Prepare cuvarbase 1.0.1 with verified external benchmark archives - #69

Open

johnh2o2 wants to merge 539 commits into

This pull request is big! We're only showing the most recent 250 commits

Commits on Jul 3, 2026

Adjust phases in custom kernel after normalizing time stamps
astrobatty
committed
17228ca
Add optimized option to transit_gpu function
astrobatty
committed
e28c5cf
Add tests for optimized kernels
astrobatty
committed
3b5175e
Merge branch 'bugfix/BLS-kernel' of https://github.com/johnh2o2/cuvarbase into bugfix/BLS-kernel
astrobatty
committed
b2d4ba2
Update eebls_gpu_custom's docstring to match results
astrobatty
committed
4e0d414
Adjust phase offset in sparse_bls_cpu to use input time scale
astrobatty
committed
a578d6d
Adjust phase offset in single_bls to use input time scale
astrobatty
committed
e9c34b0

Commits on Jul 4, 2026

Test fixture: non-zero T0 (t += 4.5) in data(); update CPU tests to original-timescale phi convention

johnh2o2
and
claude
committed
290ade5
Add v0.2.6-vs-v1.0 head-to-head benchmark + summarizer (protocol S4 fairness rules)

johnh2o2
and
claude
committed
8942477
Fix PR #65 GPU/CPU parity: single_bls fold order, custom-kernel fold precision, sparse f64 phase conversion

johnh2o2
and
claude
committed
c76db0f
Head-to-head raw results checkpoint: BLS cold/warm/loop/correctness + decomposition

johnh2o2
and
claude
committed
cfcf125
v0.2.6 vs v1.0.0 head-to-head: full results, LS/PDM, SUMMARY (RTX A5000, Jul 2026)

johnh2o2
and
claude
committed
fa7cf2c
PR #65 resolution: CHANGELOG, analysis archive (root cause + evidence + pod validation), docstring touch-up

johnh2o2
and
claude
committed
a7110c0
PR #65 resolution: force-add raw validation logs (gitignored *.log)

johnh2o2
and
claude
committed
ed545da
Merge pull request #65 from johnh2o2/bugfix/BLS-kernel
johnh2o2
authored
462a483
Merge v0.2.6 head-to-head benchmark campaign (measured baseline for v1.0.0 release notes)
johnh2o2
committed
1944bbb
Release prep: outward docs sweep

johnh2o2
and
claude
committed
ff326c2
Add survey-scale BLS benchmark + profiler driver scripts

johnh2o2
and
claude
committed
f332e05
profile driver: --freq-stride to keep ncu replay affordable

johnh2o2
and
claude
committed
2aef7dc
Add parity gate comparator (corr>0.999 + identical peak, bit-eq flag)

johnh2o2
and
claude
committed
2e335e0
profile driver: add naive variant (full public path for host-overhead tracing)

johnh2o2
and
claude
committed
af1db85
Kernel hygiene: resolve the float32 PI literals flagged by the A3 diagnosis

johnh2o2
and
claude
committed
f677a15
Docs: BLS data-hygiene note on near-zero uncertainties

johnh2o2
and
claude
committed
e132f59
bench: pin BLAS threadpools + record CFS throttle counters

johnh2o2
and
claude
committed
4bec10e
Profile evidence + measured bottleneck ranking (RTX A5000)

johnh2o2
and
claude
committed
e88741d
Opt 1: fused-noverlap BLS kernels (1.8-2.7x kernel time at survey scale)

johnh2o2
and
claude
committed
4970430
Opt 2: conflict-scatter permutation of staged LC data (TESS kernel 3.0x)

johnh2o2
and
claude
committed
3c9e6f7
Opt 3: host-path overhead -- einsum, np.max, reusable batch memory

johnh2o2
and
claude
committed
95f335b
Opt 4: occupancy-aware frequency chunking (Kepler kernel 1.32x)

