GPUTPCNNClusterizerKernels.cxx

Go to the documentation of this file.

// Copyright 2019-2020 CERN and copyright holders of ALICE O2.
// See https://alice-o2.web.cern.ch/copyright for details of the copyright holders.
// All rights not expressly granted are reserved.
//
// This software is distributed under the terms of the GNU General Public
// License v3 (GPL Version 3), copied verbatim in the file "COPYING".
//
// In applying this license CERN does not waive the privileges and immunities
// granted to it by virtue of its status as an Intergovernmental Organization
// or submit itself to any jurisdiction.
 
 
#include "GPUTPCNNClusterizerKernels.h"
#include "GPUTPCCFClusterizer.h"
#include "GPUTPCGeometry.h"
 
using namespace o2::gpu;
using namespace o2::gpu::tpccf;
 
#include "CfConsts.h"
#include "CfUtils.h"
#include "ClusterAccumulator.h"
#include "ML/3rdparty/GPUORTFloat16.h"
 
#if !defined(GPUCA_GPUCODE)
#include "GPUHostDataTypes.h"
#include "MCLabelAccumulator.h"
#endif
 
#ifdef GPUCA_GPUCODE
#include "GPUTPCCFClusterizer.inc"
#endif
 
// Defining individual thread functions for data filling, determining the class label and running the CF clusterizer
template <>
GPUdii() void GPUTPCNNClusterizerKernels::Thread<GPUTPCNNClusterizerKernels::runCfClusterizer>(int32_t nBlocks, int32_t nThreads, int32_t iBlock, int32_t iThread, GPUSharedMemory& smem, processorType& processors, uint8_t sector, int8_t dtype, int8_t withMC, uint32_t batchStart)
{
  uint32_t glo_idx = get_global_id(0);
  auto& clusterer = processors.tpcClusterer[sector];
  auto& clustererNN = processors.tpcNNClusterer[sector];
  if (clustererNN.mOutputDataClass[glo_idx] == 0) { // default clusterizer should not be called in batched mode due to mess-up with thread indices
    return;
  }
  CfArray2D<PackedCharge> chargeMap(reinterpret_cast<PackedCharge*>(clusterer.mPchargeMap));
  CPU_ONLY(MCLabelAccumulator labelAcc(clusterer));
  tpc::ClusterNative* clusterOut = (withMC) ? nullptr : clusterer.mPclusterByRow;
  o2::gpu::GPUTPCCFClusterizer::GPUSharedMemory smem_new;
  GPUTPCCFClusterizer::computeClustersImpl(get_num_groups(0), get_local_size(0), get_group_id(0), get_local_id(0), clusterer, clusterer.mPmemory->fragment, smem_new, chargeMap, clusterer.mPfilteredPeakPositions, clusterer.Param().rec, CPU_PTR(&labelAcc), clusterer.mPmemory->counters.nClusters, clusterer.mNMaxClusterPerRow, clusterer.mPclusterInRow, clusterOut, clusterer.mPclusterPosInRow);
}
 
template <>
GPUdii() void GPUTPCNNClusterizerKernels::Thread<GPUTPCNNClusterizerKernels::fillInputNNCPU>(int32_t nBlocks, int32_t nThreads, int32_t iBlock, int32_t iThread, GPUSharedMemory& smem, processorType& processors, uint8_t sector, int8_t dtype, int8_t withMC, uint32_t batchStart)
{
  auto& clusterer = processors.tpcClusterer[sector];
  auto& clustererNN = processors.tpcNNClusterer[sector];
 
  uint32_t glo_idx = get_global_id(0);
  if (glo_idx + batchStart >= clusterer.mPmemory->counters.nClusters) {
    return;
  }
 
  uint32_t write_idx = glo_idx * clustererNN.mNnClusterizerElementSize;
 
  CfArray2D<PackedCharge> chargeMap(reinterpret_cast<PackedCharge*>(clusterer.mPchargeMap));
  CfArray2D<uint8_t> isPeakMap(clusterer.mPpeakMap);
  CfChargePos peak = clusterer.mPfilteredPeakPositions[CAMath::Min(glo_idx + batchStart, (uint32_t)(clusterer.mPmemory->counters.nClusters - 1))];
  int32_t row = static_cast<int>(peak.row());
  int32_t pad = static_cast<int>(peak.pad());
  int32_t time = static_cast<int>(peak.time());
  float central_charge = static_cast<float>(chargeMap[peak].unpack());
  int32_t row_offset = GPUTPCNNClusterizerKernels::rowOffset(row, clustererNN.mNnClusterizerSizeInputRow);
 
