stillwater-sc / universal / 23689705135

85%

master: 84%

Build Type

push

github

Committed by web-flow

Commit Message

feat(ucalc): mixed-precision dot product command (#650)

* feat(ucalc): mixed-precision dot product command (#629)

Add `dot [v1] [v2] [accum=<type>]` command for computing dot products
with explicit control over element type and accumulation type.

Elements are multiplied in the active type, products are accumulated
in the accumulation type (defaults to active type). Result is compared
against a quad-double reference.

Key use case: demonstrating how accumulation precision affects results.
Example: float elements with [1e10, 1, -1e10].[1, 1, 1]:
- float accum:  0 (catastrophic cancellation, rel_error = 1.0)
- dd accum:     1.0 (exact)

All output formats: plain, --json, --csv, --quiet.
21/21 CTests pass on gcc.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix(dfloat): implement proper operator++/-- for next representable value

dfloat's operator++ was adding 1 (integer increment) instead of
advancing to the next representable value. This caused
numeric_limits::epsilon() to return 1.0 instead of 10^(1-ndigits).

The fix normalizes the significand to exactly ndigits decimal digits
before incrementing, so that +1 on the significand gives the true
successor. For example, dfloat<7>(1.0) stored as (sig=1, exp=0) is
first normalized to (sig=1000000, exp=-6), then incremented to
(sig=1000001, exp=-6) = 1.000001.

Handles boundary cases:
- Significand overflow (10^ndigits): roll to 10^(ndigits-1), exp++
- Significand underflow (below 10^(ndigits-1)): roll to 10^ndigits-1, exp--
- Zero: advance to minpos/minneg
- Exponent overflow/underflow: saturate to inf/zero

epsilon() now returns correct values:
- decimal32 (ndigits=7):  1e-6
- decimal64 (ndigits=16): 1e-15

Also reverts the ucalc numberline workaround (ulp_unreliable) since
the underlying type bug is fixed.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat(universal): add to_native() API across all number systems

Add a to_native() free function to ... (continued)

Coverage Stats

28 of 42 new or added lines in 1 file covered. (66.67%)

10 existing lines in 2 files now uncovered.

44590 of 52858 relevant lines covered (84.36%)

6101828.35 hits per line