johnh2o2
and
claude
committed
f7d336b
Campaign deliverables: SUMMARY.md, CHANGELOG, opt2-4 benchmark evidence

johnh2o2
and
claude
committed
aef8e90
Archive per-change full-suite + release-gate logs (opt1-opt4; force-add, *.log gitignored)

johnh2o2
and
claude
committed
8e872b9

Commits on Jul 6, 2026

TLS: survey-scale fast path (batch-native phase-binned kernel + exact refinement)

johnh2o2
and
claude
committed
bcf4710
TLS: measured fidelity — SDE parity with reference, matched-fidelity timing

johnh2o2
and
claude
committed
052e69a

Commits on Jul 7, 2026

TLS: apples-to-apples GTLS Fig.7 reproduction (30-171x faster at equal SDE)

johnh2o2
and
claude
committed
7e30192
TLS/GTLS: add cold single-shot (one-star, compile-included) comparison

johnh2o2
and
claude
committed
4ebf681
docs: README performance pass — TLS-vs-GTLS + corrected BLS vs-0.2.6 numbers

johnh2o2
and
claude
committed
0f7917a
TLS: fix UnboundLocalError in tls_search_batch refinement fallback (PR #68 review)

johnh2o2
and
claude
committed
78f1965
Merge pull request #67 from johnh2o2/fix/kernel-hygiene-jul2026
johnh2o2
authored
82f321f
Merge pull request #66 from johnh2o2/feature/bls-survey-speed
johnh2o2
authored
610a72a
Merge pull request #68 from johnh2o2/feature/tls-fast-survey
johnh2o2
authored
d0f7a8c

Commits on Jul 11, 2026

TLS audit (PR #68 pre-release): findings doc + fix-now items

johnh2o2
and
claude
committed
dca5de6
Docs: repair the Sphinx source tree (first buildable state since 2017)

johnh2o2
and
claude
committed
c566add
Release content: TLS becomes the v1.0.0 headline; claims-trace corrections

johnh2o2
and
claude
committed
08c13cb
Release staging: runbook, issue-sweep drafts, coordination message

johnh2o2
and
claude
committed
6f467e8
Release gate PASSED: 796/796 zero-skip suite + 14/14 gate on merged tip

johnh2o2
and
claude
committed
865bede
README: cut to a third of its length (401 -> 132 lines)

johnh2o2
and
claude
committed
ca24a80
Runbook: post-publish README install flip + consistency-test update

johnh2o2
and
claude
committed
8d7ff50
NUFFT-LRT audit (Taaki contribution): findings doc + fix-now items

johnh2o2
and
claude
committed
648a901
NUFFT-LRT validation harness: null-calibrated injection-recovery vs BLS/TLS

johnh2o2
and
claude
committed
ac8b0ec
Validation harness: adaptive epoch grids, on-grid injection, flat-PSD arm, size overrides

johnh2o2
and
claude
committed
335914b
NUFFT-LRT: implement Detector A (marginalized joint detector) + sequential baseline

johnh2o2
and
claude
committed
c8a93a6
NUFFT-LRT: drop the Lomb-Scargle arm from the validation analysis

johnh2o2
and
claude
committed
7c99ea9
Validation harness: 32 trial periods (4h -> ~1.5h campaign at equal coverage of the transition)

johnh2o2
and
claude
committed
89ad23b

Commits on Sep 4, 2026

Commits on Sep 5, 2026

BLS: page-lock only the transfer buffers and pool BLSMemory on the naive fast path (BLS-6)

johnh2o2
and
claude
committed
61ab589
LS: numpy reductions instead of Python builtins on the host path (LS-1)

johnh2o2
and
claude
committed
21161a9
CE/PDM: no unused histogram on the fast path, reuse PDM's device buffers (CE-3 / PDM-1)

johnh2o2
and
claude
committed
c2fa53f
BLS: einsum instead of np.dot in the remaining per-light-curve prologues (BLS-8)

