mx/Jungfraujoch
2
1
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 68 Wiki Activity
Files
e11c2a2b200d006d2ea82f2d97021135f8286810
Jungfraujoch/image_analysis/MXAnalysisWithoutFPGA.cpp
T
jungfrauandClaude Opus 5 2f586d2267 Stop allocating GPU and pinned memory nothing reads 
Three resource fixes and two latent bugs, none of which changes a computed
number.

The preprocessed image has a host copy that only a CPU engine ever reads. On
the GPU path every engine reads the device buffer instead, and rugnux always
runs the fused adaptive finder, so that host copy is allocated, zeroed and
PAGE-LOCKED for nothing - 72 MB per worker, 3.5 GB over 48 of them, and a
cudaHostRegister each, which the driver serializes. It is now skipped by the
same condition that already decides whether the device copies the image back.
ImagePreprocessorBuffer keeps the pixel count separately so size() still
answers when the mirror was not allocated.

ROIIntegrationGPU asked device 0 for the SM count it sizes its grid from, while
workers are pinned round-robin across the GPUs - so on a multi-GPU node it
could size a grid from a card it never launches on. It asks the current device
now, like every other engine.

~CudaRegisteredVector called a function that throws out of a destructor, and
the move-assignment did the same from a noexcept function. Either would abort
the process rather than report the failure, and teardown - after a device
reset, or while another exception unwinds - is exactly where cudaHostUnregister
fails. Both now use an unchecked unregister, as every other destructor in that
header already does for its own teardown call. The throwing form stays for
rebind()/unregister(), which are called from live code.

Measured on a 16M-pixel rotation dataset: unchanged space group, merged
reflection count and merging statistics.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-15 17:50:19 -04:00

352 lines
20 KiB
C++
Raw Blame History

// SPDX-FileCopyrightText: 2024 Filip Leonarski, Paul Scherrer Institute <filip.leonarski@psi.ch>
// SPDX-License-Identifier: GPL-3.0-only
#include "MXAnalysisWithoutFPGA.h"
#include <algorithm>
#include <spdlog/spdlog.h>
#include "spot_finding/StrongPixelSet.h"
#include "../compression/JFJochDecompress.h"
#include "spot_finding/SpotUtils.h"
#include "bragg_prediction/BraggPredictionFactory.h"
#include "image_preprocessing/ImagePreprocessorCPU.h"
#include "azint/AzIntEngineCPU.h"
#include "roi/ROIIntegrationCPU.h"
#include "spot_finding/ImageSpotFinderCPU.h"
#include "spot_finding/AdaptiveSpotFinderCPU.h"
#include "bragg_integration/BraggIntegrationEngineCPU.h"
#ifdef JFJOCH_USE_CUDA
#include "azint/AzIntEngineGPU.h"
#include "roi/ROIIntegrationGPU.h"
#include "spot_finding/ImageSpotFinderGPU.h"
#include "spot_finding/AdaptiveSpotFinderGPU.h"
#include "image_preprocessing/ImagePreprocessorGPU.h"
#include "image_preprocessing/ImagePreprocessorBufferGPU.h"
#include "bragg_integration/BraggIntegrationEngineGPU.h"
#include "../common/CUDAWrapper.h"
#endif
MXAnalysisWithoutFPGA::MXAnalysisWithoutFPGA(const DiffractionExperiment &in_experiment,
const AzimuthalIntegrationMapping &in_integration,
const PixelMask &in_mask,
IndexAndRefine &in_indexer,
bool in_enable_fused_adaptive_gpu)
: experiment(in_experiment),
integration(in_integration),
enable_fused_adaptive_gpu(in_enable_fused_adaptive_gpu),
npixels(experiment.GetPixelsNum()),
xpixels(experiment.GetXPixelsNum()),
indexer(in_indexer),
prediction(CreateBraggPrediction(experiment.IsRotationIndexing())),
mask(in_mask),
mask_resolution(experiment.GetPixelsNum(), false),
mask_high_res(-1),
mask_low_res(-1) {
#ifdef JFJOCH_USE_CUDA
if (get_gpu_count() == 0) {
#endif
preprocessor_buffer = std::make_unique<ImagePreprocessorBuffer>(experiment.GetPixelsNum());
spotFinder = std::make_unique<ImageSpotFinderCPU>(experiment.GetXPixelsNum(), experiment.GetYPixelsNum());
azint = std::make_unique<AzIntEngineCPU>(integration);
preprocessor = std::make_unique<ImagePreprocessorCPU>(in_experiment, in_mask);
bragg_engine = std::make_unique<BraggIntegrationEngineCPU>(in_experiment);
if (experiment.ROI().size() >= 1)
roi = std::make_unique<ROIIntegrationCPU>(experiment);
#ifdef JFJOCH_USE_CUDA
} else {
stream = std::make_shared<CudaStream>();
// The host copy of the preprocessed image is only read when a CPU engine wants it, which is
// the same condition that drives copy_image_to_host below. Skipping it also skips page-locking
// 4 bytes per pixel per worker.
