Files
Jungfraujoch/image_analysis/indexing/PostIndexingRefinement.cpp
T
leonarski_fandClaude Opus 5 6e805f53c0
Build Packages / build:viewer-tgz:cpu (push) Successful in 8m17s
Build Packages / build:viewer-tgz:cuda (push) Successful in 9m11s
Build Packages / build:rpm (ubuntu2404_nocuda) (push) Successful in 13m38s
Build Packages / build:rpm (rocky8_nocuda) (push) Successful in 13m57s
Build Packages / build:rpm (rocky8_sls9) (push) Successful in 13m57s
Build Packages / build:rpm (ubuntu2204_nocuda) (push) Successful in 14m13s
Build Packages / build:rpm (rocky9_nocuda) (push) Successful in 14m15s
Build Packages / build:rpm (rocky8) (push) Successful in 11m22s
Build Packages / build:rpm (rocky9_sls9) (push) Successful in 12m51s
Build Packages / XDS test (durin plugin) (push) Successful in 7m56s
Build Packages / Generate python client (push) Successful in 32s
Build Packages / Build documentation (push) Successful in 1m4s
Build Packages / Create release (push) Skipped
Build Packages / build:rpm (rocky9) (push) Successful in 13m23s
Build Packages / build:rpm (ubuntu2404) (push) Successful in 13m15s
Build Packages / build:rpm (ubuntu2204) (push) Successful in 13m53s
Build Packages / DIALS test (push) Successful in 14m21s
Build Packages / XDS test (neggia plugin) (push) Successful in 8m36s
Build Packages / XDS test (JFJoch plugin) (push) Successful in 9m16s
Build Packages / Unit tests (push) Successful in 1h15m16s
Build Packages / build:windows:nocuda (push) Failing after 2s
Build Packages / build:windows:cuda (push) Failing after 2s
image_analysis: stop paying for work that is thrown away
Three independent costs, each measured, none changing a result. Across the
37-crystal regression set the run time halves (median per crystal 2.0x, total
2.3x) and every crystal's merge statistics are unchanged.

The image copy back from the device moved the whole preprocessed frame - 72 MB
on a large detector, every frame, per worker - to serve a single host consumer
that reads only the strong pixels, at most a few hundred kilobytes of it. Give
the buffer a Gather() so that consumer asks for the values it actually wants (a
host loop on the CPU, a small kernel on the GPU), and copy the frame back only
when a CPU spot finder will genuinely read it. The copy the other way was worse:
it came from an unregistered vector, so the driver staged it through its own
pinned pool with a host-side memcpy on the calling thread, which does not overlap
and collapses under concurrency - 11.6 GB/s at one worker, 1.6 GB/s at eight.
That, not any hardware limit, is why throughput stopped improving past four to
eight workers. Pinning the decompression buffer once per worker fixes it: on a
18 Mpx dataset the image loop goes from 13.6 to 7.9 ms per image at 32 workers,
and 32 workers now beat 8 instead of losing to them.

Ceres was computing seventeen partial derivatives where five are free. The
per-image rotation refinement frees the beam and the orientation and holds
distance, detector angles, rotation axis and cell constant, but the cost
function declared all seven blocks, so every residual evaluated in Jet<17>
arithmetic. A residual exposing only the two free blocks - the same arithmetic,
the constants baked in - halves refinement, and it is exact rather than merely
close: dual coordinates evolve independently, so the residuals and the free
Jacobian columns are unchanged bit for bit.

The merge sorted an index array with a comparator that dereferenced a 1.6 GB
array of 72-byte records, i.e. a random walk over memory, single-threaded, twice
per two-pass run. Sorting a packed key instead is 2.4x. French-Wilson allocated
its integration scratch per reflection and ran serially; it now takes caller-owned
scratch and runs over chunks, 4.2x. The correction surfaces re-tested every
observation for usability and parity on each of ~22 passes and re-allocated their
accumulators each time; bucket the indices once and hoist the buffers.

