Module fp4 

Software e2m1 conversion, on u32 bit patterns only, so a backend with no 4-bit float type can still decode and encode fp4 from the bits in a word.

This is the fp4 counterpart of cubecl_core::post_processing::minifloat, and it exists for the same reason: e2m1 conversion is a CUDA intrinsic and nothing else. Every other backend either has no 4-bit float type at all or can only move one around, so a quantized kernel that reaches for e2m1x2::from_bits runs on one vendor. The arithmetic below runs everywhere.

It does not go through the general minifloat path. That one reconstructs an f32 bit pattern field by field, which is the only tractable way to cover eight exponent bits; e2m1 has four codes per sign and its eight magnitudes are {0, 0.5, 1, 1.5, 2, 3, 4, 6}, small enough that decoding is one select over the subnormal arm and encoding is a count of the midpoints a magnitude clears.

Modules

e2m1_bits_to_float

e2m1_bytes_to_f16

e2m1_packed_bits_to_float

e2m1_words_to_f16

e2m1_words_to_f16_placed

float_to_e2m1_bits

Structs

LowerFp4Cast

Lowers every cast from or to e2m1x2 onto the software codec, for a backend that has no native fp4 conversion.

Constants

E2M1_F16_LIFT

What a code placed as an f16 near the bottom of its range, its sign, exponent and mantissa moved onto the f16’s own, is short of its value by: 2^14, the gap between an e2m1 exponent of zero and an f16 one.

Functions

e2m1_bits_to_float

Decode one e2m1 code per lane, held in the low nibble of each lane of code.

e2m1_bytes_to_f16

e2m1_words_to_f16 for bytes that do not fill a word, one e2m1x2 per lane of bytes: the high nibble moves up to the second half of the word first, one more mask and shift a pair.

e2m1_packed_bits_to_float

Decode the N e2m1 codes packed into the low 4 * N bits of word, lowest nibble first.

e2m1_words_to_f16

Decode the eight codes of each word of words, lowest nibble first, to f16.

e2m1_words_to_f16_placed

The eight codes of each word of words, lowest nibble first, placed as f16 each E2M1_F16_LIFT short of its value, exactly: what a caller that multiplies the values by a factor anyway, a block scale, takes, multiplying the lift into that factor once rather than paying a multiply a value.

float_to_e2m1_bits

Encode one e2m1 code per lane into the low nibble of each lane, rounding to nearest with ties to even and saturating at ±6.

Type Aliases

LowerFp4CastPass