rontolisp:float16-bits

(rontolisp:float16-bits real)

Returns the IEEE 754 binary16 (f16) bit pattern of a real, an integer 0-65535. Unlike bfloat16, f16 is a genuinely different exponent/mantissa split from f32 -- five exponent bits and ten mantissa bits, not eight and seven -- so it trades range as well as precision for width, and a value outside its exponent range widens to an infinity rather than a rounded finite number.

The narrowing rounds to nearest, ties to even, same as rontolisp:bfloat16-bits: the two exact midpoints below round in opposite directions.

rontolisp:bits-float16 takes the pattern back. Sixteen bits fit a fixnum on every backend, so -- unlike the wider IEEE pairs float-features:single-float-bits and friends work with -- this pair needs no bignum model and is a rontolisp: primitive rather than a float-features: one; see rontolisp:widen-float-bits and rontolisp:narrow-float-bits for the bulk form a whole tensor of these patterns actually arrives as.