wyj
3deebfb2fa
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.
129 lines
8.1 KiB
Markdown
129 lines
8.1 KiB
Markdown
# Fast-mode deposition error experiment (2026-09-18)
|
||
|
||
Decision this note settles: whether `--fast-mode` should deposit each point
|
||
source image as a single nearest supersampled pixel or as 4 adjacent
|
||
(bilinear) pixels before the one global PSF convolution and downscale.
|
||
|
||
## Command
|
||
|
||
```sh
|
||
python3 scripts/fast_mode_deposit_error.py --phases 12 --grid 48 --tail 1e-6 --max-radius 64
|
||
```
|
||
|
||
## Model
|
||
|
||
Reference = pixel-area integral of the normalized circular Moffat
|
||
`M(r; α, β)`, `α = FWHM / (2 sqrt(2^(1/β) - 1))`, on the final grid, using
|
||
8-point Gauss-Legendre quadrature. Fast image with supersample factor `N`:
|
||
|
||
- deposit the event into the `N x N` supersampled grid (nearest = 1 cell,
|
||
bilinear = 4 adjacent cells);
|
||
- convolve the whole ss buffer with the single global kernel
|
||
`k[m] = ∫_ss-cell m M(z; Nα, β) dz` (4-point quadrature);
|
||
- downscale by averaging each `N x N` block, which together with the
|
||
unnormalised `N`-scaled kernel reproduces the final pixel-area integral.
|
||
|
||
Metrics: `maxErr`/`rmsErr` are whole-image errors vs the direct image at the
|
||
true position; `coreErr` is the worst central-3x3 pixel error; `FWHMerr`
|
||
compares the fast image with a direct image at the same centroid, so it is a
|
||
pure shape error (0 when deposition is exact at its snapped centre); `centErr`
|
||
is the position quantization vs the true continuous star position; `beterr` is
|
||
the deposition-induced Moffat beta change, isolated from the radial fit's own
|
||
pixel-integration bias by comparing with a reference fit (reported only for
|
||
FWHM >= 2 px). FWHM 1.0 / β 3.5 rows are sub-Nyquist and are not meaningful
|
||
for the shape metrics.
|
||
|
||
## Raw output
|
||
|
||
```text
|
||
Fast-mode deposition error experiment
|
||
grid=48 phases=12 tail=1e-06 max_radius=64
|
||
kernel quadrature = 4-point Gauss; reference = 8-point Gauss
|
||
|
||
FWHM beta N deposit R_ss retain maxErr rmsErr coreErr centErr FWHMerr beterr
|
||
---------------------------------------------------------------------------------------------------
|
||
1.00 8.00 2 nearest 10 1.00000 6.333e-01 1.714e-02 6.333e-01 3.043e-01 1.733e-02 nan
|
||
1.00 8.00 2 bilinear 10 1.00000 1.805e-01 3.972e-03 1.805e-01 2.722e-03 3.187e-01 nan
|
||
1.00 8.00 3 nearest 14 1.00000 3.907e-01 1.061e-02 3.907e-01 1.847e-01 7.491e-03 nan
|
||
1.00 8.00 3 bilinear 14 1.00000 7.587e-02 1.676e-03 7.587e-02 5.562e-03 1.816e-01 nan
|
||
1.00 8.00 4 nearest 18 1.00000 2.608e-01 7.142e-03 2.608e-01 1.237e-01 3.665e-02 nan
|
||
1.00 8.00 4 bilinear 18 1.00000 4.428e-02 9.775e-04 4.428e-02 3.395e-03 3.646e-02 nan
