private static float float16_to_float32(ushort h)
{
fp32 ans = 0;
UInt16 h_exp, h_sig;
UInt32 f_sgn, f_exp, f_sig;
h_exp = (UInt16)(h & 0x7c00u);
f_sgn = ((UInt32)h & 0x8000u) << 16;
switch (h_exp)
{
case 0x0000: /* 0 or subnormal */
h_sig = (UInt16)(h & 0x03ffu);
/* Signed zero */
if (h_sig == 0)
{
ans = f_sgn;
return(ans.f);
}
/* Subnormal */
h_sig <<= 1;
while ((h_sig & 0x0400u) == 0)
{
h_sig <<= 1;
h_exp++;
}
f_exp = ((UInt32)(127 - 15 - h_exp)) << 23;
f_sig = ((UInt32)(h_sig & 0x03ffu)) << 13;
ans = f_sgn + f_exp + f_sig;
return(ans.f);
case 0x7c00: /* inf or NaN */
/* All-ones exponent and a copy of the significand */
ans = f_sgn + 0x7f800000u + (((UInt32)(h & 0x03ffu)) << 13);
return(ans.f);
default: /* normalized */
/* Just need to adjust the exponent and shift */
ans = f_sgn + (((UInt32)(h & 0x7fffu) + 0x1c000u) << 13);
return(ans.f);
}
}
private static ushort make_float16(float value)
{
fp32 f32infty = 255U << 23;
fp32 f16infty = 31U << 23;
fp32 magic = 15U << 23;
uint sign_mask = 0x80000000U;
uint round_mask = ~0xFFFU;
fp32 @in = 0;
ushort @out = 0;
@in.f = value;
uint sign = @in.u & sign_mask;
@in.u ^= sign;
if (@in.u >= f32infty.u)
{
@out = (@in.u > f32infty.u) ? (ushort)0x7FFFU : (ushort)0x7C00U;
}
else
{
@in.u &= round_mask;
@in.f *= magic.f;
@in.u -= round_mask;
if (@in.u > f16infty.u)
{
@in.u = f16infty.u;
}
@out = (ushort)(@in.u >> 13);
}
@out |= (ushort)(sign >> 16);
return(@out);
}
