #if _ILP32 |
84 | #if _ILP32 |
82 | #define ECB_AMD64_X32 1 |
85 | #define ECB_AMD64_X32 1 |
83 | #else |
86 | #else |
84 | #define ECB_AMD64 1 |
87 | #define ECB_AMD64 1 |
85 | #endif |
88 | #endif |
… | |
… | |
147 | |
150 | |
148 | #if ECB_NO_SMP |
151 | #if ECB_NO_SMP |
149 | #define ECB_MEMORY_FENCE do { } while (0) |
152 | #define ECB_MEMORY_FENCE do { } while (0) |
150 | #endif |
153 | #endif |
151 | |
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 */ |
|
|
162 | #endif |
|
|
163 | |
152 | #ifndef ECB_MEMORY_FENCE |
164 | #ifndef ECB_MEMORY_FENCE |
153 | #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 |
154 | #if __i386 || __i386__ |
166 | #if __i386 || __i386__ |
155 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") |
167 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") |
156 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") |
168 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") |
157 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
169 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
158 | #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ |
170 | #elif ECB_GCC_AMD64 |
159 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") |
171 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") |
160 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") |
172 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") |
161 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
173 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
162 | #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ |
174 | #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ |
163 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") |
175 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") |
… | |
… | |
167 | #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ |
179 | #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ |
168 | || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ |
180 | || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ |
169 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") |
181 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") |
170 | #elif __aarch64__ |
182 | #elif __aarch64__ |
171 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory") |
183 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory") |
172 | #elif (__sparc || __sparc__) && !__sparcv8 |
184 | #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8) |
173 | #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") |
174 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") |
186 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") |
175 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") |
187 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") |
176 | #elif defined __s390__ || defined __s390x__ |
188 | #elif defined __s390__ || defined __s390x__ |
177 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") |
189 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") |
… | |
… | |
349 | #define ecb_deprecated __declspec (deprecated) |
361 | #define ecb_deprecated __declspec (deprecated) |
350 | #else |
362 | #else |
351 | #define ecb_deprecated ecb_attribute ((__deprecated__)) |
363 | #define ecb_deprecated ecb_attribute ((__deprecated__)) |
352 | #endif |
364 | #endif |
353 | |
365 | |
354 | #if __MSC_VER >= 1500 |
366 | #if _MSC_VER >= 1500 |
355 | #define ecb_deprecated_message(msg) __declspec (deprecated (msg)) |
367 | #define ecb_deprecated_message(msg) __declspec (deprecated (msg)) |
356 | #elif ECB_GCC_VERSION(4,5) |
368 | #elif ECB_GCC_VERSION(4,5) |
357 | #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg)) |
369 | #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg)) |
358 | #else |
370 | #else |
359 | #define ecb_deprecated_message(msg) ecb_deprecated |
371 | #define ecb_deprecated_message(msg) ecb_deprecated |
… | |
… | |
368 | #define ecb_unused ecb_attribute ((__unused__)) |
