Repository navigation
Releases
Kernix 0.13.1
Kernix 0.13.1 preserves the public 0.13.0 API across kernix, kernix-core and kernix-ops.
- Fix MSE returning zero for representable subnormal mean squared errors; retain existing overflow and shape errors.
- Avoid the contiguous MSE difference allocation; reduce CSR stream-merge heap work and vectorize nonzero counting.
- Reject NaN/infinite benchmark errors and validate affected result geometry and gradient lengths.
- Freeze the API in AGENTS.md and docs/API-STABILITY.md, with a fixed 0.13.0 patch compatibility gate.
Verification: 587 default, 593 all-feature and 579 no-default-feature tests; Rust 1.86, no_std/Wasm, API compatibility, lint/docs and verified workspace packages.
Nine same-host process pairs: CSR-product latency decreases about 28%, automatic CSR construction about 3.7%; training recording uses 20 instead of 22 allocation requests. Dense QR (+2.2%) and f32 GEMM (+0.9%) guard slowdowns remain visible. Intervals are exploratory and describe this host.
Source/package revision: 0f31778.
Benchmark reports retain raw samples, source/dependency identities, allocator probes and losses. Historical records are unchanged.
Release QA: https://github.com/moderubias/kernix/blob/main/docs/QA-0.13.1.md
Benchmark report: https://github.com/moderubias/kernix/blob/main/benchmarks/results/0.13.1-final/README.md
Kernix 0.13.0 · automatic dense/CSR matrices
Kernix 0.13 adds convenient finite-real rank-two matrices with automatic Dense/CSR storage selection, while preserving existing Tensor APIs.
Matrix::from_rows,from_valuesandfrom_tensorselect the smaller initialized payload, including CSR indices. Explicit dense/sparse controls remain available.matvecand all four dense/CSR matrix products dispatch without converting operands. CSR products use a row-stream heap and an optional output-entry cap.- Generic checked CSR/COO assembly adopts non-Clone elements, sums duplicate coordinates in input order and exposes borrowed CSC transpose views.
- Fallible allocation, f32/f64, signed-stride inputs and no_std + alloc are supported. No sparse gradients, factorization or GPU execution are claimed.
On the local cyclic five-diagonal fixtures, prepared sparse kernels are approximately 6.6–276× faster than Kernix dense. The 2048×2048 f64 input payload falls from 32 MiB to about 176 KiB, including indices. SciPy/faer still win several sparse comparisons. Compression costs, different finite-input contracts and actual index widths are reported separately; these results do not establish a universal sparse lead or a dense128 improvement.
Readable local report · All 98 reports · Sparse usage/costs · Release verification.
All three crates are published through ordinary Cargo workspace publication. Actual registry archives match clean signed source 7c11b2165b0cb7b523db4d938a36f1396396a138. Independent registry and Git consumers pass Rust 1.86 no_std tests/executable, 31 release native/sparse regressions and stable Wasm checks. Source/tag CI, API/security assurance, Miri and all four hardware studies pass.
Install with cargo add kernix@0.13.0. Download the full archive for standalone interactive HTML, exact tables, vector PDFs, all original raw measurements, source snapshots and paired process evidence. Previous data and commentary remain byte-identical.
Kernix 0.12.0 — faster reusable solves and compact-WY QR
Registry update, 2026-10-26: all three 0.12.0 packages are now available through the ordinary crates.io resolver. Independent Rust 1.86 no_std tests/executable and stable Wasm checks pass; signed Git installation remains verified. The rollout is complete.
0.12 preserves existing public signatures and adds owned Cholesky rank-one updates/downdates.
- Prepared-factor Cholesky solves with 64 RHS: about 2.8–5.3× faster than exact registry 0.11.1 across five measured hosts; LU about 2.1–4.5×.
- Thin QR 1024×256: about 1.5–2.0× faster, with compact-WY panels and checked native/range fallback.
- Rank-one updates/downdates in O(n²), unchanged inputs on failure, no implicit jitter or dtype conversion. These are new APIs; no fabricated historical timings.