preprocessor_buffer = std::make_unique<ImagePreprocessorBufferGPU>(
experiment.GetPixelsNum(), /*host_mirror=*/!enable_fused_adaptive_gpu);
// The preprocessed image only has to come back to the host if a CPU engine reads it. Every
// engine built below runs on the GPU, except the CPU adaptive finder that is kept when the fused
// GPU engine is off - so that is the one case that needs the copy. Every caller currently passes
// enable_fused_adaptive_gpu = true, so on the GPU path the copy is off in practice.
preprocessor = std::make_unique<ImagePreprocessorGPU>(in_experiment, in_mask, stream,
/*copy_image_to_host=*/!enable_fused_adaptive_gpu);
spotFinder = std::make_unique<ImageSpotFinderGPU>(experiment.GetXPixelsNum(), experiment.GetYPixelsNum(), stream);
azint = std::make_unique<AzIntEngineGPU>(integration, stream);
bragg_engine = std::make_unique<BraggIntegrationEngineGPU>(in_experiment, stream);
if (experiment.ROI().size() >= 1)
roi = std::make_unique<ROIIntegrationGPU>(experiment, stream);
if (enable_fused_adaptive_gpu) {
// One GPU engine that computes the azimuthal profile and the adaptive spot mask in a single
// image pass. fused_adaptive aliases it so Analyze() can lift the profile out of it.
auto fused = std::make_unique<AdaptiveSpotFinderGPU>(integration, stream);
fused_adaptive = fused.get();
adaptiveSpotFinder = std::move(fused);
}
}
#endif
if (!adaptiveSpotFinder)
adaptiveSpotFinder = std::make_unique<AdaptiveSpotFinderCPU>(integration);
}
void MXAnalysisWithoutFPGA::Analyze(DataMessage &output,
AzimuthalIntegrationProfile &profile,
const SpotFindingSettings &spot_finding_settings) {
if ((output.image.GetWidth() != xpixels)
|| (output.image.GetWidth() * output.image.GetHeight() != npixels))
throw JFJochException(JFJochExceptionCategory::InputParameterInvalid,
"Mismatch in pixel size");
// Decompress on the device where the preprocessor can, so only the compressed chunk crosses PCIe
// and the host does no decompression at all. AnalyzeCompressed says whether it took the image;
// when it declines (a CPU preprocessor, or an algorithm with no device decoder) fall through to
// the host route unchanged. The two produce the same preprocessed image.
const auto compression_start_time = std::chrono::steady_clock::now();
ImageStatistics ret{};
bool decoded_on_device = false;
try {
decoded_on_device = preprocessor->AnalyzeCompressed(*preprocessor_buffer, output.image, ret);
} catch (const JFJochException &e) {
// The device route must never be the reason a frame fails: whatever it could not handle, the
// host decoder gets its turn. If the data really is bad the host throws too and the caller
// sees the same error it saw before any of this existed - but a GPU-side problem costs speed
// rather than the acquisition.