380 | #define ecb_unused ecb_attribute ((__unused__)) |
369 | #define ecb_const ecb_attribute ((__const__)) |
381 | #define ecb_const ecb_attribute ((__const__)) |
370 | #define ecb_pure ecb_attribute ((__pure__)) |
382 | #define ecb_pure ecb_attribute ((__pure__)) |
371 | |
383 | |
372 | #if ECB_C11 || __IBMC_NORETURN |
384 | #if ECB_C11 || __IBMC_NORETURN |
373 | /* 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 */ |
374 | #define ecb_noreturn _Noreturn |
386 | #define ecb_noreturn _Noreturn |
375 | #elif ECB_CPP11 |
387 | #elif ECB_CPP11 |
376 | #define ecb_noreturn [[noreturn]] |
388 | #define ecb_noreturn [[noreturn]] |
377 | #elif _MSC_VER >= 1200 |
389 | #elif _MSC_VER >= 1200 |
378 | /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */ |
390 | /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */ |
… | |
… | |
415 | #else |
427 | #else |
416 | ecb_function_ ecb_const int ecb_ctz32 (uint32_t x); |
428 | ecb_function_ ecb_const int ecb_ctz32 (uint32_t x); |
417 | ecb_function_ ecb_const int |
429 | ecb_function_ ecb_const int |
418 | ecb_ctz32 (uint32_t x) |
430 | ecb_ctz32 (uint32_t x) |
419 | { |
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 |
420 | int r = 0; |
437 | int r = 0; |
421 | |
438 | |
422 | x &= ~x + 1; /* this isolates the lowest bit */ |
439 | x &= ~x + 1; /* this isolates the lowest bit */ |
423 | |
440 | |
424 | #if ECB_branchless_on_i386 |
441 | #if ECB_branchless_on_i386 |
… | |
… | |
434 | if (x & 0xff00ff00) r += 8; |
451 | if (x & 0xff00ff00) r += 8; |
435 | if (x & 0xffff0000) r += 16; |
452 | if (x & 0xffff0000) r += 16; |
436 | #endif |
453 | #endif |
437 | |
454 | |
438 | return r; |
455 | return r; |
|
|
456 | #endif |
439 | } |
457 | } |
440 | |
458 | |
441 | ecb_function_ ecb_const int ecb_ctz64 (uint64_t x); |
459 | ecb_function_ ecb_const int ecb_ctz64 (uint64_t x); |
442 | ecb_function_ ecb_const int |
460 | ecb_function_ ecb_const int |
443 | ecb_ctz64 (uint64_t x) |
461 | ecb_ctz64 (uint64_t x) |
444 | { |
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 |
445 | int shift = x & 0xffffffffU ? 0 : 32; |
468 | int shift = x & 0xffffffff ? 0 : 32; |
446 | return ecb_ctz32 (x >> shift) + shift; |
469 | return ecb_ctz32 (x >> shift) + shift; |
|
|
470 | #endif |
447 | } |
471 | } |
448 | |
472 | |
449 | ecb_function_ ecb_const int ecb_popcount32 (uint32_t x); |
473 | ecb_function_ ecb_const int ecb_popcount32 (uint32_t x); |
450 | ecb_function_ ecb_const int |
474 | ecb_function_ ecb_const int |
451 | ecb_popcount32 (uint32_t x) |
475 | ecb_popcount32 (uint32_t x) |
… | |
… | |
459 | } |
483 | } |
460 | |
484 | |
461 | ecb_function_ ecb_const int ecb_ld32 (uint32_t x); |
485 | ecb_function_ ecb_const int ecb_ld32 (uint32_t x); |
462 | ecb_function_ ecb_const int ecb_ld32 (uint32_t x) |
486 | ecb_function_ ecb_const int ecb_ld32 (uint32_t x) |
463 | { |
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 |
464 | int r = 0; |
493 | int r = 0; |
465 | |
494 | |
466 | if (x >> 16) { x >>= 16; r += 16; } |
495 | if (x >> 16) { x >>= 16; r += 16; } |
467 | if (x >> 8) { x >>= 8; r += 8; } |
496 | if (x >> 8) { x >>= 8; r += 8; } |
468 | if (x >> 4) { x >>= 4; r += 4; } |
497 | if (x >> 4) { x >>= 4; r += 4; } |
469 | if (x >> 2) { x >>= 2; r += 2; } |
498 | if (x >> 2) { x >>= 2; r += 2; } |
470 | if (x >> 1) { r += 1; } |
499 | if (x >> 1) { r += 1; } |
471 | |
500 | |
472 | return r; |
501 | return r; |
|
|
502 | #endif |
473 | } |
503 | } |
474 | |
504 | |
475 | ecb_function_ ecb_const int ecb_ld64 (uint64_t x); |
505 | ecb_function_ ecb_const int ecb_ld64 (uint64_t x); |
476 | ecb_function_ ecb_const int ecb_ld64 (uint64_t x) |
506 | ecb_function_ ecb_const int ecb_ld64 (uint64_t x) |
477 | { |
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 |
478 | int r = 0; |
513 | int r = 0; |
479 | |
514 | |
480 | if (x >> 32) { x >>= 32; r += 32; } |
515 | if (x >> 32) { x >>= 32; r += 32; } |
481 | |
516 | |