Also convert std::round to std::rint where the rounded value only ever enters a
squared residual. The tie rules differ - away from zero against to even - so this
is safe exactly where a tie flips the sign but not the magnitude, and unsafe
wherever the value becomes a Miller index; those sites keep std::round. Verified
over all 2^32 float bit patterns: 8388608 exact ties exist, and the squared
residual is bitwise equal for every one of them. Worth little on its own here,
because the rounding that dominates is in candidate refinement, where the value
is an index and the substitution is not available.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-01 21:35:28 +02:00

296 lines
12 KiB
C++

// SPDX-FileCopyrightText: 2025 Filip Leonarski, Paul Scherrer Institute <filip.leonarski@psi.ch>
// SPDX-License-Identifier: GPL-3.0-only
#include "../../common/JFJochMath.h"
#include "PostIndexingRefinement.h"
#include <iostream>
namespace {
struct config_ifssr final {
float threshold_contraction = .8; // contract error threshold by this value in every iteration
float max_distance = .00075; // max distance to reciprocal spots for inliers
unsigned min_spots = 8; // minimum number of spots to fit against
unsigned max_iter = 32; // max number of iterations
};
static std::pair<float, float> score_parts(float score) noexcept {
float nsp = -std::floor(score);
float s = score + nsp;
return std::make_pair(nsp - 1, s);
}
struct RefinedCandidate {
Eigen::Matrix3f cell;
float score;
float volume;
int64_t indexed_spot_count;
std::vector<uint8_t> indexed_mask;
};
static inline Eigen::MatrixX3<float> CalculateResiduals(
const Eigen::Ref<const Eigen::MatrixX3<float>> &spots,
const Eigen::Matrix3f &cell) {
Eigen::MatrixX3<float> miller = (spots * cell).array().round().matrix();
Eigen::MatrixX3<float> resid = miller * cell.inverse();
resid -= spots;
return resid;
}
static inline std::vector<uint8_t> ComputeIndexedMask(
const Eigen::Ref<const Eigen::MatrixX3<float>> &spots,
const Eigen::Matrix3f &cell,
float indexing_tolerance,
int64_t &indexed_spot_count) {
const float indexing_tolerance_sq = indexing_tolerance * indexing_tolerance;
// Compute fractional Miller indices. rint (round half to even) rather than round (round half
// away from zero): without SSE4.1 Eigen has no vector round, so each element is a libm call,
// while rint is a few inline instructions. Only the SQUARED residual is taken below and the two
// rules can differ only at an exact .5, where either leaves |frac| = 0.5 - so the mask and the
// count are the same. The refinement loop above keeps round: there the rounded value IS the
// Miller index that goes into the residual and the QR solve, so its tie rule does matter.
Eigen::MatrixX3<float> miller_frac = spots * cell;
Eigen::MatrixX3<float> miller_int = miller_frac.array().rint().matrix();
Eigen::MatrixX3<float> frac_resid = miller_frac - miller_int;
std::vector<uint8_t> mask(spots.rows(), 0);
indexed_spot_count = 0;
for (int i = 0; i < spots.rows(); ++i) {
if (frac_resid.row(i).squaredNorm() < indexing_tolerance_sq) {
mask[i] = 1;
indexed_spot_count++;
}
}
return mask;
}
template<typename MatX3, typename VecX>
static void RefineCandidateCells(const Eigen::Ref<const Eigen::MatrixX3<float>> &spots,
Eigen::DenseBase<MatX3> &cells,
Eigen::DenseBase<VecX> &scores,