spdlog::warn("Device decoding failed ({}), falling back to host decompression", e.what());
decoded_on_device = false;
}
const auto compression_end_time = std::chrono::steady_clock::now();
if (!decoded_on_device) {
const uint8_t *image_ptr = Decompress(output.image);
const auto decompressed_time = std::chrono::steady_clock::now();
if (output.image.GetCompressionAlgorithm() != CompressionAlgorithm::NO_COMPRESSION)
output.compression_time_s = std::chrono::duration<float>(decompressed_time - compression_start_time).count();
const auto preprocessing_start_time = std::chrono::steady_clock::now();
ret = preprocessor->Analyze(*preprocessor_buffer, image_ptr, output.image.GetMode());
const auto preprocessing_end_time = std::chrono::steady_clock::now();
output.preprocessing_time_s = std::chrono::duration<float>(preprocessing_end_time - preprocessing_start_time).count();
} else {
// Decode and preprocess are one device operation here, but the decompression is still a real,
// separately measurable cost - the decoder brackets it with CUDA events - so it is still
// reported as one. Leaving compression_time_s unset instead would blank the broker's
// "compression" plot trace and fill /entry/profiling/compressionTime with NaN.
const float total_s = std::chrono::duration<float>(compression_end_time - compression_start_time).count();
const float decompress_s = std::min(preprocessor->GetLastDecompressionTime_s(), total_s);
output.compression_time_s = decompress_s;
output.preprocessing_time_s = total_s - decompress_s;
}
// The fused GPU engine (rugnux offline, GPU, adaptive detection) produces the azimuthal profile as
// a byproduct of spot finding, so the separate azint pass is skipped in that case and the profile is
// lifted out of the finder below.
const bool fused = enable_fused_adaptive_gpu && spot_finding_settings.enable
&& spot_finding_settings.adaptive_threshold && fused_adaptive != nullptr;
if (!fused) {
const auto azint_start_time = std::chrono::steady_clock::now();
azint->Run(*preprocessor_buffer, profile);
const auto azint_end_time = std::chrono::steady_clock::now();
output.azint_time_s = std::chrono::duration<float>(azint_end_time - azint_start_time).count();
}
if (roi)
roi->Run(*preprocessor_buffer, output.roi);
if (spot_finding_settings.enable) {
// Update resolution mask
if (mask_high_res != spot_finding_settings.high_resolution_limit
|| mask_low_res != spot_finding_settings.low_resolution_limit)
UpdateMaskResolution(spot_finding_settings);
ImageSpotFinder &finder = spot_finding_settings.adaptive_threshold
? static_cast<ImageSpotFinder &>(*adaptiveSpotFinder)
: *spotFinder;
const auto integrate_fn = [this](const std::vector<Reflection> &predicted, size_t npredicted,
int64_t image_number) {
return bragg_engine->Run(*preprocessor_buffer, predicted, npredicted, image_number);
};
// The radial background correction has to be decided BEFORE this image is integrated, so the
// ice score is taken here rather than with the other per-image quantities at the end of the
// function. It needs the peak-excluded per-ring background, which the adaptive finder has as
// soon as it has detected - so this must be called after detection and before integration.
// Where no such background exists (no adaptive finder), auto leaves the correction off: the
// plain profile carries the Bragg peaks and cannot support an absolute threshold.
const auto decide_radial_background = [this, &spot_finding_settings, &output]() {
if (!bragg_engine->IsBackgroundRadialAuto())
return;
// Same condition as the score at the end of this function: the adaptive finder holds its
// ring background from whenever it last ran, so requiring that it ran for THIS image is
// what keeps a stale curve out.
if (!spot_finding_settings.adaptive_threshold)
return;
const std::vector<float> &ring_bkg = adaptiveSpotFinder->GetRingBackground();
if (ring_bkg.empty())
return;
output.ice_ring_score = AzimuthalIntegrationProfile::IceRingScore(
ring_bkg, integration.GetQBinCount(), integration.Settings(),
spot_finding_settings.ice_ring_width_Q_recipA);
bragg_engine->BackgroundRadial(*output.ice_ring_score
>= experiment.GetScalingSettings().GetIceMinScore());
};
// A missing min-pix (std::nullopt) means "choose it per image". This applies only to the stills
// indexing path (each frame is indexed independently); rotation indexing builds one lattice from
// all frames, so it keeps the fixed min-pix and the single-pass finder.
