File size: 2,264 Bytes
8b7c501
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
71
72
73
74
75
76
77
78
79
80
#pragma once

/*
 * This code snippet posted by user Phernost on 
 * https://stackoverflow.com/questions/1659440/32-bit-to-16-bit-floating-point-conversion
 *
 * compress and decompress methods are made "inline" for performance
 */

class Float16Compressor
{
	union Bits
	{
		float f;
		int32_t si;
		uint32_t ui;
	};

	static int const shift = 13;
	static int const shiftSign = 16;

	static int32_t const infN = 0x7F800000; // flt32 infinity
	static int32_t const maxN = 0x477FE000; // max flt16 normal as a flt32
	static int32_t const minN = 0x38800000; // min flt16 normal as a flt32
	static int32_t const signN = 0x80000000; // flt32 sign bit

	static int32_t const infC = infN >> shift;
	static int32_t const nanN = (infC + 1) << shift; // minimum flt16 nan as a flt32
	static int32_t const maxC = maxN >> shift;
	static int32_t const minC = minN >> shift;
	static int32_t const signC = signN >> shiftSign; // flt16 sign bit

	static int32_t const mulN = 0x52000000; // (1 << 23) / minN
	static int32_t const mulC = 0x33800000; // minN / (1 << (23 - shift))

	static int32_t const subC = 0x003FF; // max flt32 subnormal down shifted
	static int32_t const norC = 0x00400; // min flt32 normal down shifted

	static int32_t const maxD = infC - maxC - 1;
	static int32_t const minD = minC - subC - 1;

public:

	inline static uint16_t compress(float value)
	{
		Bits v, s;
		v.f = value;
		uint32_t sign = v.si & signN;
		v.si ^= sign;
		sign >>= shiftSign; // logical shift
		s.si = mulN;
		s.si = s.f * v.f; // correct subnormals
		v.si ^= (s.si ^ v.si) & -(minN > v.si);
		v.si ^= (infN ^ v.si) & -((infN > v.si) & (v.si > maxN));
		v.si ^= (nanN ^ v.si) & -((nanN > v.si) & (v.si > infN));
		v.ui >>= shift; // logical shift
		v.si ^= ((v.si - maxD) ^ v.si) & -(v.si > maxC);
		v.si ^= ((v.si - minD) ^ v.si) & -(v.si > subC);
		return v.ui | sign;
	}

	inline static float decompress(uint16_t value)
	{
		Bits v;
		v.ui = value;
		int32_t sign = v.si & signC;
		v.si ^= sign;
		sign <<= shiftSign;
		v.si ^= ((v.si + minD) ^ v.si) & -(v.si > subC);
		v.si ^= ((v.si + maxD) ^ v.si) & -(v.si > maxC);
		Bits s;
		s.si = mulC;
		s.f *= v.si;
		int32_t mask = -(norC > v.si);
		v.si <<= shift;
		v.si ^= (s.si ^ v.si) & mask;
		v.si |= sign;
		return v.f;
	}
};