Files
Jungfraujoch/image_analysis/spot_finding/SpotExtractorGPU.h
T
leonarski_fandClaude Opus 5.5 0a515bb086 rc179: everything on rc178 after its merge into main (#88), except the CHANGELOG
The rc178 branch continued after #88 was squash-merged; this is its content from b9ada7914 (the tree
#88 squashed) to 094eee31a, as one change on main. The individual commits stay on the rc178 branch:

- still battery: --cold evicts each set's input from the page cache; Alvra pixel masks; manifest fixes
- AGIPD: converter to one rigid panel per module, int16 photons, mask folded in; still battery reads the converted sets
- DetectorModules: panel look-up by id, plane -> panel, plane bounds and a per-pixel walk; mask size checked
- tests: [PanelLayout] equivalence harness, stacked vs assembled panel layout
- Dense per-pixel maps through the panel table (panel layout stage 1)
- Panel id on spots and the panel table in the start message
- Viewer and preview show a panel detector as it lies on the detector plane
- Panel layout stage 3: per-panel spot finding (CPU+GPU, CCL) and raster algorithms on the plane raster
- Panel layout stage 2: prediction and Bragg integration through the panel table
- XTC reader: frames handed up as stored, streams indexed in parallel
- XFEL reader: integer frames stay integer, one gain, placement by panel
- EXTERNAL_TEST_DATA_STILLS.md: images used vs the entry for every set; drop "Not in the battery"
- still battery: EuXFEL sets read the depositors' Cheetah CXI files directly; AGIPD converter removed
- float32_to_int16.py: a non-finite pixel becomes -32768, the int16 error marker, not a zero count
- CrystFEL geometry: read the pixel masks and bad regions instead of ignoring them
- Panel layout stage 5: the XFEL reader hands up the stored frame and its panel table
- Panel table persisted in NXmx; modulation grid and pre-scan memo key on the plane
- still battery: Alvra sets take their pixel mask from the CrystFEL geometry, not --mask-good
- still battery: take the deposited model as mmCIF where the RCSB has no PDB-format file
- still battery: dimple gets PDB-format coordinates converted from an mmCIF-only deposition
- stills fixed cost: geometry refinement and pre-scan do the same work in less wall time
- stills geometry sample and trial spot finding run as many workers as the image pass
- Stills integration: background ring reach measured on empty apertures, radial curvature from the whole frame
- StillsPartialityRefine: the partiality width is the run's, measured on the crystals, not the prediction window's
- MergeOnTheFly: capture term and multiplicative outlier band for the stills merge (both opt-in)
- Stills given a cell: the written cell is the mean over the frames that are merged
- Stills: the radial curvature correction decided per run on the same empty apertures
- StillsPartialityRefine: measure the run width only on shells with signal
- Stills pre-scan: report the paired z of the radial-correction comparison
- module placement: the five concurrent fits each keep every thread
- still battery: score R-free against the deposited data through the same dimple protocol
- still battery: the R-free target is the lower of the deposition's dimple and published values
- Stills: background clip follows the radial curve; signal disk from the lattice spacing
- Stills keep an overloaded observation as a lower bound instead of dropping it
- Stills: drop the lattice-spacing signal disk; the correction decision back on the plain clip
- battery: R-free against the deposition without refinement only, one protocol for rotation and stills
- still battery: EuXFEL sets run on a 2,000-image subset for now
- stills geometry refinement: fit the detector distance and the cell together
- stills geometry refinement: a distance given with --detector-distance is held in the joint fit
- Metal integration engine: back to its panel-s5 kernels; Mac stills keep the old clip, no saturation count
- Metal integration test: the CPU comparison runs on a rotation experiment
- Stills: kept overloads go back to the drop where they are not concentrated at low resolution
- stills distance search: an absolute 30-3000 mm scan, a sample found without d-limits, the fit iterated
- battery results: XFEL still battery 20261022-1358 on rc178 0cf62452c - 12 sets, cold, R_free unrefined
- battery results: still summary gains the resolution of our merge and of the deposition
- Pre-scan masks a pixel at the saturation limit on most frames, as one holding the error value
- Stills keep an overload as a lower bound only where it has pixels at the limit; run-level gate removed
- XFEL reader: a per-frame mask is read the way the frames are, its chunk inflated outside the HDF5 lock
- stills distance search: every centre scanned with a wider acceptance, a larger sample where nothing indexes