- Existing dense signatures, custom scalars, ordered paths, Rust 1.86 and no_std+alloc remain supported.
Readable local comparison · paired version intervals · 83-report catalogue · verification and limits
Ratios are exploratory medians over nine fresh-process pairs per workload, with bootstrap intervals in the reports. Ordinary dense128 remains a target; some small cases regress, and specialized peers remain faster in several workloads. No universal performance ranking is implied. Raw values, source snapshots, locks and methods are retained byte-for-byte.
Downloads contain standalone searchable HTML, vector PDF/exact tables, every workload PNG and original evidence. Unzip and open index.html. The complete archive contains all 83 studies; historical numerical data is unchanged.
Measured/tag source: 90cbbff60e5d7458ef9f89918c99a6cca94b51d1. All four final hardware jobs, nine source CI jobs and four assurance jobs passed.
All three packages are available on crates.io; actual downloaded archives and dependency lock checksums match Cargo-verified source. The registry index delay has resolved. Independent ordinary-registry Rust 1.86 no_std tests/executable and stable Wasm checks pass, as does signed Git installation.
Kernix 0.11.1 — small GEMM and readable performance evidence
Small dense f32/f64 products on AVX2/FMA x86-64 now use bounded kernels without packing. Contiguous broadcast batches use the same path. Ordinary matmul still allocates output; native hooks need no heap scratch. Public APIs, generic scalars, ordered execution, Rust 1.86 and no_std + alloc remain supported.
For [4,16,32] × [1,32,16], nine paired-process comparisons against exact registry 0.11.0 show approximately 24–34% less time for f64 and 29–42% less for f32 on the three tested x86-64 hosts. The Linux warmed probe drops from 7 to 3 allocations. The bounded path covers rows 4–32, contraction 4–64 and columns up to 16 in complete SIMD tiles. ARM/macOS retain existing paths. Large serial128 still needs work; Linux128 f64 records about 1% more time, retained in the report. These results do not establish a universal lead over other libraries.
68 reports now work directly on GitHub: every workload has a ranked zero-based chart, readable labels, explicit units and a plain-language caption. Paired version sheets identify both releases and show uncertainty. Exact tables retain every recorded digit; original measurements, source archives, locks and methods remain byte-identical. Download the complete archive and open index.html locally for search, filters and vector PDFs. The legacy release downloads for 0.8–0.11 are also refreshed with preserved data.
- Read the local report
- Read paired version changes
- Browse every report on GitHub
- Verification, package hashes, limitations and hardware evidence
All three crates and signed tag use clean 834d8fad54af5d322077941db2dabdb417328c26. Actual registry payloads match source/VCS/legal boundaries; renamed registry-only MSRV/no_std, native and wasm consumers pass. Source CI: 9 jobs, 4 assurance jobs; 4 actual hardware jobs. Local and hardware measurement identities and the test-only source difference are documented in the verification record.
ZIP downloads contain the local current/baseline/paired study, the four-platform current/baseline/paired studies, or the complete 68-report archive. The PNG is a standalone preview. The complete archive is approximately 147 MiB.
Kernix 0.11.0 — native factorization, explicit execution and readable evidence
Kernix 0.11.0 adds blocked Cholesky, measured ARM kernel selection, explicit Apple Accelerate execution and readable performance reports.
- Native Cholesky uses bounded panel workspace at n ≥ 192, preserving validation, finite checks and pivot errors. Cholesky512 is 1.8–3.5× faster than 0.10.0 across the four measured hardware hosts, using nine paired process blocks.
- On Neoverse N2, ordinary GEMM128 takes 19.6% less time for f64 and 9.0% less for f32. Independent A·B fixtures corroborate the improvement. Default serial x86 GEMM128 retains its previous engine; no new speedup is claimed there.
