stillwater-sc / universal / 36938902036

87%

master: 84%

push

github

feat(takum): opt-in fast takum<16,3,uint16_t> specialization (#1614)

TAKUM_FAST_TAKUM_16 (or TAKUM_FAST_SPECIALIZATION) replaces the codec members
of takum<16,3,uint16_t> -- convert_ieee754 and to_ieee754 for double and float
-- and the four compound arithmetic operators. Everything else, and every
other configuration, keeps the generic code. Results are bit-identical to the
generic implementation; multiply-add runs ~10x faster.

Encode: per-double-exponent table of (first pattern of the binade, shift),
round-to-nearest-even of the top p fraction bits, saturate at maxpos.
Decode: 512 KB table built on first use (function-local static), with a
constexpr formula for constant evaluation.

Verified bit for bit against takum<16,3,uint8_t> (same generic code): all 2^32
operand pairs for + - * / and comparisons, all 2^32 floats, all 2^16 unary ops,
values/midpoints/neighbours and 10^8 random doubles. tkm_takum_16_3 runs a
sampled subset in CI; tkm_takum_16_3_exhaustive is opt-in.

Signed-off-by: James Quinlan <james@stillwater-sc.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Theodore Omtzigt <theo@stillwater-sc.com>

47 of 47 new or added lines in 1 file covered. (100.0%)

45609 of 53021 relevant lines covered (86.02%)

9717542.99 hits per line