  for (int32_t r = -clustererNN.mNnClusterizerSizeInputRow; r <= clustererNN.mNnClusterizerSizeInputRow; ++r) {
    int32_t target_row = row + r;
    bool is_row_boundary = (target_row < 0) || (target_row >= o2::tpc::constants::MAXGLOBALPADROW);
    int32_t pad_offset = is_row_boundary ? 0 : GPUTPCNNClusterizerKernels::padOffset(row, target_row);
 
    for (int32_t p = -clustererNN.mNnClusterizerSizeInputPad + pad_offset; p <= clustererNN.mNnClusterizerSizeInputPad + pad_offset; ++p) {
      int32_t target_pad = pad + p;
      bool is_boundary = is_row_boundary || GPUTPCNNClusterizerKernels::isBoundary(target_row + row_offset, target_pad, clustererNN.mNnClusterizerSizeInputRow);
 
      for (int32_t t = -clustererNN.mNnClusterizerSizeInputTime; t <= clustererNN.mNnClusterizerSizeInputTime; ++t) {
        int32_t target_time = time + t;
 
        if (is_boundary || target_time < 0 || target_time >= TPC_MAX_FRAGMENT_LEN_GPU) {
          // Fill boundary value
          float boundary_value = static_cast<float>(clustererNN.mNnClusterizerBoundaryFillValue);
          if (dtype == 0) {
            clustererNN.mInputData_16[write_idx] = (OrtDataType::Float16_t)boundary_value;
          } else {
            clustererNN.mInputData_32[write_idx] = boundary_value;
          }
        } else {
          CfChargePos tmp_pos(target_row, target_pad, target_time);
          float normalized_charge = static_cast<float>(chargeMap[tmp_pos].unpack()) / central_charge;
 
          if (!clustererNN.mNnClusterizerSetDeconvolutionFlags && r == 0 && CAMath::Abs(p) < 3 && CAMath::Abs(t) < 3 && p != 0 && t != 0) {
            clustererNN.mClusterFlags[2 * glo_idx] += CfUtils::isPeak(isPeakMap[tmp_pos]);
            clustererNN.mClusterFlags[2 * glo_idx + 1] = clustererNN.mClusterFlags[2 * glo_idx];
          }
 
          if (dtype == 0) {
            clustererNN.mInputData_16[write_idx] = (OrtDataType::Float16_t)normalized_charge;
          } else {
            clustererNN.mInputData_32[write_idx] = normalized_charge;
          }
        }
        // if((CAMath::Abs(static_cast<float>(clustererNN.mInputData_16_Test[write_idx]) - static_cast<float>(clustererNN.mInputData_16[write_idx])) > 1e-4) && ((glo_idx + batchStart) < clusterer.mPmemory->counters.nClusters)) {
        //   printf("Warning: Input data mismatch at index %d, %d - row, pad, time: %d, %d, %d : %f -> %f\n", glo_idx, glo_idx + batchStart, r, p, t,
        //          static_cast<float>(clustererNN.mInputData_16_Test[write_idx]), static_cast<float>(clustererNN.mInputData_16[write_idx]));
        // }
        write_idx++;
      }
    }
  }
 
  if (clustererNN.mNnClusterizerAddIndexData) {
    if (dtype == 0) {
      clustererNN.mInputData_16[write_idx] = (OrtDataType::Float16_t)(static_cast<float>(sector) / o2::tpc::constants::MAXSECTOR);
      clustererNN.mInputData_16[write_idx + 1] = (OrtDataType::Float16_t)(static_cast<float>(row) / o2::tpc::constants::MAXGLOBALPADROW);
      clustererNN.mInputData_16[write_idx + 2] = (OrtDataType::Float16_t)(static_cast<float>(pad) / GPUTPCGeometry::NPads(row));
    } else {
      clustererNN.mInputData_32[write_idx] = static_cast<float>(sector) / o2::tpc::constants::MAXSECTOR;
      clustererNN.mInputData_32[write_idx + 1] = static_cast<float>(row) / o2::tpc::constants::MAXGLOBALPADROW;
      clustererNN.mInputData_32[write_idx + 2] = static_cast<float>(pad) / GPUTPCGeometry::NPads(row);
    }
  }
 
  if (!clustererNN.mNnClusterizerSetDeconvolutionFlags) {
    clustererNN.mClusterFlags[2 * glo_idx] = 0;
    clustererNN.mClusterFlags[2 * glo_idx + 1] = 0;
 
