Skip to content

NN clusterizer: Cleanup + fix for cluster flag setting #14445

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
davidrohr merged 3 commits into from
Jun 26, 2025
Merged

NN clusterizer: Cleanup + fix for cluster flag setting #14445

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
02ee54e
Fix for cluster flags
ChSonnabend Jun 25, 2025
c9b3ec0
Cleanup + handling of deconvolution
ChSonnabend Jun 25, 2025
e39fb22
Using only propagateMcLabels
ChSonnabend Jun 25, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
71 changes: 15 additions & 56 deletions GPU/GPUTracking/Global/GPUChainTrackingClusterizer.cxx
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -977,20 +977,10 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
GPUTPCNNClusterizer& clustererNNShadow = doGPU ? processorsShadow()->tpcNNClusterer[lane] : clustererNN;
GPUTPCNNClusterizerHost& nnApplication = nnApplications[lane];

// // bool recreateMemoryAllocator = false;
// if (lane == 0) {
// (nnApplications[lane].mModelClass).initEnvironment();
// nnApplications[lane].directOrtAllocator((nnApplications[lane].mModelClass).getEnv(), (nnApplications[lane].mModelClass).getMemoryInfo(), mRec, 0);
// }
// // recreateMemoryAllocator = true;
// (nnApplications[lane].mModelClass).initSession();
// (nnApplications[lane].mModelReg1).initSession();

int withMC = (doGPU && propagateMCLabels);

if (clustererNNShadow.mNnClusterizerUseCfRegression || (int)(nn_settings.nnClusterizerApplyCfDeconvolution)) {
// int withMC = (doGPU && propagateMCLabels);

if (nn_settings.nnClusterizerApplyCfDeconvolution) {
runKernel<GPUTPCCFDeconvolution>({GetGrid(clusterer.mPmemory->counters.nPositions, lane), {iSector}}, true);
DoDebugAndDump(RecoStep::TPCClusterFinding, GPUChainTrackingDebugFlags::TPCClustererChargeMap, clusterer, &GPUTPCClusterFinder::DumpChargeMap, *mDebugFile, "Split Charges");
} else if (clustererNNShadow.mNnClusterizerSetDeconvolutionFlags) {
runKernel<GPUTPCCFDeconvolution>({GetGrid(clusterer.mPmemory->counters.nPositions, lane), {iSector}}, false);
}
Expand All @@ -1001,15 +991,12 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
size_t iSize = CAMath::Min((uint)clustererNNShadow.mNnClusterizerBatchedMode, (uint)(clusterer.mPmemory->counters.nClusters - batchStart));

// auto start0 = std::chrono::high_resolution_clock::now();
runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::fillInputNNSingleElement>({GetGrid(iSize * clustererNNShadow.mNnClusterizerElementSize, lane), krnlRunRangeNone}, iSector, clustererNNShadow.mNnInferenceInputDType, withMC, batchStart); // Filling the data
runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::fillInputNNSingleElement>({GetGrid(iSize * clustererNNShadow.mNnClusterizerElementSize, lane), krnlRunRangeNone}, iSector, clustererNNShadow.mNnInferenceInputDType, propagateMCLabels, batchStart); // Filling the data

if (clustererNNShadow.mNnClusterizerSetDeconvolutionFlags) {
runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::publishDeconvolutionFlags>({GetGrid(iSize, lane), krnlRunRangeNone}, iSector, clustererNNShadow.mNnInferenceInputDType, withMC, batchStart); // Filling the regression data
runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::publishDeconvolutionFlags>({GetGrid(iSize, lane), krnlRunRangeNone}, iSector, clustererNNShadow.mNnInferenceInputDType, propagateMCLabels, batchStart); // Filling the regression data
}

// auto stop0 = std::chrono::high_resolution_clock::now();
// auto start1 = std::chrono::high_resolution_clock::now();

// NN evaluations
if (clustererNNShadow.mNnInferenceInputDType == 0) {
if (clustererNNShadow.mNnInferenceOutputDType == 0) {
Expand Down Expand Up @@ -1055,55 +1042,27 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
}
}