johnh2o2
and
claude
committed
4a03f46
TLS-1: tls_transit builds only the duration bounds, not the table

johnh2o2
and
claude
committed
42516b8
TLS-2: run tls_search_batch's per-lightcurve statistics sequentially

johnh2o2
and
claude
committed
a7d0960
TLS-3: memoize the fast kernel's template tables

johnh2o2
and
claude
committed
435d29e
LS: reuse memory, plans and the grid check across batched calls (LS-4)

johnh2o2
and
claude
committed
3d07ae2
LS: one stacked (nf, 2H, 2H) solve for the multiharmonic power (LS-5)

johnh2o2
and
claude
committed
fdb847b
BLS: vectorized solution re-phasing, chi2_0 from yy, cached scatter permutation (BLS-9)

johnh2o2
and
claude
committed
a87d1be
BLS: solve the Keplerian frequency-grid recursion with numpy (BLS-4) [RESULT-CHANGING]

johnh2o2
and
claude
committed
dce8c0b
LS: key the batched memory cache off the per-call kwargs (LS-4 fix round 0)

johnh2o2
and
claude
committed
6e1cf0c
docs: memory reuse and float32 reproducibility in the Lomb-Scargle guide

johnh2o2
and
claude
committed
bad3010
Merge p2-bls: kernel-cache routing, fused adaptive path, memory pooling, host-path vectorization (Phase 2)

johnh2o2
and
claude
committed
6b0d11f
Merge p2-tls: q_transit in tls_transit, no thread pool in the batch path, memoized template tables (Phase 2)

johnh2o2
and
claude
committed
bdab67b
LS: close the two majors from the Phase 2 verification

johnh2o2
and
claude
committed
43026de
Merge p2-ls: numpy host reductions, cached memory sets, stacked multiharmonic solve (Phase 2)

johnh2o2
and
claude
committed
93c1059
Merge p2-cepdm: occupancy-sized CE fast grid, skipped CE histogram, pooled PDM allocations (Phase 2)

johnh2o2
and
claude
committed
58820f8
CE/PDM: the code-level minors from the Phase 2 verification

johnh2o2
and
claude
committed
5e2890d
CHANGELOG: Phase 2 performance work, with the verifiers' corrections applied

johnh2o2
and
claude
committed
2dd12d4
Repo prune: delete the pre-1.0 process material listed in RELEASE_READINESS section 3

johnh2o2
and
claude
committed
1a532c7
Repo prune: move the GTLS/TLS-cost writeups to docs/, PROFILE_RANKING to its campaign, the a5000 JSONs into benchmarks/results/

johnh2o2
and
claude
committed
fd25ce9
Repo prune: keep only the base_envfix/opt4_chunk parity dumps (8 of 24 .npz)

johnh2o2
and
claude
committed
1a2e234
docs: delete the TLS_GPU process docs and the Dockerfile with its mentions

johnh2o2
and
claude
committed
ef752c2
scripts: merge the benchmark and RunPod how-tos into scripts/README.md; drop the dead scikit-cuda patch blocks

johnh2o2
and
claude
committed
47c49f2
Repo prune: analysis/README.md index and the exec bit on every shebang script under scripts/ and benchmarks/

johnh2o2
and
claude
committed
c21a93c
docs: phi0 is reported on the original input timescale (bls.rst)

johnh2o2
and
claude
committed
b1cbafa
docs: INSTALL/CONTRIBUTING/README consistency (CUDA 12.4, 3.9-3.14, pycuda-less scope, workflow, API policy)

johnh2o2
and
claude
committed
1e563e3
NFFT: precomp_psi=False grids with the inline-psi kernel instead of raising

johnh2o2
and
claude
committed
b937ecf
LS: Baluev d_K in batched_run_const_nfreq follows the per-call nharmonics

