… | |
… | |
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; |
… | |
… | |
153 | #endif |
153 | #endif |
154 | |
154 | |
155 | /* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/compiler_ref/compiler_builtins.html */ |
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 |
156 | #if __xlC__ && ECB_CPP |
157 | #include <builtins.h> |
157 | #include <builtins.h> |
|
|
158 | #endif |
|
|
159 | |
|
|
160 | #if 1400 <= _MSC_VER |
|
|
161 | #include <intrin.h> /* fence functions _ReadBarrier, also bit search functions _BitScanReverse */ |
158 | #endif |
162 | #endif |
159 | |
163 | |
160 | #ifndef ECB_MEMORY_FENCE |
164 | #ifndef ECB_MEMORY_FENCE |
161 | #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 |
162 | #if __i386 || __i386__ |
166 | #if __i386 || __i386__ |
… | |
… | |
423 | #else |
427 | #else |
424 | ecb_function_ ecb_const int ecb_ctz32 (uint32_t x); |
428 | ecb_function_ ecb_const int ecb_ctz32 (uint32_t x); |
425 | ecb_function_ ecb_const int |
429 | ecb_function_ ecb_const int |
426 | ecb_ctz32 (uint32_t x) |
430 | ecb_ctz32 (uint32_t x) |
427 | { |
431 | { |
|
|
432 | #if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM) |
|
|
433 | unsigned long r = 0; |
|
|
434 | _BitScanForward (&r, x); |
|
|
435 | return (int)r; |
|
|
436 | #else |
428 | int r = 0; |
437 | int r = 0; |
429 | |
438 | |
430 | x &= ~x + 1; /* this isolates the lowest bit */ |
439 | x &= ~x + 1; /* this isolates the lowest bit */ |
431 | |
440 | |
432 | #if ECB_branchless_on_i386 |
441 | #if ECB_branchless_on_i386 |
… | |
… | |
442 | if (x & 0xff00ff00) r += 8; |
451 | if (x & 0xff00ff00) r += 8; |
443 | if (x & 0xffff0000) r += 16; |
452 | if (x & 0xffff0000) r += 16; |
444 | #endif |
453 | #endif |
445 | |
454 | |
446 | return r; |
455 | return r; |
|
|
456 | #endif |
447 | } |
457 | } |
448 | |
458 | |
449 | ecb_function_ ecb_const int ecb_ctz64 (uint64_t x); |
459 | ecb_function_ ecb_const int ecb_ctz64 (uint64_t x); |
450 | ecb_function_ ecb_const int |
460 | ecb_function_ ecb_const int |
451 | ecb_ctz64 (uint64_t x) |
461 | ecb_ctz64 (uint64_t x) |
452 | { |
462 | { |
|
|
463 | #if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM) |
|
|
464 | unsigned long r = 0; |
|
|
465 | _BitScanForward64 (&r, x); |
|
|
466 | return (int)r; |
|
|
467 | #else |
453 | int shift = x & 0xffffffffU ? 0 : 32; |
468 | int shift = x & 0xffffffff ? 0 : 32; |
454 | return ecb_ctz32 (x >> shift) + shift; |
469 | return ecb_ctz32 (x >> shift) + shift; |
|
|
470 | #endif |
455 | } |
471 | } |
456 | |
472 | |
457 | ecb_function_ ecb_const int ecb_popcount32 (uint32_t x); |
473 | ecb_function_ ecb_const int ecb_popcount32 (uint32_t x); |
458 | ecb_function_ ecb_const int |
474 | ecb_function_ ecb_const int |
459 | ecb_popcount32 (uint32_t x) |
475 | ecb_popcount32 (uint32_t x) |
… | |
… | |
467 | } |
483 | } |
468 | |
484 | |
469 | ecb_function_ ecb_const int ecb_ld32 (uint32_t x); |
485 | ecb_function_ ecb_const int ecb_ld32 (uint32_t x); |
470 | ecb_function_ ecb_const int ecb_ld32 (uint32_t x) |
486 | ecb_function_ ecb_const int ecb_ld32 (uint32_t x) |
471 | { |
487 | { |
|
|
488 | #if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM) |
|
|
489 | unsigned long r = 0; |
|
|
490 | _BitScanReverse (&r, x); |
|
|
491 | return (int)r; |
|
|
492 | #else |
472 | int r = 0; |
493 | int r = 0; |
473 | |
494 | |
474 | if (x >> 16) { x >>= 16; r += 16; } |
495 | if (x >> 16) { x >>= 16; r += 16; } |
475 | if (x >> 8) { x >>= 8; r += 8; } |
496 | if (x >> 8) { x >>= 8; r += 8; } |
476 | if (x >> 4) { x >>= 4; r += 4; } |
497 | if (x >> 4) { x >>= 4; r += 4; } |