// auto stopNNs = std::chrono::high_resolution_clock::now();

// Publishing kernels
if (nnApplication.mModelClass.getNumOutputNodes()[0][1] == 1) {
runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::determineClass1Labels>({GetGrid(iSize, lane), krnlRunRangeNone}, iSector, clustererNNShadow.mNnInferenceOutputDType, withMC, batchStart); // Assigning class labels
runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::determineClass1Labels>({GetGrid(iSize, lane), krnlRunRangeNone}, iSector, clustererNNShadow.mNnInferenceOutputDType, propagateMCLabels, batchStart); // Assigning class labels
} else {
runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::determineClass2Labels>({GetGrid(iSize, lane), krnlRunRangeNone}, iSector, clustererNNShadow.mNnInferenceOutputDType, withMC, batchStart); // Assigning class labels
runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::determineClass2Labels>({GetGrid(iSize, lane), krnlRunRangeNone}, iSector, clustererNNShadow.mNnInferenceOutputDType, propagateMCLabels, batchStart); // Assigning class labels
}
if (!clustererNNShadow.mNnClusterizerUseCfRegression) {
runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::publishClass1Regression>({GetGrid(iSize, lane), krnlRunRangeNone}, iSector, clustererNNShadow.mNnInferenceOutputDType, withMC, batchStart); // Publishing class 1 regression results
runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::publishClass1Regression>({GetGrid(iSize, lane), krnlRunRangeNone}, iSector, clustererNNShadow.mNnInferenceOutputDType, propagateMCLabels, batchStart); // Publishing class 1 regression results
if (nnApplication.mModelClass.getNumOutputNodes()[0][1] > 1 && nnApplication.mModelReg2.isInitialized()) {
runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::publishClass2Regression>({GetGrid(iSize, lane), krnlRunRangeNone}, iSector, clustererNNShadow.mNnInferenceOutputDType, withMC, batchStart); // Publishing class 2 regression results
runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::publishClass2Regression>({GetGrid(iSize, lane), krnlRunRangeNone}, iSector, clustererNNShadow.mNnInferenceOutputDType, propagateMCLabels, batchStart); // Publishing class 2 regression results
}
}

// for(int i = 0; i < iSize; ++i) {
// if(clustererNNShadow.mOutputDataClass[i + batchStart] > 1) {
// LOG(info) << "WARNING ORT: Output of " << i + batchStart << " / " << clusterer.mPmemory->counters.nClusters << " is " << clustererNNShadow.mModelProbabilities_16[i].ToFloat() << " and " << clustererNNShadow.mOutputDataClass[i + batchStart] << " thresh " << clustererNNShadow.mNnClassThreshold << " instead of 0 or 1. Please check the model and the input data.";
// // std::string input = "[";
// // for(int j = 0; j < clustererNNShadow.mNnClusterizerElementSize; j++){
// // input += std::to_string(clustererNNShadow.mInputData_16[i * clustererNNShadow.mNnClusterizerElementSize + j].ToFloat()) + ", ";
// // }
// // input += "]";
// // LOG(info) << "Input is: " << input;
// }
// }

// auto stop1 = std::chrono::high_resolution_clock::now();

// time_networks += std::chrono::duration_cast<std::chrono::nanoseconds>(stopNNs - start1).count() / 1e9;
// time_clusterizer += std::chrono::duration_cast<std::chrono::nanoseconds>(stop1 - start1).count() / 1e9;
// time_fill += std::chrono::duration_cast<std::chrono::nanoseconds>(stop0 - start0).count() / 1e9;
}