    for (uint16_t i = 0; i < 8; ++i) {
      Delta2 d = cfconsts::InnerNeighbors[i];
      CfChargePos tmp_pos = peak.delta(d);
      clustererNN.mClusterFlags[2 * glo_idx] += CfUtils::isPeak(isPeakMap[tmp_pos]);
    }
    clustererNN.mClusterFlags[2 * glo_idx + 1] = clustererNN.mClusterFlags[2 * glo_idx];
  }
}
 
template <>
GPUdii() void GPUTPCNNClusterizerKernels::Thread<GPUTPCNNClusterizerKernels::fillInputNNGPU>(int32_t nBlocks, int32_t nThreads, int32_t iBlock, int32_t iThread, GPUSharedMemory& smem, processorType& processors, uint8_t sector, int8_t dtype, int8_t withMC, uint32_t batchStart)
{
  uint32_t glo_idx = get_global_id(0);
 
  auto& clusterer = processors.tpcClusterer[sector];
  auto& clustererNN = processors.tpcNNClusterer[sector];
 
  // Optimized division using bit operations
  uint32_t base_idx = glo_idx / clustererNN.mNnClusterizerRowTimeSizeFull;
  uint32_t transient_index = glo_idx - (base_idx * clustererNN.mNnClusterizerRowTimeSizeFull);
 
  // Early exit for out-of-bounds threads
  if (base_idx + batchStart >= clusterer.mPmemory->counters.nClusters) {
    return;
  }
  CfArray2D<PackedCharge> chargeMap(reinterpret_cast<PackedCharge*>(clusterer.mPchargeMap));
  CfArray2D<uint8_t> isPeakMap(clusterer.mPpeakMap);
 
  // Use dedicated neural network shared memory arrays for warp-level caching
  // First thread in each warp loads shared data
  CfChargePos peak = clusterer.mPfilteredPeakPositions[CAMath::Min(base_idx + batchStart, (uint32_t)(clusterer.mPmemory->counters.nClusters - 1))];
  float central_charge = static_cast<float>(chargeMap[peak].unpack());
  int32_t row = static_cast<int>(peak.row());
  int32_t pad = static_cast<int>(peak.pad());
  int32_t time = static_cast<int>(peak.time());
 
  // Handle index data with fewer branches
  if (clustererNN.mNnClusterizerAddIndexData && transient_index >= clustererNN.mNnClusterizerRowTimeSize) {
    int32_t data_idx = transient_index - clustererNN.mNnClusterizerRowTimeSize;
    uint32_t write_idx = base_idx * clustererNN.mNnClusterizerElementSize + clustererNN.mNnClusterizerChargeArraySize + data_idx;
 
    float index_values[3] = {
      static_cast<float>(sector) / o2::tpc::constants::MAXSECTOR,
      static_cast<float>(row) / o2::tpc::constants::MAXGLOBALPADROW,
      static_cast<float>(pad) / GPUTPCGeometry::NPads(row)};
 
    if (dtype == 0) {
      clustererNN.mInputData_16[write_idx] = (OrtDataType::Float16_t)index_values[data_idx];
    } else {
      clustererNN.mInputData_32[write_idx] = index_values[data_idx];
    }
 
    // Handle deconvolution flags only once per cluster (last thread in element)
    if (data_idx == 2 && !clustererNN.mNnClusterizerSetDeconvolutionFlags) {
      uint8_t cluster_flags = 0;
      for (uint16_t i = 0; i < 8; i++) {
        Delta2 d = cfconsts::InnerNeighbors[i];
        CfChargePos tmp_pos = peak.delta(d);
        cluster_flags += CfUtils::isPeak(isPeakMap[tmp_pos]);
      }
      clustererNN.mClusterFlags[2 * base_idx] = cluster_flags;
      clustererNN.mClusterFlags[2 * base_idx + 1] = cluster_flags;
    }
    return;
  }
 
  // Main data processing - optimize index calculations
  if (transient_index < clustererNN.mNnClusterizerRowTimeSize) {
    // Optimize 3D index calculation
    int32_t row_idx = transient_index / clustererNN.mNnClusterizerFullTimeSize;
    int32_t r_local = row_idx - clustererNN.mNnClusterizerSizeInputRow;
    int32_t time_idx = transient_index - row_idx * clustererNN.mNnClusterizerFullTimeSize;
    int32_t t_local = time_idx - clustererNN.mNnClusterizerSizeInputTime;
    int32_t write_idx = base_idx * clustererNN.mNnClusterizerElementSize + row_idx * clustererNN.mNnClusterizerPadTimeSize + time_idx;
 
