V3.96: mixed-precision ArXiv paper prep and UNUM2 number system by Ravenwater · Pull Request #506 · stillwater-sc/universal

Ravenwater · 2026-02-11T17:03:25Z

mixed-precision ArXiv paper prep and UNUM2 number system

Summary by CodeRabbit

New Features
- Added complete posit2 arithmetic operations (subtraction, multiplication, division)
- Introduced mixed-precision attention head implementation with KV cache management
- Added comprehensive test suites for posit2 conversion, assignment, and logic operators
Documentation
- Added arXiv systems and position papers with feasibility-focused narratives
- Included LaTeX scaffolding and reproducibility artifacts for research papers
- Expanded mixed-precision solver case studies documentation
Chores
- Bumped minor version to v3.96
- Updated CI workflow triggers

When exponent is a large negative (e.g. -72), the shift `1ull << -exponent` exceeds 63 bits, causing undefined behavior. Add `exponent > -64` guard so both positive and negative extremes fall through to the safe ipow() path. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Create self-contained papers/ directory for systems and position paper artifacts that can be zipped and shared with reviewers: - papers/systems-paper/iterative_refinement.cpp: Carson & Higham three-precision LU-IR across IEEE, posit, cfloat, dd, cross-family - papers/systems-paper/conjugate_gradient.cpp: CG for SPD systems with single-precision and two-precision (low preconditioner) configurations - papers/systems-paper/idrs.cpp: IDR(s) for non-symmetric systems with shadow space dimension sweep and number system comparison Move paper docs from docs/papers/ to papers/docs/ for co-location. Add UNIVERSAL_BUILD_PAPERS CMake option (wired into BUILD_ALL cascade). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

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

Plain article class (12pt) with full section structure, TODO placeholders, 29 BibTeX references (14 from JOSS + 15 new), and Makefile for local builds. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

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

…he ArXiv paper

…ites Add all four arithmetic operations (sub, mul, div plus existing add) via blocktriple pipeline, port conversion/assignment/logic regression tests from original posit, and fix three bugs discovered during testing: - convert_ieee754() extractBits too small: nbits+4 lost IEEE sticky bits causing false midpoint ties; now uses max(numeric_limits<Real>::digits, nbits+4) - Integer assignment via blocktriple had hidden-bit off-by-one in round(); rerouted through convert_ieee754(static_cast<double>(rhs)) - Literal comparison operators accessed private _block member; replaced with delegation to posit-posit comparison operators Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

positRegime::value() used manual division (1.0l / uint64_t(1) << -e2) for negative exponents, which produced wrong results under clang due to a codegen issue in this template context. Replace with std::ldexp() matching the original posit implementation. posit::setbits(uint64_t) used an uninitialized blockbinary temporary leaving upper MSU bits as garbage. Replace with _block.setbits(value) which properly masks the MSU. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

On MSVC, LONG_DOUBLE_SUPPORT is 0 but the long double comparison operators were not guarded, causing ambiguous conversion errors since the posit(long double) constructor was correctly excluded. Add matching #if LONG_DOUBLE_SUPPORT guards to friend declarations, operator implementations, and test code. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* Skeletal implementation of Unum 2.0 * Fix point multiplication bug and add support for reverse interval * Added pow() and abs() + some optimizations * Refactor according to Codacy suggestions * Added op table/matrix for efficient operations * Fix unum2 includes * Fix operation matrix bug in unum2_impl.hpp * unum2_fwd.hpp, static test and improvements * Change bitset::_Find_first() to manually finding the bit for compatibility reasons * Make unum2 friend class a little more specific on class lattice * Make lattice parameters public to bypass MSVC build fails

coderabbitai · 2026-02-11T17:03:52Z

Caution

Review failed

The pull request is closed.

📝 Walkthrough

Walkthrough

This PR adds extensive new mixed-precision applications and solver case studies, a large redesign and feature expansion of unum2 and posit2 (including arithmetic, conversion, and rounding fixes), numerous new tests and test runners, paper artifacts/LaTeX scaffolding, CMake build options for paper artifacts, and various small internal fixes (block boundaries, includes, exponent handling).

