Drop the coarse+fine zoom refinement, the SciPy FFT backend, and the
exact= audit path in favor of a single always-on full-transform peak
search (_peak_bins). Block size is now derived from a per-thread memory
budget (_fft_block_for/SRAS_FFT_PLAN_BUDGET_MB) instead of a fixed
constant, so the existing block-parallel PyFFTW pool stays memory-safe
at high pad factors without the zoom algorithm's bookkeeping. Also
removes the now-unused threadpoolctl dependency and the FFT backend
selector from the UI.
Also includes a pre-existing min_freq_mhz peak-search floor (excludes
bins below a caller-supplied frequency from the argmax) that was
already implemented and tested in the working tree.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- Extract cache mismatch logic into reusable cache_mismatch_reasons() function
for cleaner validation and consistent miss reporting
- Expose memory_budget_bytes() for external callers (aligned exporter)
- Optimize rebuild_stack() with incremental layer updates when only one angle
parameters change, avoiding full reprojection overhead
- Remove unused seed_per_angle parameter from alignment wizard
- Simplify cached_rf_image() DC masking with cleaner early exit logic
- Clean up internal state tracking with explicit origin caching
Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>
tests/test_stored_cache.py exercised two features that were never built,
so six of its tests had been failing on main. Both are now implemented.
Pad-factor caching. A padded view could not use a stored FFT cache at all:
the store was pad 1 by definition and cached_rf_image rejected any n_fft
outright, so a user working at a pad factor got nothing from batch-computing
a file. CACH tail v3 records the pad the images were resolved at, cache_file
computes at a requested pad, and the batch actions pass the viewer's own pad
down — while still refusing a store resolved at a different pad, since a
padded FFT interpolates between the natural bins and so resolves genuinely
different peak frequencies. v1/v2 tails read as pad 1 and keep working.
Stored-cache dispatch. _refresh_display only ever consulted this window's
in-session dicts, so after a batch every angle change still queued a worker
and a progress popup for an image already on disk — the exact cost the batch
was run to avoid. It now checks the file's own DC/FFT blocks first, asking
with allow_dc_recompute=False so the GUI thread never touches I/O. When the
stored cache genuinely cannot serve the view, the scan info panel says why
rather than leaving the silent recompute a mystery.
Two bugs surfaced on the way:
- The angle spinbox was wired on editingFinished, which QAbstractSpinBox
emits only on Return or focus-out — never on a step. Clicking its arrows,
the ordinary way to walk a scan, moved the number and left the image
behind. Now valueChanged with keyboard tracking off, which fires on a step
and once on commit, but not per keystroke mid-typing. Every existing GUI
test called _on_view_changed() by hand and so could not have caught this;
two new tests pin it and fail against the old wiring.
- A plain "fft" batch over a file previously cached with fft_rowavg carried
the old row_avg_n forward, labelling raw images as row-averaged. It now
writes row_avg_n=0 explicitly.
Verified: 111 passed (was 103 passed / 6 failed), and
tools/check_equivalence.py is byte-identical to the pre-change baseline.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Alignment was two disconnected menu actions. `Angle Alignment` ran the
registration with ref_angle_idx hard-coded to 0, no exposed parameters and
no way to retry; `Manual Alignment...` was a separate dialog that
deliberately refused to inherit the automatic result, so a bad fit meant
starting over by hand. Neither told you whether the alignment was any
good, and neither produced anything durable beyond an in-memory result.
Fusion -> Alignment Wizard... now covers all of it in three steps:
1. Correlate. Reference angle, DC threshold, source, rotation seed and
sign, search window, coarse step and fine grid are all on the page,
and Run/Re-run is repeatable. Angles start pre-rotated from the
stage angles in the file, so the page is informative before any
correlation runs. The verdict is a picture: every angle's DC mask
reprojected onto the shared canvas and summed, coloured by how many
angles cover each pixel, so a good alignment reads as one saturated
plateau and a bad one as a fringe of low-count halos. A per-angle
fit table flags angles that did not register or that disagree with
their stage angle by more than a degree.