- rugnux usage: --beam-center-check describes the absolute 30-3000 mm stills distance scan, not the old 60-160 % grid
- stills: detector tilt refined in the geometry pre-pass, kept where held-out frames agree
- CLAUDE.md: the changelog is written once, just before the merge into main - condensed one-liners of what users get
- broker: serve the placed model and its maps of the last rugnux live run (/rugnux/model.cif, /rugnux/map.ccp4)
- frontend: Mol* viewer of the placed model and its density on the Rugnux live page
- writer: _model.cif from rugnux live keeps the input model's metadata, as the served model.cif does
- rugnux live maps: one unit cell centred on the model; the viewer opens on all of it
- Anomalous difference map: ANOM/PHANOM in _maps.mtz, /rugnux/map.ccp4?type=anom, a magenta mesh in the viewer
- Mol* viewer: open with the maps drawn within 10 A of the centre of view
- Mol* viewer: aaredb's structure-viewer setup, on a black stage
- rugnux: the gain of a stills file that states none is read off its single-photon peak where the background resolves photons
- Mol* viewer: in line with AareDB's StructureView at its current head (8ece2c2)
- ACKNOWLEDGEMENT: the web frontend's Mol* viewer follows AareDB's setup (G. Gotthard, PSI)
- stills tilt decision: the modules stay where the run placed them
- still battery: no distance pins and no per-set tuning flags; gain from the data except where a file's baseline hides it
- battery results: rc178 6f6f69 - full still battery and full rotation battery with REFMAC
- Bragg integration: a CCD rotation sweep's sigmas at the measured gain, intensities in photons
- BackgroundGain: the gain of a frame that resolves photons is the spacing of its photon peaks, so a baseline left in the frame no longer hides it
- ShadowFinder: measure the background's azimuthal modulation instead of predicting it from the polarization
- BackgroundGain: one gain estimate for rotation and stills; the still battery no longer gives EuXFEL 152 its gain
- ShadowFinder: measure the modulation robustly under a wide one-sided shadow
- ShadowFinderTest: a one-sided shadow is not fitted as part of the background's modulation
- ShadowFinder: back to the predicted polarization factor, in the plane the background shows
- MXAnalysisAfterFPGA: the Bragg engine's edge mask in the image's layout, so raw / untransformed FPGA runs start again
- battery results: still battery on rc178 e54f55 replaces the earlier rc178 still tables
- XFELReader: per-frame mask and photon conversion without a library call or a branch per pixel
- stills geometry trial: the hypotheses a search asks are indexed side by side
- stills tilt decision: the fits start beside the pre-scan and the geometry trial
- XFELReader: a deflated frame is inflated and converted a piece at a time, while it is in the cache
- XFELReader: a byte-shuffled frame is unshuffled a piece at a time, a whole pixel per step
- GeometryRefiner: residuals carry only the blocks a fit moves; beam and distance one block
- GeometryRefiner: a spot's residual differentiates its observed and predicted halves apart
- docs: EXTERNAL_TEST_DATA lists every open-arm battery set, with corrected values and counts
- battery: doc_check.py cross-checks the manifests against the dataset pages
- stills module placement: the placement fit gets half the threads, each held-out fit an eighth
- still battery: the EuXFEL sets run on all their frames, as the published still runs did
- battery results: still battery on rc178 a0ce6c, cold on a quiet machine, replaces the earlier rc178 table
- macOS build: explicit size_t in the XFEL reader's row bounds; parentheses in the CPU preprocessor's select

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015CeEGgpKdfkCTJb3YNMJcu
2026-10-23 17:56:42 +02:00

110 lines
5.6 KiB
C++

// SPDX-FileCopyrightText: 2026 Filip Leonarski, Paul Scherrer Institute <filip.leonarski@psi.ch>
// SPDX-License-Identifier: GPL-3.0-only
#pragma once
// Device-side connected-component extraction for the GPU spot finders.
//
// The GPU finders flag strong pixels into a packed bit buffer ON THE DEVICE. Reading spots out of it
// used to mean copying that whole buffer back (2.26 MB per frame at 18 MP) and scanning it bit by bit
// on the host. This does the whole extraction where the data already is, so nothing about the image
// comes back - only the finished spot list, a few hundred entries.
//
// The algorithm is the sparse formulation the ACTS/traccc project settled on for the same problem
// (sparse silicon-detector hits): the strong pixels are compacted into a list that is sorted by flat
// index, each pixel finds its at most FOUR backward 8-neighbours by binary search in that list, and
// the resulting graph is labelled with a lock-free union-find. A dense image-wide labelling
// (Playne-equivalence, BUF/BKE, nppiLabelMarkers, cv::cuda::connectedComponents) would label 18
// million pixels to find five hundred.