482 | return r + ecb_ld32 (x); |
517 | return r + ecb_ld32 (x); |
|
|
518 | #endif |
483 | } |
519 | } |
484 | #endif |
520 | #endif |
485 | |
521 | |
486 | ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x); |
522 | ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x); |
487 | ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } |
523 | ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } |
… | |
… | |
544 | ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } |
580 | ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } |
545 | ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } |
581 | ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } |
546 | ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } |
582 | ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } |
547 | |
583 | |
548 | #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 |
549 | #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) |
588 | #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) |
|
|
589 | #endif |
550 | #define ecb_bswap32(x) __builtin_bswap32 (x) |
590 | #define ecb_bswap32(x) __builtin_bswap32 (x) |
551 | #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))) |
552 | #else |
597 | #else |
553 | ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x); |
598 | ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x); |
554 | ecb_function_ ecb_const uint16_t |
599 | ecb_function_ ecb_const uint16_t |
555 | ecb_bswap16 (uint16_t x) |
600 | ecb_bswap16 (uint16_t x) |
556 | { |
601 | { |
… | |
… | |
591 | /* but less than using pointers, and always seems to */ |
636 | /* but less than using pointers, and always seems to */ |
592 | /* successfully return a constant. */ |
637 | /* successfully return a constant. */ |
593 | /* the reason why we have this horrible preprocessor mess */ |
638 | /* the reason why we have this horrible preprocessor mess */ |
594 | /* 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 */ |
595 | /* or when using a recent enough gcc version (>= 4.6) */ |
640 | /* or when using a recent enough gcc version (>= 4.6) */ |
596 | #if ((__i386 || __i386__) && !__VOS__) || _M_X86 || __amd64 || __amd64__ || _M_X64 |
641 | #if ((__i386 || __i386__) && !__VOS__) || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64 |
597 | return 0x44; |
642 | return 0x44; |
598 | #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ |
643 | #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ |
599 | return 0x44; |
644 | return 0x44; |
600 | #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ |
645 | #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ |
601 | return 0x11; |
646 | return 0x11; |
… | |
… | |
645 | } |
690 | } |
646 | #else |
691 | #else |
647 | #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) |
692 | #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) |
648 | #endif |
693 | #endif |
649 | |
694 | |
|
|
695 | ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x); |
|
|
696 | ecb_function_ ecb_const uint32_t |
|
|
697 | ecb_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 | |
|
|
727 | ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x); |
|
|
728 | ecb_function_ ecb_const uint16_t |
|
|
729 | ecb_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 | } |
|
|
790 | |
650 | /*******************************************************************************/ |
791 | /*******************************************************************************/ |
651 | /* floating point stuff, can be disabled by defining ECB_NO_LIBM */ |
792 | /* floating point stuff, can be disabled by defining ECB_NO_LIBM */ |
652 | |
793 | |
653 | /* basically, everything uses "ieee pure-endian" floating point numbers */ |
794 | /* basically, everything uses "ieee pure-endian" floating point numbers */ |
654 | /* the only noteworthy exception is ancient armle, which uses order 43218765 */ |
795 | /* the only noteworthy exception is ancient armle, which uses order 43218765 */ |