2. Crop. An axis-aligned rectangle on that canvas, with numeric
canvas-pixel boxes synced both ways and a live size estimate.
"Fit to full overlap" uses a largest-rectangle sweep rather than a
bounding box: the overlap region of several rotated scans is roughly
a disc, whose bounding box has corners no angle covers.
3. Save. Writes the aligned, cropped stack to a new .sras.
Because the wizard replaces both actions it absorbs the old dialog's
by-eye nudge editor — without it, a scan the search cannot fit would
have no fallback at all. ManualAlignmentDialog is therefore deleted
rather than left orphaned, and its tests move to the wizard.
Two bugs found while driving it end to end and fixed here: QSpinBox
setRange clamps and emits valueChanged, which committed a 1x1 crop
before the default preset could run; and the mm round trip returns an
exact pixel boundary as 11.000000000000002, so a bare ceil() added a
spurious column on every rectangle edit.
Verified: 103 pass, the 6 test_stored_cache failures are pre-existing on
main, and tools/check_equivalence.py is byte-identical to main.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The alignment machinery never modified a scan: it produced an
AlignmentResult and every consumer resampled on the fly. That is right
for a viewer but means the aligned stack cannot leave the process, so no
other tool can read it and reopening the scan redoes the registration.
sras_align_export.write_aligned_sras bakes an alignment into a new v6
file: each angle is gathered onto the shared canvas by nearest neighbour,
so every output angle shares one grid and the file opens already aligned.
Registering the export against itself returns identity, which is the test
that pins the whole index chain.
Three details that are easy to get wrong and are now covered:
* Rounding must be floor(x + 0.5), not np.rint. scipy's order=0 rounds
halves away from zero, and the canvas is snapped to the reference's
pixel grid, so exact halves are common rather than hypothetical.
* Out-of-bounds must be tested on the fractional coordinate against
[0, n-1], not on the rounded index, or a one-pixel rim gets real data
everywhere the Aligned View shows padding.
* ...but with a tolerance, because the mm-space affine chain lands an
exactly-integer transform a few times 1e-13 off. A bare >= 0 drops
the *reference* angle's entire first row and last column.
Padding is the per-channel ADC code nearest 0 mV, not zero: zero ADC
decodes to ~+100 mV on real calibration and would masquerade as sample.
Source rows are served from sliding in-RAM bands. A rotated angle maps
one output row to a diagonal across the source, so indexing a memmap in
output order refaults nearly the whole angle per row — terabytes of
paging for a gigabyte of data.
Also in compute: crop_alignment_result (a crop is pure index translation,
so it folds into the affine's offset rather than becoming a second
transform), overlap_stats, largest_rect_at_least for a crop that stays
inside the overlap region, and seed/sign/refine knobs on
register_angle_to_reference whose defaults leave existing behaviour and
tests byte-identical.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The three block essays in sras_compute.py — memory budget / row
chunking, angle-alignment coordinate frames, and the sidecar placement
+ schema history — move to docs/design.md (joined by a new section on
the zoom FFT peak search), each replaced by a 2-4 line pointer. The
save_manual_alignment docstring no longer restates the JSON schema its
own eight lines of code construct.
Load-bearing trap notes (normalization=None, subpixel-score mixing,
memmap lazy reads) stay in place. Zero code changes — golden-hash diff
verified empty.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- SrasFile: _parse_v6 now retains per-angle x_delta and the verbatim
preamble/background byte spans, and gains public data_offset,
y_pos_per_angle, and iter_angle_blocks() (which now owns the ragged
block-offset walk used three separate places before).
- sras_edit_scans: the 70-line re-parse of the v6 header sections
(_read_v6_sections/_reread_span/_v6_angle_offsets) collapses into a
_write_v6 that consumes SrasFile directly — verified byte-identical
round-trip on v6 int8/int16 and legacy files. Eight print-and-exit
pairs become _die().