Changes

Cohort / File(s)	Summary
CI / Versioning `/.github/workflows/cmake.yml`, `CMakeLists.txt`	Bumped workflow trigger (v3.95→v3.96); added `UNIVERSAL_BUILD_PAPERS` option, auto-enable behavior when `UNIVERSAL_BUILD_ALL` is ON, and conditional inclusion of paper/application subdirs.
Mixed-Precision Attention App `applications/mixed-precision/attention/CMakeLists.txt`, `applications/mixed-precision/attention/attention.cpp`, `applications/mixed-precision/attention_head.md`	New attention application and docs: templated AttentionHead with KV cache, benchmarking across numeric types, energy/latency/accuracy estimation, and build wiring.
Posit2 Core & Tests `include/sw/universal/number/posit2/posit_impl.hpp`, `include/sw/universal/number/posit2/positRegime.hpp`, `static/posit2/...`	Major posit2 overhaul: frexp-based IEEE→posit path, unified convert paths, new normalizeMultiplication/Division, operator updates, rounding/boundary fixes, copy/construct semantics, extensive new/existing test suites for conversion, assignment, logic, addition, subtraction, multiplication, division.
Block-format Boundary Fixes `include/sw/universal/internal/blockbinary/blockbinary.hpp`, `include/sw/universal/internal/blocksignificand/blocksignificand.hpp`	Changed anyAfter() to iterate up to min(bitIndex, nbits) to correct out-of-range behavior and unify control flow.
Unum2 Lattice Redesign `include/sw/universal/number/unum2/`, `static/unum2/`	Replaced old unum2 API with a lattice-based SORN design: added `lattice<...>`, `op_matrix<T>`, new `unum2<lattice<...>>` specialization, forward headers, removed legacy manipulators/math headers, and updated tests to new model.
Solver Case Studies (papers) `papers/systems-paper/iterative_refinement.cpp`, `papers/systems-paper/conjugate_gradient.cpp`, `papers/systems-paper/idrs.cpp`	Added mixed-precision iterative refinement, preconditioned CG (including two-precision variants), and IDR(s) solver implementations with reporting and experiments across many numeric types.
Paper Artifacts & Build `papers/systems-paper/paper/*`, `papers/position-paper/CMakeLists.txt`, `papers/systems-paper/CMakeLists.txt`	Added LaTeX manuscript (`main.tex`), `references.bib`, Makefile, and CMake globs to build paper artifacts; position-paper CMake entry added.
Documentation & Changelog `CHANGELOG.md`, `docs/`, `papers/docs/`	Large additions: changelog entries, mixed-precision findings, plans/roadmaps for systems and position papers, session notes, and iterative refinement docs.
Small API/Include Fixes `include/sw/universal/number/areal/areal_impl.hpp`, `include/sw/universal/number/posit/posit_parse.hpp`, `include/sw/universal/number/shared/blocktype.hpp`, `include/sw/universal/number/support/decimal.hpp`, `applications/precision/ubit/thin_triangle.cpp`	Added `<cstdint>` includes, adjusted areal exponent fast-path bounds, updated cfloat instantiations to include ubit flag, and minor parse/format tweaks.
Removed / CMake cleanup `include/sw/universal/number/unum2/CMakeLists.txt`, `include/sw/universal/number/unum2/manipulators.hpp`, `include/sw/universal/number/unum2/math_functions.hpp`	Removed old unum2 CMake glob and deleted legacy manipulators/math_functions headers as part of unum2 redesign.

Sequence Diagram(s)

(Skipped — changes are broad and include multiple independent features rather than a single multi-component sequential flow that benefits from a sequence diagram.)

Estimated code review effort

🎯 5 (Critical) | ⏱️ ~120 minutes

Possibly related PRs

V3.86 #481 — touches the same internal multi-limb routines (blockbinary/blocksignificand anyAfter behavior); likely closely related to posit2 rounding/boundary fixes.
V3.83 #475 — overlaps with version/CMake bump and CI workflow updates (UNIVERSAL_VERSION_MINOR and workflow triggers).
V3.94: interval arithmetic for mixed-precision algorithm SDK #501 — overlaps in mixed-precision energy/instrumentation infrastructure used by the new attention benchmark and energy estimators.