    // Early boundary check for row
    int32_t target_row = row + r_local;
    int8_t is_row_boundary = (target_row < 0) || (target_row > (o2::tpc::constants::MAXGLOBALPADROW - 1));
 
    // Calculate offsets
    int32_t row_offset = GPUTPCNNClusterizerKernels::rowOffset(row, clustererNN.mNnClusterizerSizeInputRow);
    int32_t pad_offset = GPUTPCNNClusterizerKernels::padOffset(row, target_row);
    for (int32_t p_local = -clustererNN.mNnClusterizerSizeInputPad + pad_offset; p_local <= clustererNN.mNnClusterizerSizeInputPad + pad_offset; p_local++) {
      if (is_row_boundary) {
        // Use boundary fill value
        float boundary_val = static_cast<float>(clustererNN.mNnClusterizerBoundaryFillValue);
        if (dtype == 0) {
          clustererNN.mInputData_16[write_idx] = (OrtDataType::Float16_t)boundary_val;
        } else {
          clustererNN.mInputData_32[write_idx] = boundary_val;
        }
        write_idx += clustererNN.mNnClusterizerFullTimeSize; // Move to next pad position
        continue;
      }
 
      // Calculate target pad and time
      int32_t target_pad = pad + p_local;
      int32_t target_time = time + t_local;
 
      // Optimized boundary check
      int8_t is_boundary = GPUTPCNNClusterizerKernels::isBoundary(target_row + row_offset, target_pad, clustererNN.mNnClusterizerSizeInputRow) || (target_time < 0) || (target_time >= TPC_MAX_FRAGMENT_LEN_GPU);
 
      float output_value;
      if (is_boundary) {
        output_value = static_cast<float>(clustererNN.mNnClusterizerBoundaryFillValue);
      } else {
        // Coalesced memory access - create position and read charge
        CfChargePos tmp_pos(target_row, target_pad, target_time);
        output_value = static_cast<float>(chargeMap[tmp_pos].unpack()) / central_charge; // Normalize by central charge
      }
 
      // Write output with reduced branching
      if (dtype == 0) {
        clustererNN.mInputData_16[write_idx] = (OrtDataType::Float16_t)output_value;
      } else {
        clustererNN.mInputData_32[write_idx] = output_value;
      }
      write_idx += clustererNN.mNnClusterizerFullTimeSize; // Move to next pad position
    }
  }
}
 
template <>
GPUdii() void GPUTPCNNClusterizerKernels::Thread<GPUTPCNNClusterizerKernels::determineClass1Labels>(int32_t nBlocks, int32_t nThreads, int32_t iBlock, int32_t iThread, GPUSharedMemory& smem, processorType& processors, uint8_t sector, int8_t dtype, int8_t withMC, uint32_t batchStart)
{
  uint32_t glo_idx = get_global_id(0);
  if (dtype == 0) {
    processors.tpcNNClusterer[sector].mOutputDataClass[glo_idx + batchStart] = (int)((processors.tpcNNClusterer[sector].mModelProbabilities_16[glo_idx]).ToFloat() > processors.tpcNNClusterer[sector].mNnClassThreshold);
  } else if (dtype == 1) {
    processors.tpcNNClusterer[sector].mOutputDataClass[glo_idx + batchStart] = (int)(processors.tpcNNClusterer[sector].mModelProbabilities_32[glo_idx] > processors.tpcNNClusterer[sector].mNnClassThreshold);
  }
}
 
template <>
GPUdii() void GPUTPCNNClusterizerKernels::Thread<GPUTPCNNClusterizerKernels::determineClass2Labels>(int32_t nBlocks, int32_t nThreads, int32_t iBlock, int32_t iThread, GPUSharedMemory& smem, processorType& processors, uint8_t sector, int8_t dtype, int8_t withMC, uint32_t batchStart)
{
  auto& clustererNN = processors.tpcNNClusterer[sector];
  uint32_t glo_idx = get_global_id(0);
  uint32_t elem_iterator = glo_idx * clustererNN.mNnClusterizerModelClassNumOutputNodes;
  float current_max_prob = 0.f; // If the neural network doesn't contain the softmax as a last layer, the outputs can range in [-infty, infty]
  uint32_t class_label = 0;
  for (uint32_t pIdx = elem_iterator; pIdx < elem_iterator + clustererNN.mNnClusterizerModelClassNumOutputNodes; pIdx++) {
    if (pIdx == elem_iterator) {
      if (dtype == 0) {
        current_max_prob = static_cast<float>(clustererNN.mModelProbabilities_16[pIdx]);
      } else if (dtype == 1) {
        current_max_prob = clustererNN.mModelProbabilities_32[pIdx];
      }
    } else {
      if (dtype == 0) {
        current_max_prob = CAMath::Max(current_max_prob, clustererNN.mModelProbabilities_16[pIdx].ToFloat());
      } else if (dtype == 1) {
        current_max_prob = CAMath::Max(current_max_prob, clustererNN.mModelProbabilities_32[pIdx]);
      }
    }
  }
  // uint32_t class_label = std::distance(elem_iterator, std::max_element(elem_iterator, elem_iterator + clustererNN.mNnClusterizerModelClassNumOutputNodes)); // Multiple outputs of the class network are the probabilities for each class. The highest one "wins"
  clustererNN.mOutputDataClass[glo_idx + batchStart] = class_label;
  if (class_label > 1) {
    clustererNN.mClusterFlags[2 * glo_idx] = 1;
    clustererNN.mClusterFlags[2 * glo_idx + 1] = 1;
  }
}
 