- tools/make_test_sras imports the struct layouts from sras_format and
the rotation matrix from sras_compute instead of re-declaring them
(byte assembly stays independent of the reader).
- sras_compute: block_mean_2d, pixel_pitch_mm, nominal_delta_deg made
public (they were GUI-facing); registration_workers() and
default_max_workers() wrap the remaining private reach-throughs from
sras_workers.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- FFT backend and pad factor persist across sessions via QSettings
(IniFormat; tests redirect the settings path for hermeticity).
- The default backend was labelled "NumPy FFT" but always dispatched to
scipy.fft — rename the canonical value to "scipy" ("numpy" stays as a
legacy alias) and fix the dialog label.
- cache_file: DC caching fans out over angles via _parallel_map with
per-angle budgets (the DcPrecomputeWorker pattern); FFT caching stays
serial per angle because compute_rf_image now parallelises internally
over blocks. Documented that the v7 FFT cache is natural-resolution
(pad 1) by design.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
At pad 40 the old path materialised a ~9 GB padded spectrum per row,
which collapsed the chunk planner to one worker and one rfft call with
workers=1 — synthesis ran single-threaded, ~1 hour per angle on real
files.
The padded spectrum is never materialised now. Each block of 512
waveforms gets a coarse rfft at next_fast_len(2*spf); every coarse bin
within 0.7 of its row's max (plus the DC-adjacent window, which coarse
DC suppression would otherwise blind) is refined onto the exact n_fft
grid by a small complex gemm. The selected bin is bit-identical to the
full padded argmax — enforced by test_zoom_identity, a 25-seed fuzz
test over adversarial spectra, and a clean golden-hash diff against the
pre-rewrite baseline across pads {1,2,4,8,40}, masked/unmasked, bg
on/off, int8/int16, and both backends.
Blocks fan out over a persistent thread pool; pyFFTW runs through
per-thread FFTW_MEASURE builder plans with wisdom persisted to
~/.cache/sras-viewer, and threadpoolctl clamps BLAS under the pool.
compute_rf_image(exact=True) (or SRAS_FFT_EXACT=1) keeps the reference
padded path for audits.
tools/bench_fft.py measures: pad 40, 16 cores, 8192x2500 synthetic —
exact serial 717 wf/s -> zoom pool 25100 wf/s (35x, pyFFTW backend;
19x scipy), every variant verified equal to the reference.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- pyproject.toml replaces sras_viewer_requirements.txt (same pins) and
adds a dev extra with pytest.
- tools/test_refactor.py, test_alignment.py, test_gui.py become
tests/test_compute.py, tests/test_alignment.py, tests/test_gui.py with
assertions preserved verbatim. test_gui.py stays one ordered
integration sequence over a shared module-scoped window.
- check_equivalence.py: drop the dead pre-refactor monolith shim (and the
_compute_angle_alignment alias it consumed), extend the pad sweep to
(1, 2, 4, 8, 40), add legacy-v4 and big-endian int16 legs (new bps=2
option in make_test_sras) so the padded FFT path and the >i2 memmap
path are in the baseline before the FFT rewrite.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Continues the Manual Alignment work: refines the FFT cross-correlation and
mask handling, adds sras_edit_scans.py (drop/renumber bad angle scans),
tools/test_alignment.py (registration ground-truth suite), and a rotating
test fixture in make_test_sras.py.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Manual mode previously offered only Auto De-rotate (rotation from known
scan angles) plus keyboard nudging, so every angle's translation had to
be found entirely by eye — fine for a small correction, unworkable for
the tens-of-mm scatter a real many-angle scan can have between angles
before any translation search runs at all.
Add an "Auto Cross-Correlate (vs Reference)" action that, for every
non-reference angle, sets rotation to the known analytic angle (same as
Auto De-rotate) and translation to the FFT-phase-correlation best fit
against angle #0 (always the reference/ground truth). This is meant to
land every angle's outline roughly stacked on the others so keyboard
nudging only has to make small corrections afterward, per the intended
workflow: cross-correlate first, then nudge.