- Optional
accelerateexposeskernix::accelerate::AccelerateExecutorfor C-contiguous rank-two f32/f64 on macOS 15+. It verifies one BLAS thread per call and restores caller policy. On the measured M1, explicit f64 GEMM128 is 16.167 µs versus NumPy 17.291; f32 still trails NumPy (6.500 versus 5.250). This is an explicit vendor backend, separate from the default Rust CPU path. - Reports separate version changes from peer observations, show paired 95% intervals, retain losses/unknowns, filter all views and provide paginated family PDFs. Actual Firefox 320/768/1366 and A4 printing checks pass.
All three crates are published from clean 18750693908110110d85d02d0f11af8d58bfae8b, also the signed tag target. Numerical crates tree: a5645528bd0f6226fa78b13b8ceac8296a9cced7. Actual registry archives match source/VCS/legal files. Renamed no_std, six native/parallel registry tests on Rust 1.86 and a wasm consumer pass. Coverage 85.81%, API checks against 0.10, 55 AddressSanitizer checks and numerical fuzzing are documented in the verification record.
Source CI, assurance, four-host same-machine comparisons and expanded fuzz pass.
Download either report ZIP below, extract it completely and open index.html. Each contains standalone HTML, PDF/PNG charts, raw samples, exact 0.10.0 baselines, paired process evidence and build/source identities. The PNG asset previews the first chart page.
Remaining gaps stay visible: faer remains faster on Cholesky, default Mac GEMM still trails NumPy, some small unchanged paths show regressions, and native panel packing increases allocation traffic. Full GEMMFIP, blocked LU/QR/CPQR and broader numerical/GPU work remain open. No universal fastest-library claim.
Final receipts: signed-tag CI, report/main CI and report/main assurance all pass. Downloaded release assets match their upload SHA-256, immutable raw/paired evidence and every offline HTML link. Main d0392f7 preserves the exact published numerical crates tree.
Presentation refresh — 2026-10-24
The complete retained benchmark archive now has a consistent design: 53 reports / 45 measurement runs / 8 paired-process studies. Download kernix-benchmark-archive.zip, extract it completely, and open index.html to search by version, hardware or measurement type. The two 0.11.0 ZIPs and PNG preview below also use the updated presentation.
Typography, direct-label latency charts, separate process intervals, full-precision tables and vector PDF/SVG exports replace the former chart layouts. Charts explicitly mark rounding with ≈; Decimal-based tables and PDF appendices retain all recorded digits. Historical batch means stay distinct from individual-call distributions.
No measurements were rerun or changed. The archive's SHA-256 manifest confirms 202 original evidence/method/build files are byte-identical. Original commentary for all reports and the archive is retained. Every original median and paired interval was checked in the PDFs; all offline links and the four actual downloaded release assets passed verification.
29 benchmark/presentation tests, fourteen executable Markdown examples and seven assurance-tool tests pass. Real Firefox checks cover 320/768/1440 widths, current/legacy/Apple/paired layouts, exact-value tables, filters, navigation and printing. Renderer source: 4d7f867; complete archive commit: b75c591. Published numerical crate bytes and version 0.11.0 remain unchanged.
Final presentation checks: archive CI, nine jobs, final main CI, nine jobs and final main assurance, four jobs pass. Main 66cee50 additionally preserves evidence line endings across platforms; renderer/data remain 4d7f867/b75c591.
Benchmark presentation refreshed on 2026-10-25: every workload now has a readable chart, direct labels and exact-value tables. These downloads preserve their original raw measurements, source archives, driver locks and methods; no historical measurement was rerun or rewritten. Open index.html locally after unzipping. The complete archive includes all 68 retained reports.
Kernix 0.10.0 — structured RZ and CPU buffer reuse
Kernix 0.10.0 optimizes existing numerical and autodiff paths while preserving the public API, generic element support, Rust 1.86 and no_std + alloc.
- Minimum-norm solves use structured implicit RZ instead of materializing a dense right QR basis. The wide fixture takes 0.759x the 0.9.0 latency; warmed Rust allocations fall from 28 to 20 and peak requested bytes from 102,144 to 59,136.
