/usr/include/movit/fp16.h is in libmovit-dev 1.4.0-1.
This file is owned by root:root, with mode 0o644.
The actual contents of the file can be viewed below.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 | #ifndef _MOVIT_FP16_H
#define _MOVIT_FP16_H 1
#ifdef __F16C__
#include <immintrin.h>
#endif
// Code for converting to and from fp16 (from fp64), without any particular
// machine support, with proper IEEE round-to-even behavior (and correct
// handling of NaNs and infinities). This is needed because some OpenGL
// drivers don't properly round off when asked to convert data themselves.
//
// These routines are originally written by Fabian Giesen, and released by
// him into the public domain;
// see https://fgiesen.wordpress.com/2012/03/28/half-to-float-done-quic/.
// They are quite fast, and can be vectorized if need be; of course, using
// the f16c instructions (see below) will be faster still.
namespace movit {
// structs instead of ints, so that they are not implicitly convertible.
struct fp32_int_t {
unsigned int val;
};
struct fp16_int_t {
unsigned short val;
};
#ifdef __F16C__
// Use the f16c instructions from Haswell if available (and we know that they
// are at compile time).
static inline float fp16_to_fp32(fp16_int_t x)
{
return _cvtsh_ss(x.val);
}
static inline fp16_int_t fp32_to_fp16(float x)
{
fp16_int_t ret;
ret.val = _cvtss_sh(x, 0);
return ret;
}
#else
float fp16_to_fp32(fp16_int_t x);
fp16_int_t fp32_to_fp16(float x);
#endif
// Overloads for use in templates.
static inline float to_fp32(double x) { return x; }
static inline float to_fp32(float x) { return x; }
static inline float to_fp32(fp16_int_t x) { return fp16_to_fp32(x); }
template<class T> inline T from_fp32(float x);
template<> inline double from_fp32<double>(float x) { return x; }
template<> inline float from_fp32<float>(float x) { return x; }
template<> inline fp16_int_t from_fp32<fp16_int_t>(float x) { return fp32_to_fp16(x); }
template<class From, class To>
inline To convert_float(From x) { return from_fp32<To>(to_fp32(x)); }
template<class Same>
inline Same convert_float(Same x) { return x; }
} // namespace movit
#endif // _MOVIT_FP16_H
|