|
||
|
||
2.70 8.00 2 nearest 24 1.00000 1.635e-01 8.693e-03 1.635e-01 2.946e-01 2.325e-08 5.061e-07
|
||
2.70 8.00 2 bilinear 24 1.00000 4.339e-02 1.476e-03 4.339e-02 6.018e-09 4.046e-02 8.414e-02
|
||
2.70 8.00 3 nearest 35 1.00000 9.597e-02 5.229e-03 9.597e-02 1.768e-01 1.465e-08 1.067e-06
|
||
2.70 8.00 3 bilinear 35 1.00000 1.842e-02 6.387e-04 1.842e-02 2.437e-08 1.747e-02 4.029e-02
|
||
2.70 8.00 4 nearest 46 1.00000 6.308e-02 3.488e-03 6.308e-02 1.179e-01 1.563e-08 2.905e-05
|
||
2.70 8.00 4 bilinear 46 1.00000 1.011e-02 3.418e-04 1.011e-02 1.778e-08 9.355e-03 3.091e-02
|
||
|
||
6.00 8.00 2 nearest 51 1.00000 7.012e-02 7.714e-03 7.012e-02 2.946e-01 1.617e-06 3.132e-06
|
||
6.00 8.00 2 bilinear 51 1.00000 9.563e-03 6.146e-04 9.563e-03 8.749e-08 6.686e-03 1.039e-02
|
||
6.00 8.00 3 nearest 64 1.00000 4.199e-02 4.631e-03 4.199e-02 1.768e-01 7.456e-10 3.315e-08
|
||
6.00 8.00 3 bilinear 64 1.00000 4.090e-03 2.640e-04 4.090e-03 9.147e-06 2.683e-03 4.463e-03
|
||
6.00 8.00 4 nearest 64 0.99994 2.795e-02 3.088e-03 2.795e-02 1.179e-01 2.631e-07 9.889e-06
|
||
6.00 8.00 4 bilinear 64 0.99994 2.195e-03 1.409e-04 2.195e-03 9.262e-06 1.528e-03 2.354e-03
|
||
|
||
1.00 4.50 2 nearest 19 1.00000 6.006e-01 1.643e-02 6.006e-01 3.041e-01 1.516e-02 nan
|
||
1.00 4.50 2 bilinear 19 1.00000 1.745e-01 3.823e-03 1.745e-01 2.673e-03 2.485e-01 nan
|
||
1.00 4.50 3 nearest 28 1.00000 3.712e-01 1.016e-02 3.712e-01 1.846e-01 6.881e-03 nan
|
||
1.00 4.50 3 bilinear 28 1.00000 7.354e-02 1.616e-03 7.354e-02 5.429e-03 1.432e-01 nan
|
||
1.00 4.50 4 nearest 36 1.00000 2.480e-01 6.842e-03 2.480e-01 1.236e-01 3.167e-02 nan
|
||
1.00 4.50 4 bilinear 36 1.00000 4.269e-02 9.383e-04 4.269e-02 3.322e-03 3.640e-02 nan
|
||
|
||
2.70 4.50 2 nearest 49 1.00000 1.625e-01 8.537e-03 1.625e-01 2.946e-01 9.993e-07 1.210e-06
|
||
2.70 4.50 2 bilinear 49 1.00000 4.413e-02 1.439e-03 4.413e-02 5.868e-07 4.453e-02 4.843e-02
|
||
2.70 4.50 3 nearest 64 1.00000 9.571e-02 5.135e-03 9.571e-02 1.768e-01 1.482e-07 2.493e-07
|
||
2.70 4.50 3 bilinear 64 1.00000 1.871e-02 6.228e-04 1.871e-02 3.673e-06 1.867e-02 2.057e-02
|
||
2.70 4.50 4 nearest 64 0.99999 6.299e-02 3.426e-03 6.299e-02 1.179e-01 1.817e-07 4.228e-07
|
||
2.70 4.50 4 bilinear 64 0.99999 1.030e-02 3.333e-04 1.030e-02 3.726e-06 9.995e-03 1.353e-02
|
||
|
||
6.00 4.50 2 nearest 64 0.99998 6.924e-02 7.554e-03 6.924e-02 2.945e-01 1.290e-04 1.735e-04
|
||
6.00 4.50 2 bilinear 64 0.99998 9.859e-03 6.016e-04 9.859e-03 2.462e-06 7.468e-03 9.748e-03
|
||
6.00 4.50 3 nearest 64 0.99978 4.153e-02 4.535e-03 4.153e-02 1.775e-01 2.110e-08 3.706e-08
|
||
6.00 4.50 3 bilinear 64 0.99978 4.214e-03 2.585e-04 4.214e-03 7.436e-04 3.203e-03 4.204e-03
|
||
6.00 4.50 4 nearest 64 0.99872 2.767e-02 3.024e-03 2.767e-02 1.185e-01 3.310e-06 1.891e-04
|
||