template <>
GPUdii() void GPUTPCNNClusterizerKernels::Thread<GPUTPCNNClusterizerKernels::publishClass1Regression>(int32_t nBlocks, int32_t nThreads, int32_t iBlock, int32_t iThread, GPUSharedMemory& smem, processorType& processors, uint8_t sector, int8_t dtype, int8_t withMC, uint32_t batchStart)
{
  uint32_t glo_idx = get_global_id(0);
  auto& clusterer = processors.tpcClusterer[sector];
  auto& clustererNN = processors.tpcNNClusterer[sector];
 
  uint32_t maxClusterNum = clusterer.mPmemory->counters.nClusters;
  uint32_t full_glo_idx = glo_idx + batchStart;
  int32_t model_output_index = glo_idx * clustererNN.mNnClusterizerModelReg1NumOutputNodes;
 
  CfArray2D<PackedCharge> chargeMap(reinterpret_cast<PackedCharge*>(clusterer.mPchargeMap));
  CfChargePos peak = clusterer.mPfilteredPeakPositions[CAMath::Min(full_glo_idx, maxClusterNum - 1)];
  float central_charge = static_cast<float>(chargeMap[peak].unpack());
 
  CPU_ONLY(MCLabelAccumulator labelAccElem(clusterer));
  MCLabelAccumulator* labelAcc = CPU_PTR(&labelAccElem);
 
  if (full_glo_idx >= maxClusterNum) {
    if (withMC) {
      ClusterAccumulator dummy_pc;
      CPU_ONLY(labelAcc->collect(peak, central_charge));
      GPUTPCCFClusterizer::buildCluster(
        clusterer.Param().rec,
        chargeMap,
        peak,
        smem.posBcast,
        smem.buf,
        smem.innerAboveThreshold,
        &dummy_pc,
        labelAcc);
    }
    return;
  }
 
  tpc::ClusterNative* clusterOut = clusterer.mPclusterByRow;
 
  // LOG(info) << glo_idx << " -- " << model_output_index << " / " << clustererNN.outputDataReg1.size() << " / " << clustererNN.mNnClusterizerModelReg1NumOutputNodes << " -- " << clusterer.peakPositions.size() << " -- " << clusterer.centralCharges.size();
 
  if (clustererNN.mOutputDataClass[full_glo_idx] == 1 || (clustererNN.mNnClusterizerUseClassification <= 0)) {
 
    ClusterAccumulator pc;
 
    // Publishing logic is taken from default clusterizer
    if (withMC) {
      ClusterAccumulator dummy_pc;
      CPU_ONLY(labelAcc->collect(peak, central_charge));
      GPUTPCCFClusterizer::buildCluster(
        clusterer.Param().rec,
        chargeMap,
        peak,
        smem.posBcast,
        smem.buf,
        smem.innerAboveThreshold,
        &dummy_pc,
        labelAcc);
    }
    if ((clusterer.mPmemory->fragment).isOverlap(peak.time())) {
      if (clusterer.mPclusterPosInRow) {
        clusterer.mPclusterPosInRow[full_glo_idx] = clusterer.mNMaxClusterPerRow;
      }
      return;
    }
 
