stillwater-sc / universal / 29621270080

85%

master: 84%

push

github

fix(bfloat16): round float conversion to nearest-even instead of truncating (#1134)

convert_ieee754() kept the upper 16 bits of the source IEEE-754 float by a
bare `bits >> 16`, discarding the low 16 bits with no rounding. That is
round-toward-zero (truncation); Google TPUs and Intel round the float->bfloat16
conversion to nearest, ties-to-even (RNE).

Add RNE via the standard magic rounding bias `0x7FFF + lsb_of_retained_field`
before the shift, so more-than-half rounds up, less-than-half stays, and an
exact-half tie rounds toward the even significand. A carry propagates correctly
into the exponent, and a value within half a ulp of the range limit rounds to
inf. A dedicated NaN guard keeps a float NaN whose payload lives only in the
low 16 bits from collapsing to +/-inf.

Update numeric_limits<bfloat16>::round_style to round_to_nearest to match, and
add static/float/bfloat16/conversion/round_to_nearest_even.cpp covering the
issue's case (0.2691408770292272 -> 0.26953125, previously 0.267578125),
up/down rounding, tie-to-even boundaries, and inf/nan/zero preservation.

Resolves #1133

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

5 of 5 new or added lines in 1 file covered. (100.0%)

49217 of 58162 relevant lines covered (84.62%)

6677199.64 hits per line