//
// It reproduces the host StrongPixelSet::sparseccl EXACTLY, not just equivalently:
// * both make a component's root its lowest list index, so both find the same roots;
// * labels are handed out by a prefix sum over the roots in ascending order, which is the order the
// host's second scan hands them out in, so the SPOT ORDER is identical;
// * the centroid sums are accumulated per component in integers, term for term as
// DiffractionSpot::AddPixel does them, so there is no rounding for the two compilers to disagree
// about, and the order the device adds them in changes nothing.
// tests/SpotExtractorGPUParityTest.cpp holds the two to each other on realistic, occupancy-swept and
// pathological frames, and checks that repeating a frame gives byte-identical output.
#include <cstdint>
#include <memory>
#include <vector>
#include "../../common/DiffractionSpot.h"
#include "../../common/PanelBoxes.h"
#include "../indexing/CUDAMemHelpers.h"
#include "SpotFindingSettings.h"
// Per-component sums, in exactly the form DiffractionSpot holds them: x and y are sum(col*photons)
// and sum(line*photons), not a centroid.
struct SpotExtractorGPUSpot {
int64_t x;
int64_t y;
int64_t photons;
int64_t max_photons;
int32_t pixel_count;
// Longer side of the component's bounding box, for SpotShapeAccepted. It fits in what used to be
// padding, so carrying it costs nothing.
int32_t bbox_side;
int32_t panel; // PanelBox::panel of the component's pixels
};
class SpotExtractorGPU {
std::shared_ptr<CudaStream> stream;
const int32_t width;
const size_t nwords;
// Strong pixels this engine's buffers hold, and above which the extraction gives up on the frame -
// StrongPixelLimit, so it follows the detector rather than standing at a constant. 116 bytes of
// device memory apiece, 32 MB on an 18-megapixel detector.
const uint32_t max_strong;
// Spots copied back together with their count in one transfer. A frame with more than this many
// surviving spots - far past anything indexable - simply takes a second copy.
static constexpr uint32_t SPOT_PREFIX = 4096;
int compact_blocks = 0;
CudaDevicePtr<uint32_t> gpu_res_mask; // packed, bit set = pixel excluded
CudaDevicePtr<uint32_t> gpu_block_count;
CudaDevicePtr<uint32_t> gpu_block_offset;
CudaDevicePtr<uint32_t> gpu_nstrong;
CudaDevicePtr<uint32_t> gpu_index; // strong pixels, sorted by flat index
CudaDevicePtr<int32_t> gpu_value;
CudaDevicePtr<uint32_t> gpu_parent; // union-find parent
CudaDevicePtr<uint32_t> gpu_root;
CudaDevicePtr<uint32_t> gpu_label; // compact label, indexed by root
CudaDevicePtr<int4> gpu_box; // bounding box per component
CudaDevicePtr<uint32_t> gpu_nlabel; // components found
CudaDevicePtr<SpotExtractorGPUSpot> gpu_spot;
CudaDevicePtr<SpotExtractorGPUSpot> gpu_spot_out;
CudaDevicePtr<uint32_t> gpu_nspot;
CudaDevicePtr<PanelBox> gpu_panels; // the panels components stay inside
int npanels = 0;
CudaHostPtr<uint32_t> host_nstrong;
CudaHostPtr<uint32_t> host_nspot;
CudaHostPtr<SpotExtractorGPUSpot> host_spot; // SPOT_PREFIX entries, pinned
std::vector<SpotExtractorGPUSpot> overflow_spot; // only for a frame with more spots than that
public:
SpotExtractorGPU(int32_t width, int32_t height, std::shared_ptr<CudaStream> stream);
// Waits for the stream: its pooled buffers are freed next (see CudaDevicePtr).
~SpotExtractorGPU() { if (stream) cudaStreamSynchronize(*stream); }
void SetResolutionMask(const std::vector<uint32_t> &packed_mask);
// The panels of the image (PanelBoxes); default one, the whole image.
void SetPanels(const std::vector<PanelBox> &panels);
// gpu_strong is the finder's device bit buffer, gpu_image the preprocessed image it was built
// from. Fills spots with every component of at most max-pix pixels, in the same order the host
// extractor would.
void Extract(const uint32_t *gpu_strong, const int32_t *gpu_image,
const SpotFindingSettings &settings, std::vector<DiffractionSpot> &spots);
// Strong pixels the last Extract() saw, after the resolution mask. Reported whether or not the
// frame was given up on, which is the point of it: a frame at or above max_strong yields no spots
// at all, and this is what says so.
[[nodiscard]] uint32_t StrongPixelCount() const { return *host_nstrong.get(); }
};