if (clustererNNShadow.mNnClusterizerUseCfRegression) {
// auto start1 = std::chrono::high_resolution_clock::now();
runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::runCfClusterizer>({GetGrid(clusterer.mPmemory->counters.nClusters, lane), krnlRunRangeNone}, iSector, clustererNNShadow.mNnInferenceInputDType, withMC, 0); // Running the CF regression kernel - no batching needed: batchStart = 0
// auto stop1 = std::chrono::high_resolution_clock::now();
// time_clusterizer += std::chrono::duration_cast<std::chrono::nanoseconds>(stop1 - start1).count() / 1e9;
if(!nn_settings.nnClusterizerApplyCfDeconvolution) {
runKernel<GPUTPCCFDeconvolution>({GetGrid(clusterer.mPmemory->counters.nPositions, lane), {iSector}}, true);
}
DoDebugAndDump(RecoStep::TPCClusterFinding, GPUChainTrackingDebugFlags::TPCClustererChargeMap, clusterer, &GPUTPCClusterFinder::DumpChargeMap, *mDebugFile, "Split Charges");
runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::runCfClusterizer>({GetGrid(clusterer.mPmemory->counters.nClusters, lane), krnlRunRangeNone}, iSector, clustererNNShadow.mNnInferenceInputDType, propagateMCLabels, 0); // Running the CF regression kernel - no batching needed: batchStart = 0
}
// if (clustererNNShadow.mNnClusterizerVerbosity < 3) {
// int acceptedClusters = 0;
// for (size_t i = 0; i < clusterer.mPmemory->counters.nClusters; ++i) {
// if(clustererNNShadow.mOutputDataClass[i] > 1 || clustererNNShadow.mOutputDataClass[i] < 0) {
// LOG(info) << "WARNING ORT 2: " << clustererNNShadow.mOutputDataClass[i] << " for index " << i << " / " << clusterer.mPmemory->counters.nClusters;
// }
// acceptedClusters += clustererNNShadow.mOutputDataClass[i];
// }
// LOG(info) << "[NN CF] Apply NN (fragment " << fragment.index << ", lane: " << lane << ", sector: " << iSector << "): filling data " << time_fill << "s ; networks: " << time_networks << "s ; clusterizer: " << time_clusterizer << "s ; " << clusterer.mPmemory->counters.nClusters << " clusters, " << acceptedClusters << " accepted. --> " << (int32_t)clusterer.mPmemory->counters.nClusters / (time_fill + time_clusterizer) << " clusters/s";
// }
#else
GPUFatal("Project not compiled with neural network clusterization. Aborting.");
#endif
Expand Down
4 changes: 4 additions & 0 deletions GPU/GPUTracking/TPCClusterFinder/GPUTPCNNClusterizerKernels.cxx
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -127,6 +127,8 @@ GPUdii() void GPUTPCNNClusterizerKernels::Thread<GPUTPCNNClusterizerKernels::fil
if (clustererNN.mNnClusterizerAddIndexData && (int32_t)transient_index == (clustererNN.mNnClusterizerElementSize - 1)) {
uint32_t top_idx = (base_idx + 1) * clustererNN.mNnClusterizerElementSize;
if (!clustererNN.mNnClusterizerSetDeconvolutionFlags) { // Only if deconvolution flags are not set
clustererNN.mClusterFlags[2 * base_idx] = 0;
clustererNN.mClusterFlags[2 * base_idx + 1] = 0;
for (uint16_t i = 0; i < 8; i++) { // This solution needs testing. It is not the same as the deconvolution flags
Delta2 d = cfconsts::InnerNeighbors[i];
CfChargePos tmp_pos = peak.delta(d);
Expand Down Expand Up @@ -470,6 +472,8 @@ GPUdii() void GPUTPCNNClusterizerKernels::Thread<GPUTPCNNClusterizerKernels::pub
CfArray2D<PackedCharge> chargeMap(reinterpret_cast<PackedCharge*>(clusterer.mPchargeMap));
CfChargePos peak = clusterer.mPfilteredPeakPositions[idx + batchStart];

clustererNN.mClusterFlags[2 * idx] = 0;
clustererNN.mClusterFlags[2 * idx + 1] = 0;
for (int i = 0; i < 8; i++) {
Delta2 d = cfconsts::InnerNeighbors[i];
CfChargePos tmp_pos = peak.delta(d);
Expand Down