    if (dtype == 0) {
      pc.setFull(central_charge * clustererNN.mOutputDataReg1_16[model_output_index + 4].ToFloat(),
                 static_cast<float>(peak.pad()) + clustererNN.mOutputDataReg1_16[model_output_index].ToFloat(),
                 clustererNN.mOutputDataReg1_16[model_output_index + 2].ToFloat(),
                 (clusterer.mPmemory->fragment).start + static_cast<float>(peak.time()) + clustererNN.mOutputDataReg1_16[model_output_index + 1].ToFloat(),
                 clustererNN.mOutputDataReg1_16[model_output_index + 3].ToFloat(),
                 clustererNN.mClusterFlags[2 * glo_idx],
                 clustererNN.mClusterFlags[2 * glo_idx + 1]);
    } else if (dtype == 1) {
      pc.setFull(central_charge * clustererNN.mOutputDataReg1_32[model_output_index + 4],
                 static_cast<float>(peak.pad()) + clustererNN.mOutputDataReg1_32[model_output_index],
                 clustererNN.mOutputDataReg1_32[model_output_index + 2],
                 (clusterer.mPmemory->fragment).start + static_cast<float>(peak.time()) + clustererNN.mOutputDataReg1_32[model_output_index + 1],
                 clustererNN.mOutputDataReg1_32[model_output_index + 3],
                 clustererNN.mClusterFlags[2 * glo_idx],
                 clustererNN.mClusterFlags[2 * glo_idx + 1]);
    }
 
    tpc::ClusterNative myCluster;
    bool rejectCluster = !pc.toNative(peak, central_charge, myCluster, clusterer.Param(), chargeMap);
    if (rejectCluster) {
      if (clusterer.mPclusterPosInRow) {
        clusterer.mPclusterPosInRow[full_glo_idx] = clusterer.mNMaxClusterPerRow;
      }
      return;
    }
 
    uint32_t rowIndex = 0;
    if (clusterOut != nullptr) {
      rowIndex = GPUTPCCFClusterizer::sortIntoBuckets(
        clusterer,
        myCluster,
        peak.row(),
        clusterer.mNMaxClusterPerRow,
        clusterer.mPclusterInRow,
        clusterOut);
      if (clusterer.mPclusterPosInRow != nullptr) {
        clusterer.mPclusterPosInRow[full_glo_idx] = rowIndex;
      }
    } else if (clusterer.mPclusterPosInRow) {
      rowIndex = clusterer.mPclusterPosInRow[full_glo_idx];
    }
    CPU_ONLY(labelAcc->commit(peak.row(), rowIndex, clusterer.mNMaxClusterPerRow));
  } else {
    if (clusterer.mPclusterPosInRow) {
      clusterer.mPclusterPosInRow[full_glo_idx] = clusterer.mNMaxClusterPerRow;
    }
    return;
  }
}
 
template <>
GPUdii() void GPUTPCNNClusterizerKernels::Thread<GPUTPCNNClusterizerKernels::publishClass2Regression>(int32_t nBlocks, int32_t nThreads, int32_t iBlock, int32_t iThread, GPUSharedMemory& smem, processorType& processors, uint8_t sector, int8_t dtype, int8_t withMC, uint32_t batchStart)
{
  uint32_t glo_idx = get_global_id(0);
  auto& clusterer = processors.tpcClusterer[sector];
  auto& clustererNN = processors.tpcNNClusterer[sector];
 
  uint32_t maxClusterNum = clusterer.mPmemory->counters.nClusters;
  CfArray2D<PackedCharge> chargeMap(reinterpret_cast<PackedCharge*>(clusterer.mPchargeMap));
  CfChargePos peak = clusterer.mPfilteredPeakPositions[CAMath::Min(glo_idx + batchStart, (uint32_t)(clusterer.mPmemory->counters.nClusters - 1))];
  float central_charge = static_cast<float>(chargeMap[peak].unpack());
 
  CPU_ONLY(MCLabelAccumulator labelAccElem(clusterer));
  MCLabelAccumulator* labelAcc = CPU_PTR(&labelAccElem);
  tpc::ClusterNative* clusterOut = (withMC) ? nullptr : clusterer.mPclusterByRow;
  uint32_t full_glo_idx = glo_idx + batchStart;
 
  if (full_glo_idx >= maxClusterNum) {
    if (withMC) {
      ClusterAccumulator dummy_pc;
      CPU_ONLY(labelAcc->collect(peak, central_charge));
      GPUTPCCFClusterizer::buildCluster(
        clusterer.Param().rec,
        chargeMap,
        peak,
        smem.posBcast,
        smem.buf,
        smem.innerAboveThreshold,
        &dummy_pc,
        labelAcc);
    }
    return;
  }
 
  uint32_t model_output_index = glo_idx * clustererNN.mNnClusterizerModelReg2NumOutputNodes;
 
  if ((clustererNN.mOutputDataClass[full_glo_idx] > 0) || (clustererNN.mNnClusterizerUseClassification <= 0)) {
 