Poem

🐰 I hopped through lattice, bit and byte,
I cached each KV token bright,
Posit2 rounded, unum danced,
Papers sprouted, tests advanced,
Energy counted, metrics light ✨

✨ Finishing touches

📝 Generate docstrings

🧪 Generate unit tests (beta)

Create PR with unit tests
Post copyable unit tests in a comment
Commit unit tests in branch v3.96

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

coveralls · 2026-02-11T17:32:11Z

Pull Request Test Coverage Report for Build 21914780216

Warning: This coverage report may be inaccurate.

This pull request's base commit is no longer the HEAD commit of its target branch. This means it includes changes from outside the original pull request, including, potentially, unrelated coverage changes.

For more information on this, see Tracking coverage changes with pull request builds.
To avoid this issue with future PRs, see these Recommended CI Configurations.
For a quick fix, rebase this PR at GitHub. Your next report should be accurate.

Details

495 of 570 (86.84%) changed or added relevant lines in 9 files are covered.
14 unchanged lines in 3 files lost coverage.
Overall coverage increased (+0.09%) to 84.914%

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
include/sw/universal/number/posit2/posit_impl.hpp	97	101	96.04%
include/sw/universal/number/unum2/op_matrix.hpp	30	35	85.71%
include/sw/universal/number/unum2/lattice.hpp	69	79	87.34%
include/sw/universal/number/unum2/unum2_impl.hpp	291	347	83.86%

Files with Coverage Reduction	New Missed Lines	%
include/sw/universal/number/posit2/positExponent.hpp	1	83.61%
include/sw/universal/internal/blocktriple/blocktriple.hpp	3	87.12%
include/sw/universal/number/posit2/posit_impl.hpp	10	94.61%

Totals
Change from base Build 21833272337:	0.09%
Covered Lines:	36345
Relevant Lines:	42802

💛 - Coveralls

Ravenwater and others added 15 commits February 9, 2026 13:18

Incrementing SEMVER to v3.96.1

8155422

weird difference between double and duble

0e5bfb3

systems and position paper roadmaps

7cb79ba

Add changelog and session doc for paper artifact tree and solver studies

6ed33fa

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

Add LaTeX scaffolding for arXiv systems paper

f404aee

Plain article class (12pt) with full section structure, TODO placeholders, 29 BibTeX references (14 from JOSS + 15 new), and Makefile for local builds. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Add changelog and session doc for LaTeX paper scaffolding

750973a

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

arxive paper first draft

f811c07

arxiv systems paper draft v3

6d3d9f3

Adding a mixed-precision Attention head with KV cache test case for t…

a0ad9f5

…he ArXiv paper

Ravenwater self-assigned this Feb 11, 2026

Ravenwater merged commit 6daf46f into main Feb 11, 2026
14 of 16 checks passed

Ravenwater deleted the v3.96 branch February 11, 2026 17:05

coderabbitai bot mentioned this pull request Feb 12, 2026

V3.97: hygiene release #507

Merged

This was referenced Feb 24, 2026

V3.99: supporting cross-platform BlockType == uint64_t #511

Merged

feat(dfloat): portable blockbinary storage, string I/O, and formatting #524

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

V3.96: mixed-precision ArXiv paper prep and UNUM2 number system#506

V3.96: mixed-precision ArXiv paper prep and UNUM2 number system#506
Ravenwater merged 15 commits intomainfrom
v3.96

Ravenwater commented Feb 11, 2026 •

edited by coderabbitai bot

Loading

Uh oh!

coderabbitai bot commented Feb 11, 2026 •

edited

Loading

Review failed

Walkthrough

Changes

Sequence Diagram(s)

Estimated code review effort

Possibly related PRs

Poem

Uh oh!

Uh oh!

coveralls commented Feb 11, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

Conversation

Ravenwater commented Feb 11, 2026 • edited by coderabbitai bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary by CodeRabbit

Uh oh!

coderabbitai bot commented Feb 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Review failed

Walkthrough

Changes

Sequence Diagram(s)

Estimated code review effort

Possibly related PRs

Poem

Uh oh!

Uh oh!

coveralls commented Feb 11, 2026

Pull Request Test Coverage Report for Build 21914780216

Warning: This coverage report may be inaccurate.

Details

💛 - Coveralls

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Ravenwater commented Feb 11, 2026 •

edited by coderabbitai bot

Loading

coderabbitai bot commented Feb 11, 2026 •

edited

Loading