We read every piece of feedback, and take your input very seriously.
To see all available qualifiers, see our documentation.
Add GPU Speedup Scaling Model wiki page
Updated gels QR vs Normal Equation Cholesky (markdown)
Performance-Benchmarks: add GalapagosSenDT128 large-scene row Adds a new "## GalapagosSenDT128" section after the Fernandina table and links report_large_scene.md @ 019ceba. Headlines: 36.4× step wall / 44.4× internal speedup on --dostep invert_network (cpu 6189 s → torch 170.06 s), abs RMS max ~16 µm. Code under test: MintPy fork 11b94388. Closes Issue #6 and the wiki entry for it.
Add GPU Quick Start wiki page Reproducible recipe for verifying the [gpu] extras install path and the torch solver end-to-end on FernandinaSenDT128, distilled from the Issue #5 docs-surfaces verification run. Calls out the cfg-basename gotcha that breaks correct_LOD with KeyError: 'PLATFORM' if the project name omits a recognised sensor substring (Sen / Env / Alos / ...). Assisted-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Performance-Benchmarks: add solver comparison row (cholesky vs lstsq, 2026-05-03)
Performance-Benchmarks: add 2026-05-03 torch.profiler kernel breakdown row Adds the kernel-level breakdown finding from mintpy-benchmark cea7573: gels QR is host-launch-overhead bound (~1.88M micro-kernel launches per chunk for our 19,403-pixel chunk), motivating the Phase 2 (#4) switch to batched Cholesky on the normal-equation form. The previous py-spy row stays at 45dd50e since each entry pins the SHA at which its findings were committed. Also extends the Methodology section to cover the torch.profiler harness (run_profile_torch.sh + profile_torch.py + parse_trace.py) and the shared lib/setup_ulimit.sh safety net. Assisted-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add gels QR vs. Normal Equation + Cholesky reference page Captures the linear-algebra reasoning and the library-implementation trade-offs behind the Phase 2 (#4) decision to switch the per-pixel WLS solver from cusolver gels to batched Cholesky on the normal equation form. Separates the mathematical layer (cond squaring, unavoidable) from the implementation layer (cusolver's batched parallelism for Cholesky vs. per-problem for QR), so future design decisions are not confused. Links into the empirical profile data (mintpy-benchmark/report_profile.md) and the Phase 2 tracking issue. Assisted-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Performance-Benchmarks: refresh links after history rewrite + sibling repo split Bench assets are now in s-sasaki-earthsea-wizard/mintpy-benchmark (sibling repo). Pre-rewrite SHAs in the MintPy fork no longer exist; updated all links to post-rewrite SHAs and added entries for torch backend, chunk sweep, and profile reports. Assisted-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add Upstream PR Checklist page Distill upstream's CONTRIBUTING.md into a fork-aware checklist: what changes belong upstream vs stay fork-only, branch naming differences, pre-commit + overall test requirements, and the rebase-on-upstream/main sync flow. Assisted-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add Performance Benchmarks index page Link to the FernandinaSenDT128 baseline report committed at 85ae4f2 (PR #1) and document the benchmark methodology. Reports are kept in-repo and pinned by SHA so figures stay in sync with the code that produced them.
Initial Home page