rontolisp:widen-float-bits

(rontolisp:widen-float-bits bits format dst &key (start 0))

Widens a packed (unsigned-byte 16) vector of bits -- f16 or bfloat16 bit patterns, chosen by format (:float16 or :bfloat16) -- into dst, a packed float array (single-float, double-float or bfloat16, any rank), row-major from flat index start. Returns dst.

This is the bulk form of rontolisp:bits-float16/rontolisp:bits-bfloat16: a published checkpoint's tensors arrive as a whole vector of sixteen-bit patterns, never one element at a time, and widening them one call at a time would cost a function-call boundary per element. Every element gets exactly the scalar primitive's answer -- widening is total and exact for both formats, so this never rounds.

A bfloat16 destination stores bit patterns rather than values. With format :bfloat16 that makes it a plain copy -- the patterns already are what the array holds, so nothing rounds and no NaN payload changes. With format :float16 it is one rounding, the same one rontolisp:bfloat16-bits applies, and it is the way to load a published f16 checkpoint into the narrow width without allocating a single-float array on the way.

:start is where a chunked read lands its next slice inside a whole tensor's destination -- see checkpoint:stage-float-bits for the pattern. Works on the interpreter, the JVM and both WASM backends; --no-gc has no packed float array model and refuses at compile time. A bfloat16 destination is the interpreter and the JVM only, since no WASM backend carries that array width. rontolisp:narrow-float-bits is the inverse direction.