const bool adaptive_min_pix = !spot_finding_settings.min_pix_per_spot.has_value()
&& spot_finding_settings.indexing
&& !experiment.IsRotationIndexing();
if (adaptive_min_pix) {
// Choose the per-image min-pix adaptively instead of a fixed one. min-pix filters connected
// components AFTER detection, so BOTH the detection (the expensive per-pixel pass) and the
// connected-component search run ONCE and only the filter is repeated; the azimuthal
// profile is the one Detect() computed.
// Index at 3/2/1 (index-only, no integration/accumulation), keep whichever maximises
// n_indexed^2 / n_total (indexed count weighted by indexed fraction) together with its spot
// list, and integrate that one.
const auto detect_start_time = std::chrono::steady_clock::now();
finder.Detect(*preprocessor_buffer, spot_finding_settings);
const auto &components = finder.ExtractComponents(*preprocessor_buffer, spot_finding_settings);
float spot_finding_time_s =
std::chrono::duration<float>(std::chrono::steady_clock::now() - detect_start_time).count();
float indexing_time_s = 0.0f;
SpotFindingSettings s = spot_finding_settings;
std::vector<DiffractionSpot> best_spots;
int best_mp = 0;
double best_score = -1.0;
for (int mp : {3, 2, 1}) {
s.min_pix_per_spot = mp;
const auto extract_start_time = std::chrono::steady_clock::now();
std::vector<DiffractionSpot> spots = ImageSpotFinder::Filter(components, s);
spot_finding_time_s +=
std::chrono::duration<float>(std::chrono::steady_clock::now() - extract_start_time).count();
SpotAnalyze(experiment, s, spots, output);
const bool indexed = indexer.IndexFrameOnly(output, s);
indexing_time_s += output.indexing_time_s.value_or(0.0f);
if (indexed) {
const double n_idx = static_cast<double>(output.spot_count_indexed.value_or(0));
const double n_tot = static_cast<double>(std::max<int64_t>(1, output.spot_count.value_or(1)));
const double score = n_idx * n_idx / n_tot;
if (score > best_score) {
best_score = score;
best_mp = mp;
best_spots = std::move(spots);
}
}
}
if (best_mp != 0) {
// Index and integrate the winning spot list; no spot finding left to do.
s.min_pix_per_spot = best_mp;
SpotAnalyze(experiment, s, best_spots, output);
decide_radial_background();
indexer.ProcessImage(output, s, *prediction, integrate_fn);
indexing_time_s += output.indexing_time_s.value_or(0.0f);
}
// Each indexer call reports only its own time, so the escalation's total is summed here.
output.spot_finding_time_s = spot_finding_time_s;
output.indexing_time_s = indexing_time_s;
} else {
const auto spot_finding_start_time = std::chrono::steady_clock::now();
const std::vector<DiffractionSpot> spots = finder.Run(*preprocessor_buffer, spot_finding_settings);
SpotAnalyze(experiment, spot_finding_settings, spots, output);
output.spot_finding_time_s = std::chrono::duration<float>(std::chrono::steady_clock::now() - spot_finding_start_time).count();
decide_radial_background();
if (spot_finding_settings.indexing)
indexer.ProcessImage(output, spot_finding_settings, *prediction, integrate_fn);
}
#ifdef JFJOCH_USE_CUDA
if (fused) {
// Lift the azimuthal profile the fused engine computed in the same detection pass; its azint
// cost is folded into spot_finding_time_s above.
