9a7cc396e7126d02c8d7a9c4f97e9226f0712328
2 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Thomas Ales
|
9a7cc396e7 |
Bound nested parallelism, make compute cancellable, harden batch pool
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
|
||
|
Thomas Ales
|
55a4c5e42f |
Split into four modules, deduplicate, and parallelize the compute paths
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
|