#define ECB_H
/* 16 bits major, 16 bits minor */
-#define ECB_VERSION 0x00010004
+#define ECB_VERSION 0x00010005
#ifdef _WIN32
typedef signed char int8_t;
#endif
#else
#include <inttypes.h>
- #if UINTMAX_MAX > 0xffffffffU
+ #if (defined INTPTR_MAX ? INTPTR_MAX : ULONG_MAX) > 0xffffffffU
#define ECB_PTRSIZE 8
#else
#define ECB_PTRSIZE 4
#include <builtins.h>
#endif
+#if 1400 <= _MSC_VER
+ #include <intrin.h> /* fence functions _ReadBarrier, also bit search functions _BitScanReverse */
+#endif
+
#ifndef ECB_MEMORY_FENCE
#if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
#if __i386 || __i386__
#define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("")
#elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__
#define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory")
+ #elif defined __ARM_ARCH_2__ \
+ || defined __ARM_ARCH_3__ || defined __ARM_ARCH_3M__ \
+ || defined __ARM_ARCH_4__ || defined __ARM_ARCH_4T__ \
+ || defined __ARM_ARCH_5__ || defined __ARM_ARCH_5E__ \
+ || defined __ARM_ARCH_5T__ || defined __ARM_ARCH_5TE__ \
+ || defined __ARM_ARCH_5TEJ__
+ /* should not need any, unless running old code on newer cpu - arm doesn't support that */
#elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \
- || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__
+ || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ \
+ || defined __ARM_ARCH_6T2__
#define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory")
#elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
- || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__
+ || defined __ARM_ARCH_7R__ || defined __ARM_ARCH_7M__
#define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
#elif __aarch64__
#define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
ecb_function_ ecb_const int
ecb_ctz32 (uint32_t x)
{
+#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
+ unsigned long r;
+ _BitScanForward (&r, x);
+ return (int)r;
+#else
int r = 0;
x &= ~x + 1; /* this isolates the lowest bit */
#endif
return r;
+#endif
}
ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
ecb_function_ ecb_const int
ecb_ctz64 (uint64_t x)
{
- int shift = x & 0xffffffffU ? 0 : 32;
+#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
+ unsigned long r;
+ _BitScanForward64 (&r, x);
+ return (int)r;
+#else
+ int shift = x & 0xffffffff ? 0 : 32;
return ecb_ctz32 (x >> shift) + shift;
+#endif
}
ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
{
+#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
+ unsigned long r;
+ _BitScanReverse (&r, x);
+ return (int)r;
+#else
int r = 0;
if (x >> 16) { x >>= 16; r += 16; }
if (x >> 1) { r += 1; }
return r;
+#endif
}
ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
{
+#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
+ unsigned long r;
+ _BitScanReverse64 (&r, x);
+ return (int)r;
+#else
int r = 0;
if (x >> 32) { x >>= 32; r += 32; }
return r + ecb_ld32 (x);
+#endif
}
#endif
/* try to tell the compiler that some condition is definitely true */
#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
-ecb_inline ecb_const unsigned char ecb_byteorder_helper (void);
-ecb_inline ecb_const unsigned char
+ecb_inline ecb_const uint32_t ecb_byteorder_helper (void);
+ecb_inline ecb_const uint32_t
ecb_byteorder_helper (void)
{
/* the union code still generates code under pressure in gcc, */
/* the reason why we have this horrible preprocessor mess */
/* is to avoid it in all cases, at least on common architectures */
/* or when using a recent enough gcc version (>= 4.6) */
-#if ((__i386 || __i386__) && !__VOS__) || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64
- return 0x44;
-#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
- return 0x44;
-#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
- return 0x11;
+#if (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
+ || ((__i386 || __i386__ || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64) && !__VOS__)
+ #define ECB_LITTLE_ENDIAN 1
+ return 0x44332211;
+#elif (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) \
+ || ((__AARCH64EB__ || __MIPSEB__ || __ARMEB__) && !__VOS__)
+ #define ECB_BIG_ENDIAN 1
+ return 0x11223344;
#else
union
{
- uint32_t i;
- uint8_t c;