profile.Clear(integration);
profile += fused_adaptive->GetProfile();
output.azint_time_s = 0.0f;
}
#endif
}
output.max_viable_pixel_value = ret.max_value;
output.min_viable_pixel_value = ret.min_value;
output.error_pixel_count = ret.error_pixel_count;
output.saturated_pixel_count = ret.saturated_pixel_count;
output.az_int_profile = profile.GetResult();
output.az_int_profile_count = profile.GetPixelCount();
output.az_int_profile_std = profile.GetStd();
output.bkg_estimate = profile.GetBkgEstimate(integration.Settings());
// The ice score wants a radial profile with the Bragg peaks taken OUT of it. The azimuthal profile
// is a plain per-ring mean, so a strong low-resolution reflection landing in a ring's bin is
// indistinguishable from ice sitting there - measured, that alone lifts clean crystals to a score
// of 1.5-4.2, right into the range real ice occupies. The adaptive spot finder already computes
// exactly what is wanted: a sigma-clipped per-ring background, in the same bins, from which the
// peaks have been removed (an ice ring is azimuthally smooth, so it survives the clip). It is in
// raw counts rather than corrected ones, which the score does not care about - it is a ratio to the
// background interpolated under the ring, and the corrections are smooth in radius.
const std::vector<float> &ring_bkg = adaptiveSpotFinder->GetRingBackground();
const bool have_ring_bkg = spot_finding_settings.enable && spot_finding_settings.adaptive_threshold
&& !ring_bkg.empty();
output.ice_ring_score = AzimuthalIntegrationProfile::IceRingScore(
have_ring_bkg ? ring_bkg : profile.GetResult1D(), integration.GetQBinCount(),
integration.Settings(), spot_finding_settings.ice_ring_width_Q_recipA);
}
void MXAnalysisWithoutFPGA::RebuildROI() {
if (experiment.ROI().empty()) {
roi.reset();
return;
}
#ifdef JFJOCH_USE_CUDA
if (stream) {
roi = std::make_unique<ROIIntegrationGPU>(experiment, stream);
return;
}
#endif
roi = std::make_unique<ROIIntegrationCPU>(experiment);
}
void MXAnalysisWithoutFPGA::AnalyzeROIOnly(DataMessage &output) {
if ((output.image.GetWidth() != xpixels)
|| (output.image.GetWidth() * output.image.GetHeight() != npixels))
throw JFJochException(JFJochExceptionCategory::InputParameterInvalid,
"Mismatch in pixel size");
const uint8_t *image_ptr = Decompress(output.image);
preprocessor->Analyze(*preprocessor_buffer, image_ptr, output.image.GetMode());
RunROIOnly(output);
}
const uint8_t *MXAnalysisWithoutFPGA::Decompress(const CompressedImage &image) {
// An uncompressed image is read straight out of the message and never touches decompression_buffer,
// so it stays in pageable memory - the buffer is only worth page-locking when it is actually used.
if (image.GetCompressionAlgorithm() != CompressionAlgorithm::NO_COMPRESSION)
preprocessor->PinInputBuffer(decompression_buffer, image.GetUncompressedSize());
return image.GetUncompressedPtr(decompression_buffer);
}
void MXAnalysisWithoutFPGA::RunROIOnly(DataMessage &output) {
output.roi.clear();
if (roi)
roi->Run(*preprocessor_buffer, output.roi);
}
void MXAnalysisWithoutFPGA::UpdateMaskResolution(const SpotFindingSettings &settings) {
mask_low_res = settings.low_resolution_limit;
mask_high_res = settings.high_resolution_limit;
// An unset limit masks nothing at that end. At the high-resolution end 0 does that on its own - no
// pixel has d < 0, and the detector's own edge is where the pixels stop anyway; at the low-resolution
// end every pixel lies above any finite stand-in, so it takes an infinite one.
const float high_res = mask_high_res.value_or(0.0f);
const float low_res = mask_low_res.value_or(INFINITY);
auto const &resolution_map = integration.Resolution();
for (int i = 0; i < mask_resolution.size(); i++)
mask_resolution[i] = (resolution_map[i] > low_res) || (resolution_map[i] < high_res);
// The finders keep their own copy (the GPU ones a bit-packed device copy), so the mask is handed
// over here - when the limits change - rather than with every image.
spotFinder->SetResolutionMask(mask_resolution);
adaptiveSpotFinder->SetResolutionMask(mask_resolution);
}
Reference in New Issue View Git Blame Copy Permalink