Files
Jungfraujoch/image_analysis/spot_finding/SpotExtractorGPU.h
T
leonarski_fandClaude Opus 5 4bdb229fb8
Build Packages / build:viewer-tgz:cpu (push) Successful in 7m46s
Build Packages / build:viewer-tgz:cuda (push) Successful in 9m14s
Build Packages / build:rpm (ubuntu2404_nocuda) (push) Successful in 13m51s
Build Packages / build:rpm (rocky8_nocuda) (push) Successful in 14m17s
Build Packages / build:rpm (ubuntu2204_nocuda) (push) Successful in 14m14s
Build Packages / build:rpm (rocky9_nocuda) (push) Successful in 14m43s
Build Packages / build:rpm (rocky8_sls9) (push) Successful in 14m45s
Build Packages / build:rpm (rocky8) (push) Successful in 11m44s
Build Packages / build:rpm (rocky9_sls9) (push) Successful in 13m24s
Build Packages / XDS test (durin plugin) (push) Successful in 8m33s
Build Packages / Generate python client (push) Successful in 28s
Build Packages / Build documentation (push) Successful in 1m4s
Build Packages / Create release (push) Skipped
Build Packages / build:rpm (rocky9) (push) Successful in 12m45s
Build Packages / build:rpm (ubuntu2204) (push) Successful in 12m25s
Build Packages / build:rpm (ubuntu2404) (push) Successful in 13m1s
Build Packages / DIALS test (push) Successful in 14m29s
Build Packages / XDS test (neggia plugin) (push) Successful in 8m17s
Build Packages / XDS test (JFJoch plugin) (push) Successful in 9m5s
Build Packages / Unit tests (push) Successful in 1h16m19s
Build Packages / build:windows:nocuda (push) Failing after 2s
Build Packages / build:windows:cuda (push) Failing after 3s
spot_finding: find connected components on the GPU
The spot finder flagged strong pixels on the device and then labelled them on the
host, so every frame sent the packed bitmask back - 2.26 MB on a large detector -
and the host walked all of it to recover a few hundred pixels. Do the labelling on
the device instead: compact the bitmask into a flat-index-sorted list, find each
pixel's backward neighbours by binary search, union them lock-free with path
halving, then label, accumulate and filter in one kernel. Only the spot list comes
back, and only one stream synchronisation per frame.

The gain in the ordinary case is modest - about a quarter off per-image spot
finding - because the host algorithm is genuinely fast on a normal frame. What
justifies it is the frame that is not ordinary. The host labels a sorted sparse
list through a window spanning two detector lines, so its cost is quadratic in how
many strong pixels share a line. A lit band of detector rows - a hot module, a
panel edge - costs 33 ms at two rows and 377 ms at fifteen, all of it under the
pixel cap that was supposed to bound this, and none of it maskable when the cause
is a diffraction ring rather than a defect: a ring runs tangent to a row at its
top and bottom, which is exactly the shape that hurts. The device version is flat
at 0.05 to 0.64 ms across every geometry tried, so an online run no longer stalls
a quarter of a second on an ice ring. Rejecting an over-cap frame is now free too,
since the count is known before any pixel is written.

Also label once and filter three times. The per-image minimum-pixel search runs the
extraction at three settings, but that setting only decides which components are
kept - it does not change the components - so the search itself need not be
repeated. This helps the host path as much as the device one.

The resolution mask moves to the device as a bit mask, uploaded when the limits
change rather than per frame, since the compaction needs it there.

Parity is asserted permanently rather than argued: five cases covering realistic
frames, occupancy from a hundred pixels to past the cap, the pathological
geometries including rings, the resolution mask, and a hundred-repeat determinism
check - requiring the same partition, the same spot order, and identical counts.
The centroid is a float sum and therefore order-dependent, so the device walks each
component from its root in ascending order and fuses its multiply-add the way the
host's does; note that whether the host fuses at all depends on the architecture
flags, so exact centroid equality is asserted where the compiler fuses and a
two-ulp bound otherwise. Making those accumulators integer would remove that
dependence entirely and is worth doing separately.

Regression set: all 37 crystals identical to the last printed digit. Unit suite
passes with the new cases.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-02 13:34:43 +02:00

92 lines
4.4 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 ascending list order, in float, term for
// term as DiffractionSpot::AddPixel does them, with the rounding spelled out (see the comment at
// the sum itself - the two compilers do not contract a multiply-add alike).
// 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 "../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 {
float x;
float y;
int64_t photons;
int64_t max_photons;
int32_t pixel_count;
int32_t padding;
};
class SpotExtractorGPU {
std::shared_ptr<CudaStream> stream;
const int32_t width;
const size_t nwords;
// The connected-component search gives up above this many strong pixels (see
// StrongPixelSet::FindComponentsImage), so nothing larger is ever built.
static constexpr uint32_t MAX_STRONG = UINT16_MAX;
// 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<int32_t> gpu_count; // pixels per component
CudaDevicePtr<SpotExtractorGPUSpot> gpu_spot;
CudaDevicePtr<SpotExtractorGPUSpot> gpu_spot_out;
CudaDevicePtr<uint32_t> gpu_nspot;
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);
void SetResolutionMask(const std::vector<uint32_t> &packed_mask);
// 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);
};