stillwater-sc / universal / 32330082276

85%

master: 84%

push

github

fix(bfloat16): report the true scale of subnormals (#1359)

scale() read the biased exponent field and subtracted the bias. That is right
for normals and wrong for every subnormal: their exponent field is zero, so it
answered -127 for all 254 of them, where the true scale runs from -127 down to
-133 with the leading one of the 7-bit fraction. Understated by up to 6 binades.

bfloat16's own suites did not catch it because none of them asked a subnormal for
its scale. It surfaced from elreal, where block<FpType>::scale_of_v() calls
scale() and the block's combined exponent is scale_of_v() + exp -- so a block
holding a subnormal carried a combined exponent up to 6 binades wrong, and the
0-overlap accounting built on that exponent was wrong with it.

The new regression is exhaustive rather than sampled: bfloat16 has 65536
encodings, so every one is compared against std::ilogb of the exactly-equal
double. It also fails if the sweep sees no subnormal at all, since a run that
never reaches one would pass vacuously. Mutation-tested: restoring the old
expression reports 126 failures.

This came out of an attempt at #1051 that did not work. The attempt and the
reason are recorded in docs/design/elreal-narrow-host-blocks.md so the next
person does not repeat it: scale-normalised blocks behave exactly as designed and
keep double bit-identical, but bfloat16 then violates the 0-overlap invariant.
Not via the comparison -- zero_overlap uses exponent(), which normalise()
preserves -- but via operand alignment. twoAdd brings two blocks to a common
scale by ldexp'ing v by the exp difference, and moving the scale out of v into
exp makes that difference large enough to underflow the operand before the EFT
sees it. Keeping the scale in v is what makes alignment expressible in host
arithmetic; the denormal floors are the price, and they are the cheaper side of
the tension. A real fix has to attack the alignment instead, and the doc names
where to start.

12/12 bfloat16 tes... (continued)

8 of 8 new or added lines in 1 file covered. (100.0%)

43051 of 50432 relevant lines covered (85.36%)

6897551.45 hits per line