unum+saloon: finish posit transcendentals; wire %unum into Saloon#58
Merged
Conversation
/lib/unum gains the five elementary functions it was missing (factorial, cbrt, atan, asin, acos), mirroring /lib/math's algorithms in posit arithmetic with NaR domain guards. (The reference SoftPosit has no transcendentals, so these follow the same naive Taylor/AGM series as /lib/math; #18's Chebyshev rewrite will harden both real and posit layers together.) Saloon's Tier-B dispatch (+trans-scalar / +fun-scalar) gains a %unum branch (posits are bloq 3/4/5/6; unum's sqrt is +sqt), so exp/sin/cos/tan/log/sqrt/ cbrt/factorial/pow/pow-n now work on %unum rays, not just %i754. Oracle + tests: - tools/posit_check.py: a posit-arithmetic series replica + check_transcendental that confirms the series is correctly rounded vs mpmath at posit8 (11/11). - tests/lib/unum-fns: the new arms + domain guards at posit8, and the full transcendental set at posit8/16/32 (Hoon == replica at all three widths). - tests/lib/saloon-unum: ray-level %unum dispatch through Saloon's Tier B. All green on ~hex; %i754 eig and the lagoon suites unaffected. Complex transcendentals are penciled in (a TODO in /lib/complex with the closed forms) as the next follow-on. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…loon /lib/complex gains complex elementary functions (cexp, clog, csqrt, csin, ccos, ctan, cpow) in all four width doors (cs/cd/ch/cq). Kept self-contained -- like /lib/unum, /lib/fixed, /lib/twoc -- by giving each door its OWN real exp/sin/ cos/log/atan as naive Taylor/AGM series over the stdlib component float door, rather than depending on /lib/math (which would couple the kind-primitive to the special-functions lib and pull it into Lagoon's build). Saloon's Tier-B dispatch (+trans-scalar / +fun-scalar) gains a %cplx branch (complex bloqs 5/6/7/8 = ch/cs/cd/cq), so exp/sin/cos/tan/log/sqrt/pow now work on %cplx rays. Real-only / ordered functions (factorial, log-10, log-2, cbrt, pow-n) are undefined on complex and crash. Tests (validated vs numpy, all green on ~hex): - tests/lib/complex-fns: the full set at @cs, exp/log/sqrt at @cd, exp at @ch/@cq (cd matches complex128 to ~15 digits, cq to ~34). - tests/lib/saloon-cplx: ray-level %cplx dispatch through Saloon Tier B, plus factorial/cbrt crash-on-complex guards. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Edge-case testing surfaced a bug: the self-contained real +rlog lacked the x<=0 domain guard the unum +log has, so clog(0+0i) ran the atanh series on y=-1 and returned garbage (~-5.4) instead of the -inf singularity. Add a log(0) -> -inf guard to +rlog in all four doors (|z| is always >= 0, so 0 is the only singular input). Add /tests/lib/complex-edge: origin, real/imaginary axes, the csqrt/clog branch cut (-1, +-i), the csqrt sign branch (sqrt(-i) = 0.707 - 0.707i), the clog(0) singularity, off-axis (1+1i, 3+4i), and two large-argument cases locked as documented limitations (the naive series is ~0.2% off far from the origin). All validated against numpy complex64. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
complex-edge now covers the same edges across all four widths -- @cd/@ch/@cq get origin, the csqrt/clog branch cut (-1, -i), the clog(0)=-inf singularity, and the imaginary axis -- plus: - cpow edges: 2^0=1, 2^2=4, i^2=-1, (-1)^0.5=i, and the 0^2 corner (documented: yields NaN, since cpow routes through clog(0)=-inf -> exp(-inf), which the naive series can't represent -- locked as a known limitation, not correctness). - rounding: cexp(1+2i) under %n/%u/%d/%z gives four distinct bit patterns, confirming the rounding mode propagates bit-exactly through the complex ops. 39 cases total, all validated against numpy complex64/128. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds tests/lib/unum-edge (28 cases), the %unum analogue of complex-edge: - exact identities (exp0=1, cos0=1, sqt4=2, fact5=120, log1=0, acos1=0, ...) - sign (exp(-1)=1/e, atan(-1)=-pi/4) - NaR propagation (exp/sin/sqt/factorial of NaR -> NaR) - domain -> NaR (sqrt(-1)) - a posit8 cross-width spot (identity + NaR propagation) - large-argument breakdown / saturation, locked as DOCUMENTED limitations: exp(10)~21991 vs 22026; exp(50)/exp(100) saturate near maxpos; sin(10) blows up out of [-1,1] (no range reduction); log (atanh) holds up (log(100)~4.605). No bug found (the unum domain guards from the transcendental PR already hold); posits round-to-nearest-even so there is no rounding-mode axis to sweep. The large-arg cases are the #18 hazard made explicit, not asserted as correct. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
unum: posit transcendental edge-case sweep
complex+saloon: implement complex transcendentals; wire %cplx into Saloon
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Completes the
%unumtranscendental story so Saloon's Tier B works on posits, not just%i754./lib/unum— the five missing elementary functionsfactorial,cbrt,atan,asin,acos, mirroring/lib/math's algorithms in posit arithmetic withNaRdomain guards (negative/out-of-range → NaR per the unum convention). The reference SoftPosit has no transcendentals, so — like/lib/math— these are the naive Taylor/AGM series; #18's Chebyshev rewrite will harden the real and posit layers together.Saloon —
%unumdispatch+trans-scalar/+fun-scalargain a%unumbranch (posits are bloq 3/4/5/6, and unum's sqrt is+sqt), soexp/sin/cos/tan/log/sqrt/cbrt/factorial/pow/pow-nnow accept%unumrays.Oracle + tests (all green on
~hex)tools/posit_check.py: a posit-arithmetic series replica +check_transcendental— confirms the series is correctly rounded vs mpmath at posit8 (11/11). (asin/acos/small-argatansit ~1 ULP off and are excluded from the exact gate, as noted.)tests/lib/unum-fns: new arms + domain guards at posit8, and the full transcendental set at posit8/16/32 (Hoon == replica at all three widths).tests/lib/saloon-unum: ray-level%unumdispatch through Saloon Tier B.%i754eig and all lagoon suites unaffected.Complex — penciled in
A
TODOin/lib/complexwith the closed forms (cexp/clog/csqrt/…) for the next follow-on, which would then wire a%cplxbranch into Saloon the same way.🤖 Generated with Claude Code