Add the production-density 4K Galactic-center comparison to the 2026-09-25 benchmark record, including the per-image error decomposition, and refine the fast-mode positioning in usage.md, the design document, and README.
- Document that nearest/bilinear deposit accuracy is set by the deposit mode and N rather than output resolution, and that the still-frame error is dominated by a small number of resolved bright cores while the flux-dominant faint texture is much better than the global relative L2.
- Add the temporal-coherence caveat: nearest deposition jumps by 1/N output pixel per axis across supersampled-cell boundaries, so fast mode is not qualified as a temporally coherent movie path; bilinear removes the centroid jump but broadens the profile.
- Add scripts/fast_mode_error_decomposition.py, which ranks the largest |fast-base| RGB samples and partitions the error by a disjoint base-value range.
Replace the nested spatial global convolution in fast_psf_accumulator_resolve with a reusable double-precision FFTW linear convolution on the CPU PSF backend.
- Add the private src/fast_psf_fftw.{c,h} module: zero-padded R2C/C2R plans, cached kernel spectrum and planar scratch, exact 1/(Pwidth*Pheight) and 1/N^2 normalization, (R,R) crop, and additive HDR output.
- Keep the previous nested loops as fast_psf_accumulator_resolve_spatial_reference for tests/benchmarks only; it is not a runtime fallback.
- Cache the circular row spans on FastPsfAccumulator and report one-time plan, kernel transform, scratch, and per-frame stage timings.
- Require fftw3_omp for CPU builds; HIP and dummy builds do not link FFTW.
- Namespace test/helper binaries by spacetime and build tag, and reject make test / psf-capture for non-CPU backends.
- Add tests/test_fast_psf_fftw.c (FFTW versus spatial), tests/benchmark_fast_psf_fftw.c, an FFTW CLI smoke check, and the 2026-09-25 benchmark record.
Add an optional CPU preview path that deposits each point-source image as a
supersampled delta and resolves the whole frame with one global Moffat
convolution plus an N x N box average, instead of splatting a per-event PSF.
- optics: FastPsfAccumulator builds the pixel-area-integral kernel of the
target Moffat at the supersampled scale (width N*alpha, same beta). The
1/N^2 box average then reproduces the final pixel-area integral, so the
requested FWHM and beta are preserved without renormalisation. Deposits are
per-cell atomic adds; resolve accumulates into the caller's HDR buffer.
- frame: fast branch in frame_splat_catalog with one shared supersampled
buffer and a single resolve per frame; the accumulator is reused across
movie frames and built from the map dimensions on lens-map import.
- main: --fast-mode, --fast-supersample N (1..8, default 2) and
--fast-deposit nearest|bilinear (default nearest). CPU-only and rejected in
the HIP/dummy backends; --psf-min-y still applies per event while
--max-cache-psf-flux does not.
- The deposition scheme was chosen by scripts/fast_mode_deposit_error.py:
nearest keeps the PSF shape exactly with <= 0.5/N px position quantization;
bilinear keeps the exact centroid but broadens FWHM and beta. Recorded in
benchmarks/fast_mode_deposit_2026-09-18.md.
- tests/test_frame.c covers fast nearest vs the direct evaluator at the snapped
centre, flux conservation, bilinear centroid, min-Y discard, frame plumbing,
and HDR accumulation onto a non-zero background.
- benchmarks/fast_mode_cpu_2026-09-18.md records a ~10x speedup on the 2MASS
galactic-centre field with small tone-mapped differences.
Document the producer-private adaptive path, the dummy support workload
diagnostics, the negative result of the old serial binning and the
2026-09-13 full-day comparison (splat 282.944 s vs atomic 481.417 s).
Add the corresponding raw logs, hotspot fixtures and runner evidence,
and lay out the staged plan toward a sub-240 s full-day render without
changing the default atomic path.
Pin OMP_DYNAMIC=FALSE so thread counts stay deterministic, record it in
the command line, and locate the binary hash correctly when the command
is wrapped in env with inline assignments.
Preserve the protected full-catalog dummy run, its initial harness failure, input and output hashes, distribution quantiles, and the resulting sparse-tile scheduling conclusions.
Reuse production catalog mapping and per-worker chunk boundaries to report triangle and spatial distributions without initializing HIP or writing image outputs. Include a protected full-catalog runner and build documentation.
Add a bounded per-chunk center-tile selector to the real-event replay prototype and retain the paired timing, mixed-boundary, CPU, HIP, and sandbox-failure evidence.
Archive fixed catalog subsets and captured events, manifests, reproducible commands, raw timing and regression logs, and device resource evidence. Document distribution-dependent tile reduction gains and independent indexed producer throughput.
Update the GPU plan and renderer design while retaining the production atomic path. Preserve historical measurement metadata and exclude generated Python bytecode.
Preserve user-provided complete CPU and HIP logs and associate the optimized implementation with e1ec480669. Record matching event counts, user-confirmed bit-identical PNG output, and the 2.77x workflow speedup with tracing/import differences noted.
Mark the pre-artifact-fix CPU benchmark as historical and link the new paired result.
Cooperate across 32 lanes per PSF and use two completion-protected staging slots with complete batch timing. Restore coarse OpenMP event production while serializing shared GPU submissions and direct fallback boundaries.
Add bounded benchmarks, streaming and renderer regressions, and preserve validation evidence and ownership documentation.
Schedule ray-pool advancement in dynamic chunks of 32 slots to distribute frame-grouped active rays across workers while preserving stable endpoint destinations.
Add 1/4/16-thread lens-map equivalence coverage and document schedule comparisons with complete commands, input data and original logs. The 393-frame 64x36 production-source benchmark improves from 62.13 to 38.69 seconds with byte-identical PNGs and lens maps. Release and Debug test suites pass.
wyj