    ClusterAccumulator pc;
 
    if (withMC) {
      ClusterAccumulator dummy_pc;
      CPU_ONLY(labelAcc->collect(peak, central_charge));
      GPUTPCCFClusterizer::buildCluster(
        clusterer.Param().rec,
        chargeMap,
        peak,
        smem.posBcast,
        smem.buf,
        smem.innerAboveThreshold,
        &dummy_pc,
        labelAcc);
    }
    if ((clusterer.mPmemory->fragment).isOverlap(peak.time())) {
      if (clusterer.mPclusterPosInRow) {
        clusterer.mPclusterPosInRow[full_glo_idx] = clusterer.mNMaxClusterPerRow;
      }
      return;
    }
 
    // Cluster 1
    if (dtype == 0) {
      pc.setFull(central_charge * clustererNN.mOutputDataReg2_16[model_output_index + 8].ToFloat(),
                 static_cast<float>(peak.pad()) + clustererNN.mOutputDataReg2_16[model_output_index].ToFloat(),
                 clustererNN.mOutputDataReg2_16[model_output_index + 4].ToFloat(),
                 (clusterer.mPmemory->fragment).start + static_cast<float>(peak.time()) + clustererNN.mOutputDataReg2_16[model_output_index + 2].ToFloat(),
                 clustererNN.mOutputDataReg2_16[model_output_index + 6].ToFloat(),
                 clustererNN.mClusterFlags[2 * glo_idx],
                 clustererNN.mClusterFlags[2 * glo_idx + 1]);
    } else if (dtype == 1) {
      pc.setFull(central_charge * clustererNN.mOutputDataReg2_32[model_output_index + 8],
                 static_cast<float>(peak.pad()) + clustererNN.mOutputDataReg2_32[model_output_index],
                 clustererNN.mOutputDataReg2_32[model_output_index + 4],
                 (clusterer.mPmemory->fragment).start + static_cast<float>(peak.time()) + clustererNN.mOutputDataReg2_32[model_output_index + 2],
                 clustererNN.mOutputDataReg2_32[model_output_index + 6],
                 clustererNN.mClusterFlags[2 * glo_idx],
                 clustererNN.mClusterFlags[2 * glo_idx + 1]);
    }
 
    tpc::ClusterNative myCluster;
    bool rejectCluster = !pc.toNative(peak, central_charge, myCluster, clusterer.Param(), chargeMap);
    if (rejectCluster) {
      if (clusterer.mPclusterPosInRow) {
        clusterer.mPclusterPosInRow[full_glo_idx] = clusterer.mNMaxClusterPerRow;
      }
      return;
    }
 
    uint32_t rowIndex = 0;
    if (clusterOut != nullptr) {
      rowIndex = GPUTPCCFClusterizer::sortIntoBuckets(
        clusterer,
        myCluster,
        peak.row(),
        clusterer.mNMaxClusterPerRow,
        clusterer.mPclusterInRow,
        clusterOut);
      if (clusterer.mPclusterPosInRow != nullptr) {
        clusterer.mPclusterPosInRow[full_glo_idx] = rowIndex;
      }
    } else if (clusterer.mPclusterPosInRow) {
      rowIndex = clusterer.mPclusterPosInRow[full_glo_idx];
    }
    CPU_ONLY(labelAcc->commit(peak.row(), rowIndex, clusterer.mNMaxClusterPerRow));
 
    // Cluster 2
    if (dtype == 0) {
      pc.setFull(central_charge * clustererNN.mOutputDataReg2_16[model_output_index + 9].ToFloat(),
                 static_cast<float>(peak.pad()) + clustererNN.mOutputDataReg2_16[model_output_index + 1].ToFloat(),
                 clustererNN.mOutputDataReg2_16[model_output_index + 5].ToFloat(),
                 (clusterer.mPmemory->fragment).start + static_cast<float>(peak.time()) + clustererNN.mOutputDataReg2_16[model_output_index + 3].ToFloat(),
                 clustererNN.mOutputDataReg2_16[model_output_index + 7].ToFloat(),
                 clustererNN.mClusterFlags[2 * glo_idx],
                 clustererNN.mClusterFlags[2 * glo_idx + 1]);
    } else if (dtype == 1) {
      pc.setFull(central_charge * clustererNN.mOutputDataReg2_32[model_output_index + 9],
                 static_cast<float>(peak.pad()) + clustererNN.mOutputDataReg2_32[model_output_index + 1],
                 clustererNN.mOutputDataReg2_32[model_output_index + 5],
                 (clusterer.mPmemory->fragment).start + static_cast<float>(peak.time()) + clustererNN.mOutputDataReg2_32[model_output_index + 3],
                 clustererNN.mOutputDataReg2_32[model_output_index + 7],
                 clustererNN.mClusterFlags[2 * glo_idx],
                 clustererNN.mClusterFlags[2 * glo_idx + 1]);
    }
 
