stillwater-sc / universal / 22929553272

84%

master: 84%

Build Type

push

github

Committed by web-flow

Commit Message

feat(lns): add fused dot product for lns and dbns via generalized quire (#554)

* feat(lns,dbns): add fused dot product via generalized quire

Implement quire_mul, quire_resolve, fdp, fdp_stride, and fdp_qc for
both lns and dbns number systems. Since these types store values in
log domain, quire_mul bridges to linear domain via double conversion,
decomposes the product with std::frexp, and packs into a blocktriple
for quire accumulation.

The double-precision path is exact for lns/dbns types with <=52
exponent fractional bits. For wider types, a native log-to-linear
conversion using extended precision would be needed.

Test suites cover: quire_mul products, basic fdp, 1024-element vectors,
limb-boundary carry/borrow, catastrophic cancellation, strided fdp,
quire continuation, multiple configs (lns<8,4>, <12,6>, <16,8>, <32,16>;
dbns<8,4>, <9,5>, <10,5>), and accuracy vs double reference.

Resolves #549
Relates to #345

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* ci: add dbns to allowed conventional commit scopes

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(lns): use BLAS-style strided loop semantics in fdp_qc and fdp_stride

Change strided loops from `for (ix=0; ix<n; ix+=incx)` to
`for (i=0; i<n; ++i) { ix = i*incx; }` so n is the product count,
not an index bound. This matches BLAS conventions where n specifies
how many dot-product terms to compute.

Fix applies to both lns and dbns fdp_qc and fdp_stride functions.
Test stride parameters updated accordingly (n=2 for 2 products at stride 2).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(lns): guard quire_mul double conversion for wide lns types

Replace inline double conversion with if constexpr split: small lns
types (2*max_exp <= 1023) use fast double bridge, wide types compute
exponent sum directly and extract significand via std::pow(2, frac).
Add static_assert to dbns quire_mul confirming double range is safe.
Use std::ldexp for fractional expo... (continued)

Run Details

128 of 142 new or added lines in 2 files covered. (90.14%)

6 existing lines in 3 files now uncovered.

42687 of 50622 relevant lines covered (84.32%)

6137449.06 hits per line