477 | if (x >> 2) { x >>= 2; r += 2; } |
498 | if (x >> 2) { x >>= 2; r += 2; } |
478 | if (x >> 1) { r += 1; } |
499 | if (x >> 1) { r += 1; } |
479 | |
500 | |
480 | return r; |
501 | return r; |
|
|
502 | #endif |
481 | } |
503 | } |
482 | |
504 | |
483 | ecb_function_ ecb_const int ecb_ld64 (uint64_t x); |
505 | ecb_function_ ecb_const int ecb_ld64 (uint64_t x); |
484 | ecb_function_ ecb_const int ecb_ld64 (uint64_t x) |
506 | ecb_function_ ecb_const int ecb_ld64 (uint64_t x) |
485 | { |
507 | { |
|
|
508 | #if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM) |
|
|
509 | unsigned long r = 0; |
|
|
510 | _BitScanReverse64 (&r, x); |
|
|
511 | return (int)r; |
|
|
512 | #else |
486 | int r = 0; |
513 | int r = 0; |
487 | |
514 | |
488 | if (x >> 32) { x >>= 32; r += 32; } |
515 | if (x >> 32) { x >>= 32; r += 32; } |
489 | |
516 | |
490 | return r + ecb_ld32 (x); |
517 | return r + ecb_ld32 (x); |
|
|
518 | #endif |
491 | } |
519 | } |
492 | #endif |
520 | #endif |
493 | |
521 | |
494 | 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); |
495 | 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)); } |
… | |
… | |
661 | return N; |
689 | return N; |
662 | } |
690 | } |
663 | #else |
691 | #else |
664 | #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) |
692 | #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) |
665 | #endif |
693 | #endif |
|
|
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 | } |
666 | |
790 | |
667 | /*******************************************************************************/ |
791 | /*******************************************************************************/ |
668 | /* floating point stuff, can be disabled by defining ECB_NO_LIBM */ |
792 | /* floating point stuff, can be disabled by defining ECB_NO_LIBM */ |
669 | |
793 | |
670 | /* basically, everything uses "ieee pure-endian" floating point numbers */ |
794 | /* basically, everything uses "ieee pure-endian" floating point numbers */ |
… | |
… | |
713 | #else |
837 | #else |
714 | #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e)) |
838 | #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e)) |
715 | #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e)) |
839 | #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e)) |
716 | #endif |
840 | #endif |
717 | |
841 | |
718 | /* converts an ieee half/binary16 to a float */ |
|
|
719 | ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x); |
|
|
720 | ecb_function_ ecb_const float |
|
|
721 | ecb_binary16_to_float (uint16_t x) |
|
|
722 | { |
|
|
723 | int e = (x >> 10) & 0x1f; |
|
|
724 | int m = x & 0x3ff; |
|
|
725 | float r; |
|
|
726 | |
|
|
727 | if (!e ) r = ecb_ldexpf (m , -24); |
|
|
728 | else if (e != 31) r = ecb_ldexpf (m + 0x400, e - 25); |
|
|
729 | else if (m ) r = ECB_NAN; |
|
|
730 | else r = ECB_INFINITY; |
|
|
731 | |
|
|
732 | return x & 0x8000 ? -r : r; |
|
|
733 | } |
|
|
734 | |
|
|
735 | /* convert a float to ieee single/binary32 */ |
842 | /* convert a float to ieee single/binary32 */ |
736 | 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); |
737 | ecb_function_ ecb_const uint32_t |
844 | ecb_function_ ecb_const uint32_t |
738 | ecb_float_to_binary32 (float x) |
845 | ecb_float_to_binary32 (float x) |
739 | { |
846 | { |
… | |
… | |
870 | #endif |
977 | #endif |
871 | |
978 | |
872 | return r; |
979 | return r; |
873 | } |
980 | } |
874 | |
981 | |
875 | #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 | } |
876 | |
989 | |
877 | #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 | } |
878 | |
997 | |
|
|
998 | #endif |
|
|
999 | |
|
|
1000 | #endif |
|
|
1001 | |