655 | #if 0 \ |
796 | #if 0 \ |
656 | || __i386 || __i386__ \ |
797 | || __i386 || __i386__ \ |
657 | || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ |
798 | || ECB_GCC_AMD64 \ |
658 | || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ |
799 | || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ |
659 | || defined __s390__ || defined __s390x__ \ |
800 | || defined __s390__ || defined __s390x__ \ |
660 | || defined __mips__ \ |
801 | || defined __mips__ \ |
661 | || defined __alpha__ \ |
802 | || defined __alpha__ \ |
662 | || defined __hppa__ \ |
803 | || defined __hppa__ \ |
663 | || defined __ia64__ \ |
804 | || defined __ia64__ \ |
664 | || defined __m68k__ \ |
805 | || defined __m68k__ \ |
665 | || defined __m88k__ \ |
806 | || defined __m88k__ \ |
666 | || defined __sh__ \ |
807 | || defined __sh__ \ |
667 | || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 \ |
808 | || defined _M_IX86 || defined ECB_MSVC_AMD64 || defined _M_IA64 \ |
668 | || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \ |
809 | || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \ |
669 | || defined __aarch64__ |
810 | || defined __aarch64__ |
670 | #define ECB_STDFP 1 |
811 | #define ECB_STDFP 1 |
671 | #include <string.h> /* for memcpy */ |
812 | #include <string.h> /* for memcpy */ |
672 | #else |
813 | #else |
… | |
… | |
690 | #define ECB_NAN ECB_INFINITY |
831 | #define ECB_NAN ECB_INFINITY |
691 | #endif |
832 | #endif |
692 | |
833 | |
693 | #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L |
834 | #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L |
694 | #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)) |
695 | #else |
837 | #else |
696 | #define ecb_ldexpf(x,e) (float) ldexp ((x), (e)) |
838 | #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e)) |
|
|
839 | #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e)) |
697 | #endif |
840 | #endif |
698 | |
|
|
699 | /* converts an ieee half/binary16 to a float */ |
|
|
700 | ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x); |
|
|
701 | ecb_function_ ecb_const float |
|
|
702 | ecb_binary16_to_float (uint16_t x) |
|
|
703 | { |
|
|
704 | int e = (x >> 10) & 0x1f; |
|
|
705 | int m = x & 0x3ff; |
|
|
706 | float r; |
|
|
707 | |
|
|
708 | if (!e ) r = ecb_ldexpf (m , -24); |
|
|
709 | else if (e != 31) r = ecb_ldexpf (m + 0x400, e - 25); |
|
|
710 | else if (m ) r = ECB_NAN; |
|
|
711 | else r = ECB_INFINITY; |
|
|
712 | |
|
|
713 | return x & 0x8000 ? -r : r; |
|
|
714 | } |
|
|
715 | |
841 | |
716 | /* convert a float to ieee single/binary32 */ |
842 | /* convert a float to ieee single/binary32 */ |
717 | 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); |
718 | ecb_function_ ecb_const uint32_t |
844 | ecb_function_ ecb_const uint32_t |
719 | ecb_float_to_binary32 (float x) |
845 | ecb_float_to_binary32 (float x) |
… | |
… | |
730 | if (x == 0e0f ) return 0x00000000U; |
856 | if (x == 0e0f ) return 0x00000000U; |
731 | if (x > +3.40282346638528860e+38f) return 0x7f800000U; |
857 | if (x > +3.40282346638528860e+38f) return 0x7f800000U; |
732 | if (x < -3.40282346638528860e+38f) return 0xff800000U; |
858 | if (x < -3.40282346638528860e+38f) return 0xff800000U; |
733 | if (x != x ) return 0x7fbfffffU; |
859 | if (x != x ) return 0x7fbfffffU; |
734 | |
860 | |
735 | m = frexpf (x, &e) * 0x1000000U; |
861 | m = ecb_frexpf (x, &e) * 0x1000000U; |
736 | |
862 | |
737 | r = m & 0x80000000U; |
863 | r = m & 0x80000000U; |
738 | |
864 | |
739 | if (r) |
865 | if (r) |
740 | m = -m; |
866 | m = -m; |
… | |
… | |
851 | #endif |
977 | #endif |
852 | |
978 | |
853 | return r; |
979 | return r; |
854 | } |
980 | } |
855 | |
981 | |
856 | #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 | } |
857 | |
989 | |
858 | #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 | } |
859 | |
997 | |
|
|
998 | #endif |
|
|
999 | |
|
|
1000 | #endif |
|
|
1001 | |