- Adjacent C-contiguous contractions avoid per-element multidimensional indexing, retaining the same scalar bounds, clone count and accumulation order. The measured fixture takes 0.059x the baseline latency, about 17x faster. NumPy remains faster on this fixture.
- Shared backward accumulation reuses private buffers. The fixture uses 8 rather than 10 allocations and 49,768 rather than 66,152 peak requested bytes. The latency interval includes no change; no stable speedup is claimed.
- Explicit two-worker reused-output matmul retains native-eligible row chunks on x86 V3. 128/256 f64 latency ratios are 0.774/0.736. These are separate from the ordinary allocating single-thread API, whose 128 f64/f32 intervals include no change.
These are nine fresh-process paired blocks against exact registry 0.9.0 with the same current driver. Every interval, loss, raw sample and allocation probe is retained. CPU frequency and shared-host load remain uncontrolled. No universal performance lead is claimed; Apple/Windows serial GEMM and factorization gaps remain.
Verification: 548 default / 551 all-feature / 541 no-default tests, 52 sanitizer checks, strict Clippy/rustdoc, MSRV and wasm/no_std consumers, API compatibility against 0.9.0, 85.78% library line coverage, bounded fuzzing, plus passing CI and 108-case evidence on Linux x86/ARM, macOS and Windows. Human-readable Markdown, searchable standalone HTML and charts retain explicit allocation/thread policies; actual Firefox filtering and footer links pass.
Verification and numerical limits · Local readable comparison · Paired comparison · Hardware workflow
The benchmark ZIP contains the current report, exact 0.9.0 baseline, paired evidence and reproduction metadata. Open reports/0.10.0-final/report.html after extracting the whole bundle. The hardware ZIP contains all four actual host reports. Crates contain library sources/tests/legal material; research and local plans are excluded.
Full GEMMFIP packing, blocked QR/CPQR/LU/Cholesky, vendor and accelerator backends remain separate work; this release implements the validated compatible subsets described above.
Published packages and signed tag use a47540c14faaef260207dd92d09c514c64528602; numerical crates tree bded13e32892c0dcae27c172b1c2ff4674284b45 matches both measured revisions. Actual downloaded registry archives (48/60/27 files), legal material and clean VCS bytes match source. Renamed no_std, native/parallel Rust 1.86 and wasm registry consumers pass using exactly all three 0.10.0 packages.
Release-source CI · Assurance · Signed-tag CI
Release closure: signed-tag CI passes all nine jobs. Final QA/main 5ad3f6187aecab5604ec6c6dd52573df3c5fcaea CI passes all nine jobs and assurance passes all four. Actual downloaded benchmark/hardware ZIP and PNG assets match local bytes; HTML footer links and raw evidence remain intact.
Benchmark presentation refreshed on 2026-10-25: every workload now has a readable chart, direct labels and exact-value tables. These downloads preserve their original raw measurements, source archives, driver locks and methods; no historical measurement was rerun or rewritten. Open index.html locally after unzipping. The complete archive includes all 68 retained reports.
Kernix 0.9.0 — pivoted QR and minimum-norm least squares
Kernix 0.9.0 adds column-pivoted QR and minimum-norm least-squares solves for rank-deficient and underdetermined systems. The rank policy is explicit; full Q/R reconstruct the input independently of truncation. A separate basic solve avoids the additional right factorization when minimum norm is unnecessary.
Both ordinary and pivoted QR now preserve Q orthogonality when subnormal norm rounding would otherwise corrupt Householder vectors. R and solves retain the input dtype and its range limitations. Existing unpivoted QR behavior, array-valued literals, Rust 1.86 and no_std + alloc remain supported. No normal equations, implicit promotion or runtime dependency was added.
Verification: 542 default / 544 all-feature / 535 no-default tests, 26 AddressSanitizer tests, MSRV/downstream consumers, independent SVD/null-space checks and fresh 85.76% library line coverage.