const config_ifssr &cifssr,
unsigned block = 0, unsigned nblocks = 1) {
using namespace Eigen;
using Mx3 = MatrixX3<float>;
using M3 = Matrix3<float>;
const unsigned nspots = spots.rows();
const unsigned ncells = scores.rows();
VectorX<bool> below{nspots};
MatrixX3<bool> sel{nspots, 3u};
Mx3 resid{nspots, 3u};
Mx3 miller{nspots, 3u};
M3 cell;
const unsigned blocksize = (ncells + nblocks - 1u) / nblocks;
const unsigned startcell = block * blocksize;
const unsigned endcell = std::min(startcell + blocksize, ncells);
for (unsigned j = startcell; j < endcell; j++) {
if (nspots < cifssr.min_spots) {
scores(j) = float{1.};
continue;
}
cell = cells.block(3u * j, 0u, 3u, 3u).transpose(); // cell: col vectors
const float scale = cell.colwise().norm().minCoeff();
float threshold = score_parts(scores[j]).second / scale;
for (unsigned niter = 1; niter < cifssr.max_iter && threshold > cifssr.max_distance; niter++) {
miller = (spots * cell).array().round().matrix();
resid = miller * cell.inverse();
resid -= spots;
below = (resid.rowwise().norm().array() < threshold);
if (below.count() < cifssr.min_spots)
break;
threshold *= cifssr.threshold_contraction;
sel.colwise() = below;
HouseholderQR<Mx3> qr{sel.select(spots, .0f)};
cell = qr.solve(sel.select(miller, .0f));
}
resid = CalculateResiduals(spots, cell);
ArrayX<float> dist = resid.rowwise().norm();
auto nth = std::begin(dist) + (cifssr.min_spots - 1);
std::nth_element(std::begin(dist), nth, std::end(dist));
scores(j) = *nth;
cells.block(3u * j, 0u, 3u, 3u) = cell.transpose();
}
}
}
std::vector<CrystalLattice> Refine(const std::vector<Coord> &in_spots,
size_t nspots,
Eigen::MatrixX3<float> &oCell,
Eigen::VectorX<float> &scores,
RefineParameters &p) {
std::vector<CrystalLattice> ret;
Eigen::MatrixX3<float> spots(in_spots.size(), 3u);
for (int i = 0; i < in_spots.size(); i++) {
spots(i, 0u) = in_spots[i].x;
spots(i, 1u) = in_spots[i].y;
spots(i, 2u) = in_spots[i].z;
}
config_ifssr cifssr{
.min_spots = static_cast<uint32_t>(p.viable_cell_min_spots)
};
RefineCandidateCells(spots.topRows(nspots), oCell, scores, cifssr);
std::vector<RefinedCandidate> candidates;
for (int i = 0; i < scores.size(); i++) {
Eigen::Matrix3f cell_rows = oCell.block(3u * i, 0u, 3u, 3u);
Eigen::Matrix3f cell_cols = cell_rows.transpose();
Eigen::Vector3f row_norms = cell_rows.rowwise().norm();
if (p.reference_unit_cell) {
std::array<float, 3> obs = {row_norms(0), row_norms(1), row_norms(2)};
std::array<float, 3> ref = {
static_cast<float>(p.reference_unit_cell->a),
static_cast<float>(p.reference_unit_cell->b),
static_cast<float>(p.reference_unit_cell->c)
};
std::sort(obs.begin(), obs.end());
std::sort(ref.begin(), ref.end());
bool lengths_ok = true;
for (int k = 0; k < 3; ++k) {
const float denom = std::max(ref[k], REFINE_MIN_REFERENCE_LENGTH_EPSILON);
const float rel_dev = std::abs(obs[k] - ref[k]) / denom;
if (rel_dev > p.dist_tolerance_vs_reference) {
lengths_ok = false;
break;
}
}
if (!lengths_ok)
continue;
// Also require the angles to match the reference. Fold each to its acute complement
// (min(x,180-x)) so the obtuse/acute setting choice is irrelevant, then compare the
// sorted triples. Guards against a right-edges/wrong-angle cell (a pseudo-symmetric
// near-metric, e.g. a monoclinic beta refined to the wrong value) passing on lengths.
