ClusterNative.h

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.
 
#ifndef ALICEO2_DATAFORMATSTPC_CLUSTERNATIVE_H
#define ALICEO2_DATAFORMATSTPC_CLUSTERNATIVE_H
#ifndef GPUCA_GPUCODE_DEVICE
#include <climits>
#include <cstdint>
#include <cstddef> // for size_t
#include <utility>
#endif
#include "DataFormatsTPC/Constants.h"
#include "GPUCommonDef.h"
 
namespace o2
{
class MCCompLabel;
namespace dataformats
{
template <class T>
class ConstMCTruthContainer;
template <class T>
class ConstMCTruthContainerView;
} // namespace dataformats
} // namespace o2
 
namespace o2
{
namespace tpc
{
struct ClusterNative {
  // NOTE: These states must match those from GPUTPCGMMergedTrackHit!
  enum clusterState { flagSplitPad = 0x1,  // Split in pad direction
                      flagSplitTime = 0x2, // Split in time direction
                      flagEdge = 0x4,      // At edge of TPC sector
                      flagSingle = 0x8 };  // Single pad or single time-bin cluster
 
  static GPUglobalconstexpr() int scaleTimePacked = 64;      //< ~50 is needed for 0.1mm precision, but leads to float rounding artifacts around 20ms
  static GPUglobalconstexpr() int scalePadPacked = 64;       //< ~60 is needed for 0.1mm precision, but power of two avoids rounding
  static GPUglobalconstexpr() int scaleSigmaTimePacked = 32; // 1/32nd of pad/timebin precision for cluster size
  static GPUglobalconstexpr() int scaleSigmaPadPacked = 32;
  static GPUglobalconstexpr() int scaleSaturatedQtot = 8;
  static GPUglobalconstexpr() int maxRegularQtot = 25 * 1024;
  static GPUglobalconstexpr() int maxSaturatedQtot = (USHRT_MAX - maxRegularQtot) * scaleSaturatedQtot;
 
  uint32_t timeFlagsPacked; //< Contains the time in the lower 24 bits in a packed format, contains the flags in the
                            // upper 8 bits
  uint16_t padPacked;       //< Contains the pad in a packed format
  uint8_t sigmaTimePacked;  //< Sigma of the time in packed format
  uint8_t sigmaPadPacked;   //< Sigma of the pad in packed format
  uint16_t qMax;            //< QMax of the cluster
  uint16_t qTotPacked;      //< Total charge of the cluster
 
  GPUd() static uint16_t packPad(float pad) { return (uint16_t)(pad * scalePadPacked + 0.5); }
  GPUd() static uint32_t packTime(float time) { return (uint32_t)(time * scaleTimePacked + 0.5); }
  GPUd() static float unpackPad(uint16_t pad) { return float(pad) * (1.f / scalePadPacked); }
  GPUd() static float unpackTime(uint32_t time) { return float(time) * (1.f / scaleTimePacked); }
 
  GPUdDefault() ClusterNative() = default;
  GPUd() ClusterNative(uint32_t time, uint8_t flags, uint16_t pad, uint8_t sigmaTime, uint8_t sigmaPad, uint16_t qmax, uint16_t qtotPacked) : padPacked(pad), sigmaTimePacked(sigmaTime), sigmaPadPacked(sigmaPad), qMax(qmax), qTotPacked(qtotPacked)
  {
    setTimePackedFlags(time, flags);
  }
 
  GPUd() uint16_t getQmax() const { return qMax; }
  GPUd() uint32_t getQtot() const { return isSaturated() ? getSaturatedQtot() : (uint32_t)qTotPacked; }
  GPUd() uint8_t getFlags() const { return timeFlagsPacked >> 24; }
  GPUd() uint32_t getTimePacked() const { return timeFlagsPacked & 0xFFFFFF; }
  GPUd() void setTimePackedFlags(uint32_t timePacked, uint8_t flags)
  {
    timeFlagsPacked = (timePacked & 0xFFFFFF) | (uint32_t)flags << 24;
  }
  GPUd() void setTimePacked(uint32_t timePacked)
  {
    timeFlagsPacked = (timePacked & 0xFFFFFF) | (timeFlagsPacked & 0xFF000000);
  }
  GPUd() void setFlags(uint8_t flags) { timeFlagsPacked = (timeFlagsPacked & 0xFFFFFF) | ((uint32_t)flags << 24); }
  GPUd() float getTime() const { return unpackTime(timeFlagsPacked & 0xFFFFFF); }
  GPUd() void setTime(float time)
  {
    timeFlagsPacked = (packTime(time) & 0xFFFFFF) | (timeFlagsPacked & 0xFF000000);
  }
  GPUd() void setTimeFlags(float time, uint8_t flags)
  {
    timeFlagsPacked = (packTime(time) & 0xFFFFFF) | ((decltype(timeFlagsPacked))flags << 24);
  }
 