Exposes two tunable options, since real data may correlate better one
way than the other: "Correlate on" (raw CH4 signal, minus its own
minimum -- the default, robust to a shared threshold not suiting every
angle's real signal level -- or the thresholded binary mask), and
"Search margin" (how far a shift the phase correlation searches before
wraparound bias becomes a risk).
Runs on a background thread (new CrossCorrelateWorker in
sras_workers.py) since correlating many high-resolution angles can take
long enough to visibly freeze the dialog otherwise.
The underlying per-angle correlation math is factored out of
compute_angle_alignment's Step 2 into a new, reusable
correlate_translation_mm, so the existing automatic Fusion -> Angle
Alignment action and the new manual button now share one implementation
instead of two copies of the same phase-correlation logic.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The pivot was still derived from the binary CH4 >= dc_threshold_mv mask,
so a threshold that happens to leave a real angle's mask empty (signal
levels vary scan to scan across a many-angle acquisition) silently fell
back to the raw scan-window bbox center — reproducing the exact "aligned
to the scan window, not the sample" scatter the centroid pivot exists to
fix, without any visible error.
Replace the binary-mask centroid with an intensity-weighted centroid of
the continuous CH4 signal, which is always well-defined regardless of the
RF-mask threshold in effect. The threshold still controls what the manual
dialog's overlay and CH1 masking show; it no longer has any bearing on
where alignment pivots.
Also bump the sidecar schema version: a file saved before today's pivot
and rotation-sign fixes stores numbers for a since-corrected transform, so
loading it unchanged would reintroduce the same scatter. Old sidecars are
now treated as absent rather than silently reused.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Rotation was pivoting on each angle's raw scan-window bbox center, which
ties alignment to wherever the window happened to sit in microscope/global
XY space rather than to the sample itself, leaving a residual orbital drift
between angles. Pivot on each angle's own CH4-threshold mask centroid
instead (its own sample footprint), for both the automatic phase-
correlation path and the manual dialog's Auto De-rotate/live preview.
Also negate the rotation used everywhere (_theta_deg): the GR stage's
reported angle increases in the opposite rotational sense from this
module's CCW math convention, so de-rotating by the raw angles_deg delta
was making misalignment worse rather than better.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Automatic alignment's phase-correlation translation search is unreliable, so
add a Fusion -> Manual Alignment dialog: every angle's CH4 threshold mask
overlaid at once with distinct colors/opacity, a selector for the active
angle, arrow keys to nudge translation and Q/E to nudge rotation, an Auto
De-rotate button that snaps to the known scan angle, and Save/Clear
controls backed by a JSON sidecar that's restored automatically on reload.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Three defects found by testing the parallel paths end to end.
Nested parallelism was unbounded. DcPrecomputeWorker and the alignment
driver both parallelise over angles, and each angle's compute_dc_image
then parallelised over rows again: 14 x 14 threads, and — worse — every
angle sized its chunk against the *whole* memory budget, so worst-case
live buffers were ~13.9 GB against a 1 GB budget. Capping the inner
worker count alone does not fix it; the chunk is still sized against the
full budget. plan_angle_level now hands each angle both max_workers=1
and its share of the budget, which bounds the real file at 1002 MB.
Chunk planning could not produce concurrency at all. _plan_chunks sized
chunk_rows first, and _chunk_rows_for spends whatever budget it is given
— so a single chunk consumed the lot and budget//chunk_bytes came back
as one worker. It now picks the worker count first and sizes the chunk to
it, giving 7-16 workers on the real file instead of 1.
Cancellation was too coarse to shut down. stop() was only checked between
angles, and one angle of the real scan is ~40 s, so closing the window
sat through it and returned with threads still running. should_stop is
now polled per row chunk and closeEvent signals every worker before
waiting: close went from 4.0 s with two live threads to 1.04 s with both
finished. A cancelled ComputeWorker emits None so a half-filled image is
never cached.