6.00 4.50 4 bilinear 64 0.99872 2.264e-03 1.485e-04 2.264e-03 7.451e-04 1.709e-03 2.399e-03
|
||
|
||
1.00 3.50 2 nearest 35 1.00000 5.809e-01 1.600e-02 5.809e-01 3.038e-01 1.370e-02 nan
|
||
1.00 3.50 2 bilinear 35 1.00000 1.708e-01 3.732e-03 1.708e-01 2.601e-03 2.192e-01 nan
|
||
1.00 3.50 3 nearest 52 1.00000 3.594e-01 9.895e-03 3.594e-01 1.843e-01 6.403e-03 nan
|
||
1.00 3.50 3 bilinear 52 1.00000 7.210e-02 1.579e-03 7.210e-02 5.255e-03 1.252e-01 nan
|
||
1.00 3.50 4 nearest 64 1.00000 2.402e-01 6.660e-03 2.402e-01 1.234e-01 2.661e-02 nan
|
||
1.00 3.50 4 bilinear 64 1.00000 4.170e-02 9.143e-04 4.170e-02 3.223e-03 3.644e-02 nan
|
||
|
||
2.70 3.50 2 nearest 64 1.00000 1.619e-01 8.440e-03 1.619e-01 2.946e-01 2.790e-05 2.602e-05
|
||
2.70 3.50 2 bilinear 64 1.00000 4.461e-02 1.417e-03 4.461e-02 3.036e-07 4.689e-02 3.892e-02
|
||
2.70 3.50 3 nearest 64 0.99997 9.553e-02 5.077e-03 9.553e-02 1.769e-01 3.421e-07 3.302e-07
|
||
2.70 3.50 3 bilinear 64 0.99997 1.889e-02 6.133e-04 1.889e-02 8.121e-05 2.016e-02 1.646e-02
|
||
2.70 3.50 4 nearest 64 0.99989 6.293e-02 3.387e-03 6.293e-02 1.179e-01 8.537e-07 5.529e-06
|
||
2.70 3.50 4 bilinear 64 0.99989 1.042e-02 3.282e-04 1.042e-02 8.142e-05 1.066e-02 1.038e-02
|
||
|
||
Worst case over all tested configurations, by deposit scheme:
|
||
nearest: maxErr=6.333e-01 rmsErr=1.714e-02 coreErr=6.333e-01 centErr=3.043e-01 FWHMerr=3.665e-02 beterr=1.891e-04
|
||
bilinear: maxErr=1.805e-01 rmsErr=3.972e-03 coreErr=1.805e-01 centErr=5.562e-03 FWHMerr=3.187e-01 beterr=8.414e-02
|
||
```
|
||
|
||
## Conclusion
|
||
|
||
For the production PSF (FWHM 2.7, β 4.5):
|
||
|
||
| N | deposit | core HDR err | image RMS | centroid err | FWHM err | beta change |
|
||
|---|---|---|---|---|---|---|
|
||
| 2 | nearest | 16.3% | 0.85% | 0.295 px | ~0 | ~1.2e-6 |
|
||
| 2 | bilinear | 4.4% | 0.14% | ~0 | +4.45% | +4.8% |
|
||
| 3 | nearest | 9.6% | 0.51% | 0.177 px | ~0 | ~2.5e-7 |
|
||
| 3 | bilinear | 1.9% | 0.06% | ~0 | +1.87% | +2.1% |
|
||
| 4 | nearest | 6.3% | 0.34% | 0.118 px | ~0 | ~4.2e-7 |
|
||
| 4 | bilinear | 1.0% | 0.03% | ~0 | +1.0% | +1.4% |
|
||
|
||
- Nearest preserves the specified FWHM/β exactly (kernel-only shape and beta
|
||
error at the numerical floor); its only error is a bounded position
|
||
quantization of `0.5/N` output pixels.
|
||
- Bilinear recovers the sub-pixel centroid exactly and roughly quarters the
|
||
HDR error, but broadens the profile because it interpolates the kernel
|
||
samples rather than the continuous kernel; it therefore changes the
|
||
specified FWHM and beta unless separately compensated.
|
||
- A single fixed global convolution cannot provide both. `--fast-mode` default
|
||
is therefore `nearest`; `--fast-deposit bilinear` is available as an
|
||
explicit position-over-shape approximation.
|