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

Comparing libecb/ecb.h (file contents):
Revision 1.167 by root, Sat Nov 21 16:53:50 2015 UTC vs.
Revision 1.172 by root, Tue Nov 24 22:12:00 2015 UTC

… …
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;
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;
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;
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;
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
494ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x); 522ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x);
495ecb_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)); }
… …
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
666 694
667ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint16_t x); 695ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
668ecb_function_ ecb_const uint32_t 696ecb_function_ ecb_const uint32_t
669ecb_binary16_to_binary32 (uint16_t x) 697ecb_binary16_to_binary32 (uint32_t x)
670{ 698{
671 unsigned int s = (x & 0x8000) << (31 - 15); 699 unsigned int s = (x & 0x8000) << (31 - 15);
672 int e = (x >> 10) & 0x001f; 700 int e = (x >> 10) & 0x001f;
673 unsigned int m = x & 0x03ff; 701 unsigned int m = x & 0x03ff;
674 702
… …
724 752
725 /* handle large numbers and infinity */ 753 /* handle large numbers and infinity */
726 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000)) 754 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
727 return s | 0x7c00; 755 return s | 0x7c00;
728 756
729 /* handle zero and subnormals */ 757 /* handle zero, subnormals and small numbers */
730 if (ecb_expect_true (x < 0x38800000)) 758 if (ecb_expect_true (x < 0x38800000))
731 { 759 {
732 /* zero */ 760 /* zero */
733 if (ecb_expect_true (!x)) 761 if (ecb_expect_true (!x))
734 return s; 762 return s;
735 763
736 /* handle subnormals */ 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;
737 769
738 m |= 0x00800000; /* make implicit bit explicit */ 770 m |= 0x00800000; /* make implicit bit explicit */
739 771
740 /* very tricky - we need to round to the nearest e (+10) bit value */ 772 /* very tricky - we need to round to the nearest e (+10) bit value */
741 { 773 {

Diff Legend

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