… | |
… | |
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; |
… | |
… | |
448 | |
448 | |
449 | ecb_function_ ecb_const int ecb_ctz64 (uint64_t x); |
449 | ecb_function_ ecb_const int ecb_ctz64 (uint64_t x); |
450 | ecb_function_ ecb_const int |
450 | ecb_function_ ecb_const int |
451 | ecb_ctz64 (uint64_t x) |
451 | ecb_ctz64 (uint64_t x) |
452 | { |
452 | { |
453 | int shift = x & 0xffffffffU ? 0 : 32; |
453 | int shift = x & 0xffffffff ? 0 : 32; |
454 | return ecb_ctz32 (x >> shift) + shift; |
454 | return ecb_ctz32 (x >> shift) + shift; |
455 | } |
455 | } |
456 | |
456 | |
457 | ecb_function_ ecb_const int ecb_popcount32 (uint32_t x); |
457 | ecb_function_ ecb_const int ecb_popcount32 (uint32_t x); |
458 | ecb_function_ ecb_const int |
458 | ecb_function_ ecb_const int |
… | |
… | |
661 | return N; |
661 | return N; |
662 | } |
662 | } |
663 | #else |
663 | #else |
664 | #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) |
664 | #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) |
665 | #endif |
665 | #endif |
|
|
666 | |
|
|
667 | ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint16_t x); |
|
|
668 | ecb_function_ ecb_const uint32_t |
|
|
669 | ecb_binary16_to_binary32 (uint16_t x) |
|
|
670 | { |
|
|
671 | unsigned int s = (x & 0x8000) << (31 - 15); |
|
|
672 | int e = (x >> 10) & 0x001f; |
|
|
673 | unsigned int m = x & 0x03ff; |
|
|
674 | |
|
|
675 | if (ecb_expect_false (e == 31)) |
|
|
676 | /* infinity or NaN */ |
|
|
677 | e = 255 - (127 - 15); |
|
|
678 | else if (ecb_expect_false (!e)) |
|
|
679 | { |
|
|
680 | if (ecb_expect_true (!m)) |
|
|
681 | /* zero, handled by code below by forcing e to 0 */ |
|
|
682 | e = 0 - (127 - 15); |
|
|
683 | else |
|
|
684 | { |
|
|
685 | /* subnormal, renormalise */ |
|
|
686 | unsigned int s = 10 - ecb_ld32 (m); |
|
|
687 | |
|
|
688 | m = (m << s) & 0x3ff; /* mask implicit bit */ |
|
|
689 | e -= s - 1; |
|
|
690 | } |
|
|
691 | } |
|
|
692 | |
|
|
693 | /* e and m now are normalised, or zero, (or inf or nan) */ |
|
|
694 | e += 127 - 15; |
|
|
695 | |
|
|
696 | return s | (e << 23) | (m << (23 - 10)); |
|
|
697 | } |
|
|
698 | |
|
|
699 | ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x); |
|
|
700 | ecb_function_ ecb_const uint16_t |
|
|
701 | ecb_binary32_to_binary16 (uint32_t x) |
|
|
702 | { |
|
|
703 | unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */ |
|
|
704 | unsigned int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */ |
|
|
705 | unsigned int m = x & 0x007fffff; |
|
|
706 | |
|
|
707 | x &= 0x7fffffff; |
|
|
708 | |
|
|
709 | /* if it's within range of binary16 normals, use fast path */ |
|
|
710 | if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff)) |
|
|
711 | { |
|
|
712 | /* mantissa round-to-even */ |
|
|
713 | m += 0x00000fff + ((m >> (23 - 10)) & 1); |
|
|
714 | |
|
|
715 | /* handle overflow */ |
|
|
716 | if (ecb_expect_false (m >= 0x00800000)) |
|
|
717 | { |
|
|
718 | m >>= 1; |
|
|
719 | e += 1; |
|
|
720 | } |
|
|
721 | |
|
|
722 | return s | (e << 10) | (m >> (23 - 10)); |
|
|
723 | } |
|
|
724 | |
|
|
725 | /* handle large numbers and infinity */ |
|
|
726 | if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000)) |
|
|
727 | return s | 0x7c00; |
|
|
728 | |
|
|
729 | /* handle zero and subnormals */ |
|
|
730 | if (ecb_expect_true (x < 0x38800000)) |
|
|
731 | { |
|
|
732 | /* zero */ |
|
|
733 | if (ecb_expect_true (!x)) |
|
|
734 | return s; |
|
|
735 | |
|
|
736 | /* handle subnormals */ |
|
|
737 | |
|
|
738 | m |= 0x00800000; /* make implicit bit explicit */ |
|
|
739 | |
|
|
740 | /* very tricky - we need to round to the nearest e (+10) bit value */ |
|
|
741 | { |
|
|
742 | unsigned int bits = 14 - e; |
|
|
743 | unsigned int half = (1 << (bits - 1)) - 1; |
|
|
744 | unsigned int even = (m >> bits) & 1; |
|
|
745 | |
|
|
746 | /* if this overflows, we will end up with a normalised number */ |
|
|
747 | m = (m + half + even) >> bits; |
|
|
748 | } |
|
|
749 | |
|
|
750 | return s | m; |
|
|
751 | } |
|
|
752 | |
|
|
753 | /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */ |
|
|
754 | m >>= 13; |
|
|
755 | |
|
|
756 | return s | 0x7c00 | m | !m; |
|
|
757 | } |
666 | |
758 | |
667 | /*******************************************************************************/ |
759 | /*******************************************************************************/ |
668 | /* floating point stuff, can be disabled by defining ECB_NO_LIBM */ |
760 | /* floating point stuff, can be disabled by defining ECB_NO_LIBM */ |
669 | |
761 | |
670 | /* basically, everything uses "ieee pure-endian" floating point numbers */ |
762 | /* basically, everything uses "ieee pure-endian" floating point numbers */ |
… | |
… | |
713 | #else |
805 | #else |
714 | #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e)) |
806 | #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e)) |
715 | #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e)) |
807 | #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e)) |
716 | #endif |
808 | #endif |
717 | |
809 | |
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 */ |
810 | /* convert a float to ieee single/binary32 */ |
736 | ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x); |
811 | ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x); |
737 | ecb_function_ ecb_const uint32_t |
812 | ecb_function_ ecb_const uint32_t |
738 | ecb_float_to_binary32 (float x) |
813 | ecb_float_to_binary32 (float x) |
739 | { |
814 | { |
… | |
… | |
870 | #endif |
945 | #endif |
871 | |
946 | |
872 | return r; |
947 | return r; |
873 | } |
948 | } |
874 | |
949 | |
875 | #endif |
950 | /* convert a float to ieee half/binary16 */ |
|
|
951 | ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x); |
|
|
952 | ecb_function_ ecb_const uint16_t |
|
|
953 | ecb_float_to_binary16 (float x) |
|
|
954 | { |
|
|
955 | return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x)); |
|
|
956 | } |
876 | |
957 | |
877 | #endif |
958 | /* convert an ieee half/binary16 to float */ |
|
|
959 | ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x); |
|
|
960 | ecb_function_ ecb_const float |
|
|
961 | ecb_binary16_to_float (uint16_t x) |
|
|
962 | { |
|
|
963 | return ecb_binary32_to_float (ecb_binary16_to_binary32 (x)); |
|
|
964 | } |
878 | |
965 | |
|
|
966 | #endif |
|
|
967 | |
|
|
968 | #endif |
|
|
969 | |