johnh2o2
and
claude
committed
09eaed7
CHANGELOG: the two Phase 2 verification carry-over fixes (NFFT, LS)

johnh2o2
and
claude
committed
68d03a8
Packaging: pyproject.toml is the single source of metadata

johnh2o2
and
claude
committed
68cf407
Remove the orphan wavelet.cu kernel and guard the kernel inventory

johnh2o2
and
claude
committed
d518780
Repo metadata: track the gate logs, add .gitattributes and .mailmap

johnh2o2
and
claude
committed
7015bbb
CI: 3.9-3.14 matrix, read-only permissions, wheel/sdist legs, docs job

johnh2o2
and
claude
committed
f3bc3bf
CHANGELOG/release notes: Phase 3 packaging, CI and repo-metadata entries

johnh2o2
and
claude
committed
8cdd2aa
analysis: point the two kept records at the moved GTLS/TLS-cost docs and pdm_a5000.json

johnh2o2
and
claude
committed
9986ae3
tests: move the pycuda-stub conftest into cuvarbase/tests (ships in the wheel)

johnh2o2
and
claude
committed
1c55f0d
Runbook: rewrite for the Sep-2026 release sequence (Phases 4-7)

johnh2o2
and
claude
committed
f4de96d
Release staging: refresh the @astrobatty coordination draft for September 2026

johnh2o2
and
claude
committed
00c93d1
Release staging: draft the @xiaziyna message about NUFFT-LRT in 1.0

johnh2o2
and
claude
committed
bd1ee3c
Release staging: update the issue-sweep drafts for the Sep-2026 state

johnh2o2
and
claude
committed
9a48799
API freeze: frozen top-level namespace, NUFFT-LRT quarantined

johnh2o2
and
claude
committed
d6ea7bb
tests: wheel-safe layout checks, scoped nfft importorskip, no swallow-all BLS tests, flake8 F cleanup

johnh2o2
and
claude
committed
972dbb5
docs: release notes -- Ada/V100 TLS ratios, measured cache numbers, LS-only batch_size, NUFFT-LRT status, Sep-2026 audit section

johnh2o2
and
claude
committed
b26b37a
docs: BENCHMARK_RESULTS -- July 2026 BLS survey-throughput table, 5-13x, post-prune paths

johnh2o2
and
claude
committed
da0fbcf
docs: Sphinx -- nufft_lrt (experimental) and base.context on the API page, real landing page, plot hygiene

johnh2o2
and
claude
committed
652253e
docs: NUFFT-LRT page as docs/source/nufft_lrt.rst; CHANGELOG line edits; example pointer

johnh2o2
and
claude
committed
f3fa360
API freeze: compatibility shims per decision D3 (kept for 1.x, removed in 2.0)

johnh2o2
and
claude
committed
24b45db
API freeze: remove what never shipped (decision D3) + the three 0.2.5 utils helpers

johnh2o2
and
claude
committed
d1134ba
docs: do not index the lazily re-exported package members on the API page

johnh2o2
and
claude
committed
8aef1e1
Repo prune: verifier fixes (NUFFT-LRT run pointer, archive notes, notebook version)

johnh2o2
and
claude
committed
a6a8b15
API freeze: keyword-only markers on the 1.0-new signatures (finding 136)

johnh2o2
and
claude
committed
680fada
API freeze: explicit __all__ in every user-facing module (finding 135)

johnh2o2
and
claude
committed
d6fcc33
tests: port the stranded script checks, new entry-point tests, cufinufft device test, examples compile check, -rs and gate preflight

johnh2o2
and
claude
committed
082c8a7
Source hygiene: unused imports/locals, 'resource', shebang, mutable defaults (section 5)

johnh2o2
and
claude
committed
3fba9fc

Commits on Sep 6, 2026

Commits on Sep 10, 2026

Make observation-level TLS the default and publish audited benchmarks
johnh2o2
committed
b834d4e

Commits on Sep 28, 2026