On the measured Intel host, minimum-norm fixtures take 49.903/76.670 µs versus NumPy 63.224/115.946 µs. NumPy uses lstsq and nalgebra uses SVD; the algorithms and rank policies differ. Large pivoted QR still trails faer (511.262 vs 466.149 µs). Nine paired process blocks show no stable GEMM 128 improvement over 0.8.0. These results do not establish universal leadership.
- Readable 100-case report, charts and raw observations
- Exact-registry 0.8.0 baseline: 88 comparable cases
- Paired current/baseline comparisons
SVD, pseudoinverse, eigenproblems and optional vendor/GPU backends remain future work. Download the reports archive for standalone searchable HTML, charts, source/lock snapshots and complete raw evidence.
Publication, registry verification and complete QA records all package hashes and limitations. Source/tag CI, assurance, bounded fuzzing and all four actual hardware jobs pass. Each hardware report contains 100 workload/library pairs.
Hosted results retain the losses: Apple M1 GEMM 128 f64 is 102.541 µs versus NumPy/Accelerate 17.667 µs under verified single-thread policy; Windows rank-16 minimum norm is 90.800 versus NumPy 85.600 µs. Pivoted QR trails faer on the hosted runners. Windows selected a different CPU than the previous release, so its changed GEMM latency is not a version speedup. Default execution remains pure Rust.
Latest readable-report fixes are verified by 14 report/evidence tests and an actual Firefox filtering/link session. Final main CI passes all nine jobs; final assurance passes all four. Published numerical source remains the signed release tag. Report rendering preserves all raw observations; downloaded release archives were checked against the final local artifacts.
Benchmark presentation refreshed on 2026-10-25: every workload now has a readable chart, direct labels and exact-value tables. These downloads preserve their original raw measurements, source archives, driver locks and methods; no historical measurement was rerun or rewritten. Open index.html locally after unzipping. The complete archive includes all 68 retained reports.
Kernix 0.8.0 — conditioning, refinement and decomposition improvements
Kernix now provides explicit one-norm conditioning and LU refinement with a correction budget, componentwise backward error and a stopping reason. Refinement stays in the input dtype; convergence and a forward-error bound are not promised. Existing tensor literal, layout and error contracts are preserved.
Cholesky reuses its input copy and improves memory traversal/accumulation. QR removes an intermediate C-order copy and returns empty thin factors without iterating enormous empty axes. QR peak requested bytes fall from 106,752 to 82,176 for 96×32 and from 557,568 to 426,496 for 256×64.
On one Intel i3-1215U, nine randomized fresh-process pairs against exact crates.io 0.7.3 show Cholesky speedups of about 1.15× / 1.48× / 2.03× at 64/128/256. QR 256×64 is about 4% faster; QR 96×32 is about 4% slower. GEMM 128 f64 is about 1.5% slower in that paired run; its engine is unchanged. Factorizations still trail the best comparators. These are exploratory, shared-machine results, not universal rankings.
Validation: 532 default / 534 all-feature / 525 no-default tests; Rust 1.86, no_std registry consumer and wasm checks; 17 ASan tests; API compatibility; actual published source/license/archive verification. All three crates are published. Library advisory count is zero; the documented, expiring benchmark-only Candle/zip exceptions and paste maintenance warning remain visible.
The attached archive contains HTML, charts, raw observations, manifests/locks and source snapshots for both versions. Extract before opening report.html.
Additional hardware reports cover AMD/Linux, Neoverse N2/Linux, AMD/Windows and Apple M1/macOS. The macOS rerun verifies Accelerate's single-thread policy through its API. These expose substantial gaps: GEMM 128 f64 is 103.166 µs versus NumPy/Accelerate 16.125 µs on M1, and 137.700 versus NumPy/OpenBLAS 100.800 µs on the Windows VM. The hardware ZIP contains all four complete reports.
Complete release verification, coverage and hardware limitations.
Benchmark presentation refreshed on 2026-10-25: every workload now has a readable chart, direct labels and exact-value tables. These downloads preserve their original raw measurements, source archives, driver locks and methods; no historical measurement was rerun or rewritten. Open index.html locally after unzipping. The complete archive includes all 68 retained reports.