    rejectCluster = !pc.toNative(peak, central_charge, myCluster, clusterer.Param(), chargeMap);
    if (rejectCluster) {
      if (clusterer.mPclusterPosInRow) {
        clusterer.mPclusterPosInRow[full_glo_idx] = clusterer.mNMaxClusterPerRow;
      }
      return;
    }
 
    if (clusterOut != nullptr) {
      rowIndex = GPUTPCCFClusterizer::sortIntoBuckets(
        clusterer,
        myCluster,
        peak.row(),
        clusterer.mNMaxClusterPerRow,
        clusterer.mPclusterInRow,
        clusterOut);
      if (clusterer.mPclusterPosInRow != nullptr) {
        clusterer.mPclusterPosInRow[full_glo_idx] = rowIndex;
      }
    } else if (clusterer.mPclusterPosInRow) {
      rowIndex = clusterer.mPclusterPosInRow[full_glo_idx];
    }
    // CPU_ONLY(labelAcc->commit(peak.row(), rowIndex, clusterer.mNMaxClusterPerRow)); // -> Is this needed? How to handle MC labels for split clusters?
  } else {
    if (clusterer.mPclusterPosInRow) {
      clusterer.mPclusterPosInRow[full_glo_idx] = clusterer.mNMaxClusterPerRow;
    }
    return;
  }
}
 
// ---------------------------------
template <>
GPUdii() void GPUTPCNNClusterizerKernels::Thread<GPUTPCNNClusterizerKernels::publishDeconvolutionFlags>(int32_t nBlocks, int32_t nThreads, int32_t iBlock, int32_t iThread, GPUSharedMemory& smem, processorType& processors, uint8_t sector, int8_t dtype, int8_t withMC, uint batchStart)
{
  // Implements identical publishing logic as the heuristic clusterizer and deconvolution kernel
  uint32_t idx = get_global_id(0);
  auto& clusterer = processors.tpcClusterer[sector];
  auto& clustererNN = processors.tpcNNClusterer[sector];
  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);
    PackedCharge charge = chargeMap[tmp_pos];
    clustererNN.mClusterFlags[2 * idx] += (d.y != 0 && charge.isSplit());
    clustererNN.mClusterFlags[2 * idx + 1] += (d.x != 0 && charge.isSplit());
  }
  for (int i = 0; i < 16; i++) {
    Delta2 d = cfconsts::OuterNeighbors[i];
    CfChargePos tmp_pos = peak.delta(d);
    PackedCharge charge = chargeMap[tmp_pos];
    clustererNN.mClusterFlags[2 * idx] += (d.y != 0 && charge.isSplit() && !charge.has3x3Peak());
    clustererNN.mClusterFlags[2 * idx + 1] += (d.x != 0 && charge.isSplit() && !charge.has3x3Peak());
  }
}
 
// THe following arithmetic is done because the network is trained with a split between IROC and OROC boundary
GPUd() int32_t GPUTPCNNClusterizerKernels::padOffset(int32_t row_ref, int32_t row_current)
{
  if (row_current < 0 || row_current >= o2::tpc::constants::MAXGLOBALPADROW) {
    return 0; // Short-circuit for negative rows
  } else {
    return (int)((GPUTPCGeometry::NPads(row_current) - GPUTPCGeometry::NPads(row_ref)) / 2);
  }
}
 
GPUd() int32_t GPUTPCNNClusterizerKernels::rowOffset(int32_t row, int32_t offset)
{
  return (row > 62 ? offset : 0);
}
 
GPUd() bool GPUTPCNNClusterizerKernels::isBoundary(int32_t row, int32_t pad, int32_t offset)
{
  if (pad < 0 || row < 0) { // Faster short-circuit
    return true;
  } else if (row < 63) {
    return ((pad < 0) || (pad >= static_cast<int>(GPUTPCGeometry::NPads(row))));
  } else if (row < (63 + offset)) { // to account for the gap between IROC and OROC. Charge will be set to the boundary fill value in order to signal boundaries to the neural network
    return true;
  } else if (row < (o2::tpc::constants::MAXGLOBALPADROW + offset)) {
    return ((pad < 0) || (pad >= static_cast<int>(GPUTPCGeometry::NPads(row - offset))));
  } else {
    return true;
  }
}