Module minifloat 

Software fp8 conversion, on u32 bit patterns only so that a backend with no 8- or 16-bit types can still call it on the bits in a word.

Modules

f32_to_fp8_bits

f32_to_ue8m0_bits

fp8_bits_to_f32

ue8m0_bits_to_f32

Structs

LowerMinifloatCast

Lowers every cast from or to an fp8 format the backend does not convert natively onto the software polyfill, through f32.

LowerMinifloatCompare

Lowers fp8 equality onto the lanes’ bit patterns.

Enums

Fp8Container

How a backend without a native fp8 type holds the bytes of an fp8 vector.

Functions

f32_to_fp8_bits

Round to nearest even; overflow and infinities saturate to the largest finite value, as the host codecs do.

f32_to_ue8m0_bits

Encode to ue8m0, rounding up to the next power of two and saturating at both ends.

fp8_bits_to_f32

Bits above the low byte are ignored.

ue8m0_bits_to_f32

Decode ue8m0 codes: a bare exponent, so the code is the f32 exponent field.

Type Aliases

LowerMinifloatCastPass

LowerMinifloatComparePass