Also: BatchCacheWorker no longer reports every file as failed when the
process pool itself dies (unguarded __main__, frozen build, sandbox) —
it retries those files in-process. Batches below 512 MB skip the pool
entirely, since interpreter startup would otherwise make small jobs
slower. cache_file rejects an unknown mode instead of silently treating
it as "fft".
Measured on the real 496 GB scan, angle 3, 32 rows:
DC 2.94s -> 0.75s RF 6.10s -> 1.99s peak RSS 1.40 -> 1.93 GB
Batch convert, 24 files / 1 GB: 4.03s -> 2.36s.
Memory budget defaults to 1024 MB (SRAS_MEM_BUDGET_MB), the measured knee.
Testing: tools/test_refactor.py (70 checks: cache round-trip and block
carry-forward, partial-cache handling, parallel-vs-serial identity,
no-mask path, ROI mask vs full-grid, v2/v3/v4 parsing, sras_average
round-trip incl. remainder handling) and tools/test_gui.py (46 checks
driving the real widgets and workers headlessly). check_equivalence.py
still reports all 167 outputs identical to ed0eba4.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Structure
sras_format.py parsing/writing, calibration, axes (numpy + struct)
sras_compute.py DC/FFT images, alignment, batch cache (+ scipy)
sras_workers.py Qt background workers
sras_viewer.py ROI, canvases, dialog, main window
format+compute import in 0.59s with no Qt or matplotlib (vs 2.96s for the
full app), which is what makes a spawn-based process pool worth using.
Deduplication
- SrasFile.cal() and compute.dc_image_mv() replace the ADC->mV
calibration incantation that appeared at ten call sites.
- _read_preambles/_read_background/_set_calibration are shared by the
legacy and v6 parsers instead of duplicated.
- v5 PREC images are normalised into the same ragged per-angle list
v6/v7 uses, removing four isinstance() branches and shortening
compute_rf_image's fast path.
- _run_worker replaces five copies of the QThread setup and five
near-identical teardown methods; the two lifetime hazards they
guarded against are now documented once, authoritatively.
- _on_channel_changed defers to _update_controls_enabled rather than
re-deriving the same six enable rules.
- _corners_in_ref_frame, _CHANNEL_DISPLAY dict, unified progress-dialog
helper, dead _on_roi_angle_edited stub removed.
- sras_average.py builds on SrasFile instead of carrying a second copy
of the header format, and streams per angle instead of loading the
whole file (was a ~3x file-size peak).
Executable lines: 2390 -> 2254, despite adding all of the below.
Parallelism
- compute_rf_image/compute_dc_image map row chunks over a thread pool.
Worker count is derived from the memory budget rather than the core
count: on a large scan chunk_rows is already floored at 1 row (~75 MB
of float32 at 7507x2500), so only the worker count can bound peak RAM.
- DcPrecomputeWorker computes angles on a pool, emitting each result
from its own QThread as it lands.
- BatchCacheWorker runs one process per file via cache_file(); only
paths and scalars cross the boundary. Per-process thread counts are
divided so the two levels don't oversubscribe. Drops the old pass 1,
which fully parsed every file just to weight a progress bar.
- dc_threshold_mv=None means "no mask" and skips the CH4 read entirely
— the batch FFT job previously read all of CH4 to compare against a
threshold of -1e9.
- Alignment mask and correlation stages map over angles; fft2/ifft2 use
workers=-1.
Fixes found on the way
- A partially-cached v5 file showed an all-zero image for uncached
angles: the fast-path check was file-wide, not per-angle.
- v7 cached images were read-only big-endian views; now native float32.
Verified: tools/check_equivalence.py produces byte-identical hashes for
167 outputs (DC/FFT across channels, angles, bg-sub, pad factors and
thresholds, plus the full alignment result) against ed0eba4, on both
synthetic files and a real 496 GB 17-angle v6 scan.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Thomas Ales [M S E]