- } u = { 0x11223344 };
- return u.c;
+ uint8_t c[4];
+ uint32_t u;
+ } u = { 0x11, 0x22, 0x33, 0x44 };
+ return u.u;
#endif
}
ecb_inline ecb_const ecb_bool ecb_big_endian (void);
-ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; }
+ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11223344; }
ecb_inline ecb_const ecb_bool ecb_little_endian (void);
-ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; }
+ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44332211; }
#if ECB_GCC_VERSION(3,0) || ECB_C99
#define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
#define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
#endif
+ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
+ecb_function_ ecb_const uint32_t
+ecb_binary16_to_binary32 (uint32_t x)
+{
+ unsigned int s = (x & 0x8000) << (31 - 15);
+ int e = (x >> 10) & 0x001f;
+ unsigned int m = x & 0x03ff;
+
+ if (ecb_expect_false (e == 31))
+ /* infinity or NaN */
+ e = 255 - (127 - 15);
+ else if (ecb_expect_false (!e))
+ {
+ if (ecb_expect_true (!m))
+ /* zero, handled by code below by forcing e to 0 */
+ e = 0 - (127 - 15);
+ else
+ {
+ /* subnormal, renormalise */
+ unsigned int s = 10 - ecb_ld32 (m);
+
+ m = (m << s) & 0x3ff; /* mask implicit bit */
+ e -= s - 1;
+ }
+ }
+
+ /* e and m now are normalised, or zero, (or inf or nan) */
+ e += 127 - 15;
+
+ return s | (e << 23) | (m << (23 - 10));
+}
+
+ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
+ecb_function_ ecb_const uint16_t
+ecb_binary32_to_binary16 (uint32_t x)
+{
+ unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
+ unsigned int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
+ unsigned int m = x & 0x007fffff;
+
+ x &= 0x7fffffff;
+
+ /* if it's within range of binary16 normals, use fast path */
+ if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
+ {
+ /* mantissa round-to-even */
+ m += 0x00000fff + ((m >> (23 - 10)) & 1);
+
+ /* handle overflow */
+ if (ecb_expect_false (m >= 0x00800000))
+ {
+ m >>= 1;
+ e += 1;
+ }
+
+ return s | (e << 10) | (m >> (23 - 10));
+ }
+
+ /* handle large numbers and infinity */
+ if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
+ return s | 0x7c00;
+
+ /* handle zero, subnormals and small numbers */
+ if (ecb_expect_true (x < 0x38800000))
+ {
+ /* zero */
+ if (ecb_expect_true (!x))
+ return s;
+
+ /* handle subnormals */
+
+ /* too small, will be zero */
+ if (e < (14 - 24)) /* might not be sharp, but is good enough */
+ return s;
+
+ m |= 0x00800000; /* make implicit bit explicit */
+
+ /* very tricky - we need to round to the nearest e (+10) bit value */
+ {
+ unsigned int bits = 14 - e;
+ unsigned int half = (1 << (bits - 1)) - 1;
+ unsigned int even = (m >> bits) & 1;
+
+ /* if this overflows, we will end up with a normalised number */
+ m = (m + half + even) >> bits;
+ }
+
+ return s | m;
+ }
+
+ /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
+ m >>= 13;
+
+ return s | 0x7c00 | m | !m;
+}
+
/*******************************************************************************/
/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
#define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
#endif
- /* converts an ieee half/binary16 to a float */
- ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
- ecb_function_ ecb_const float
- ecb_binary16_to_float (uint16_t x)
- {
- int e = (x >> 10) & 0x1f;
- int m = x & 0x3ff;
- float r;
-
- if (!e ) r = ecb_ldexpf (m , -24);
- else if (e != 31) r = ecb_ldexpf (m + 0x400, e - 25);
- else if (m ) r = ECB_NAN;
- else r = ECB_INFINITY;
-
- return x & 0x8000 ? -r : r;
- }
-
/* convert a float to ieee single/binary32 */
ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
ecb_function_ ecb_const uint32_t
return r;
}
+ /* convert a float to ieee half/binary16 */
+ ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
+ ecb_function_ ecb_const uint16_t
+ ecb_float_to_binary16 (float x)
+ {
+ return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
+ }
+
+ /* convert an ieee half/binary16 to float */
+ ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
+ ecb_function_ ecb_const float
+ ecb_binary16_to_float (uint16_t x)
+ {
+ return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
+ }
+
#endif
#endif