auto fold = [](float deg) { return std::min(deg, 180.0f - deg); };
auto row_angle = [&](int i, int j) {
return std::acos(std::clamp(cell_rows.row(i).normalized().dot(cell_rows.row(j).normalized()),
-1.0f, 1.0f)) * 180.0f / PI;
};
std::array<float, 3> obs_ang = {fold(row_angle(1, 2)), fold(row_angle(0, 2)), fold(row_angle(0, 1))};
std::array<float, 3> ref_ang = {
fold(static_cast<float>(p.reference_unit_cell->alpha)),
fold(static_cast<float>(p.reference_unit_cell->beta)),
fold(static_cast<float>(p.reference_unit_cell->gamma))
};
std::sort(obs_ang.begin(), obs_ang.end());
std::sort(ref_ang.begin(), ref_ang.end());
bool angles_ok = true;
for (int k = 0; k < 3; ++k) {
if (std::abs(obs_ang[k] - ref_ang[k]) > REFINE_ANGLE_TOLERANCE_VS_REFERENCE_DEG) {
angles_ok = false;
break;
}
}
if (!angles_ok)
continue;
} else {
if (row_norms.minCoeff() < p.min_length_A || row_norms.maxCoeff() > p.max_length_A)
continue;
}
// Filter for wrong angles
float alpha = std::acos(cell_rows.row(1).normalized().dot(cell_rows.row(2).normalized())) * 180.0f / PI;
float beta = std::acos(cell_rows.row(0).normalized().dot(cell_rows.row(2).normalized())) * 180.0f / PI;
float gamma = std::acos(cell_rows.row(0).normalized().dot(cell_rows.row(1).normalized())) * 180.0f / PI;
if (alpha < p.min_angle_deg || alpha > p.max_angle_deg ||
beta < p.min_angle_deg || beta > p.max_angle_deg ||
gamma < p.min_angle_deg || gamma > p.max_angle_deg)
continue;
int64_t indexed_spot_count = 0;
auto indexed_mask = ComputeIndexedMask(spots.topRows(nspots), cell_cols, p.indexing_tolerance, indexed_spot_count);
if (indexed_spot_count < p.viable_cell_min_spots)
continue;
candidates.emplace_back(RefinedCandidate{
.cell = cell_rows,
.score = scores(i),
.volume = std::abs(cell_rows.determinant()),
.indexed_spot_count = indexed_spot_count,
.indexed_mask = std::move(indexed_mask)
});
}
std::sort(candidates.begin(), candidates.end(),
[](const RefinedCandidate &a, const RefinedCandidate &b) {
const auto max_spots = std::max(a.indexed_spot_count, b.indexed_spot_count);
const auto min_spots = std::min(a.indexed_spot_count, b.indexed_spot_count);
const bool spot_counts_close = (max_spots > 0)
&& (static_cast<float>(min_spots) / static_cast<float>(max_spots)
>= REFINE_CANDIDATE_SPOT_COUNT_RATIO_THRESHOLD);
if (!spot_counts_close)
return a.indexed_spot_count > b.indexed_spot_count;
const float max_volume = std::max(a.volume, b.volume);
const float min_volume = std::max(std::min(a.volume, b.volume), REFINE_MIN_VOLUME_EPSILON);
const bool volume_differs = (max_volume / min_volume) > REFINE_CANDIDATE_VOLUME_RATIO_THRESHOLD;
if (volume_differs)
return a.volume < b.volume;
if (a.score != b.score)
return a.score < b.score;
return a.indexed_spot_count > b.indexed_spot_count;
});
std::vector<RefinedCandidate> accepted;
for (const auto &candidate: candidates) {
int64_t overlap = 0;
// Check all already selected lattices and see how many spots are already indexed for the candidate
// If the overlap is more than 40% of indexed spots - we assume the lattice doesn't bring anything new
for (const auto &selected: accepted) {
for (size_t i = 0; i < candidate.indexed_mask.size(); ++i) {
if (candidate.indexed_mask[i] && selected.indexed_mask[i])
overlap++;
}
}
if (overlap < static_cast<int64_t>(REFINE_CANDIDATE_OVERLAP_RATIO_THRESHOLD
* static_cast<float>(candidate.indexed_spot_count))) {
accepted.emplace_back(candidate);
}
}
ret.reserve(accepted.size());
for (auto &candidate: accepted) {
auto cell = candidate.cell;
if (cell.determinant() < .0f)
cell = -cell;
ret.emplace_back(
Coord(cell(0, 0), cell(0, 1), cell(0, 2)),
Coord(cell(1, 0), cell(1, 1), cell(1, 2)),
Coord(cell(2, 0), cell(2, 1), cell(2, 2))
);
}
return ret;
}