  GPUd() float getPad() const { return unpackPad(padPacked); }
  GPUd() void setPad(float pad) { padPacked = packPad(pad); }
  GPUd() float getSigmaTime() const
  {
    if (isSaturated()) [[unlikely]] {
      return 0;
    }
    return float(sigmaTimePacked) * (1.f / scaleSigmaTimePacked);
  }
  GPUd() void setSigmaTime(float sigmaTime)
  {
    uint32_t tmp = sigmaTime * scaleSigmaTimePacked + 0.5;
    if (tmp > 0xFF) {
      tmp = 0xFF;
    }
    sigmaTimePacked = tmp;
  }
  GPUd() float getSigmaPad() const { return float(sigmaPadPacked) * (1.f / scaleSigmaPadPacked); }
  GPUd() void setSigmaPad(float sigmaPad)
  {
    uint32_t tmp = sigmaPad * scaleSigmaPadPacked + 0.5;
    if (tmp > 0xFF) {
      tmp = 0xFF;
    }
    sigmaPadPacked = tmp;
  }
 
  GPUd() bool isSaturated() const { return qTotPacked > maxRegularQtot; }
 
  GPUd() void setSaturatedQtot(uint32_t qtot)
  {
    this->qTotPacked = USHRT_MAX;
    if (qtot < maxSaturatedQtot) {
      this->qTotPacked = ((qtot + scaleSaturatedQtot / 2) / scaleSaturatedQtot) + maxRegularQtot;
    }
  }
 
  GPUd() uint32_t getSaturatedQtot() const
  {
    return uint32_t(qTotPacked - maxRegularQtot) * scaleSaturatedQtot;
  }
 
  GPUd() void setSaturatedTailLength(uint32_t tail)
  {
    sigmaTimePacked = encodeTailLength(tail);
  }
 
  GPUd() uint32_t getSaturatedTailLength() const
  {
    return decodeTailLength(sigmaTimePacked);
  }
 
  GPUd() bool operator<(const ClusterNative& rhs) const
  {
    if (this->getTimePacked() != rhs.getTimePacked()) {
      return (this->getTimePacked() < rhs.getTimePacked());
    } else if (this->padPacked != rhs.padPacked) {
      return (this->padPacked < rhs.padPacked);
    } else if (this->sigmaTimePacked != rhs.sigmaTimePacked) {
      return (this->sigmaTimePacked < rhs.sigmaTimePacked);
    } else if (this->sigmaPadPacked != rhs.sigmaPadPacked) {
      return (this->sigmaPadPacked < rhs.sigmaPadPacked);
    } else if (this->qMax != rhs.qMax) {
      return (this->qMax < rhs.qMax);
    } else if (this->qTotPacked != rhs.qTotPacked) {
      return (this->getQtot() < rhs.getQtot());
    } else {
      return (this->getFlags() < rhs.getFlags());
    }
  }
 
  GPUd() bool operator==(const ClusterNative& rhs) const
  {
    return this->getTimePacked() == rhs.getTimePacked() &&
           this->padPacked == rhs.padPacked &&
           this->sigmaTimePacked == rhs.sigmaTimePacked &&
           this->sigmaPadPacked == rhs.sigmaPadPacked &&
           this->qMax == rhs.qMax &&
           this->qTotPacked == rhs.qTotPacked &&
           this->getFlags() == rhs.getFlags();
  }
 
 private:
  static constexpr GPUd() uint32_t decodeTailLength(uint8_t code)
  {
    // Quantize tail length into 8bits.
    // Max expected length is 1500 tbs.
    // But allow outliers up to 8000 tbs.
    //
    // Full code layout is:
    //
    // | Code range | Decoded values |  Step | Codes |
    // | ---------: | -------------: | ----: | ----: |
    // |    `0..63` |        `0..63` |   `1` |  `64` |
    // |   `64..95` |      `64..126` |   `2` |  `32` |
    // |  `96..127` |     `128..252` |   `4` |  `32` |
    // | `128..159` |     `256..504` |   `8` |  `32` |
    // | `160..223` |    `512..1520` |  `16` |  `64` |
    // | `224..239` |   `1552..2032` |  `32` |  `16` |
    // | `240..255` |   `2048..8048` | `400` |  `16` |
    //
 
    if (code < 64) {
      return code;
    }
 
    if (code < 160) {
      uint32_t q = (uint32_t)code - 64u;
      uint32_t exponent = (q >> 5) + 1u; // 1, 2, 3
      uint32_t mantissa = q & 31u;       // 0..31
 
      return (32u + mantissa) << exponent;
    }
 
    if (code < 224) {
      return 512u + 16u * ((uint32_t)code - 160u);
    }
 
    if (code < 240) {
      return 1552u + 32u * ((uint32_t)code - 224u);
    }
 
    return 2048u + 400u * ((uint32_t)code - 240u);
  }
 
  static constexpr GPUd() uint8_t encodeTailLength(uint32_t value)
  {
    // Saturate above representable range.
    if (value >= decodeTailLength(255)) [[unlikely]] {
      return 255;
    }
 
    // Binary search for the first code whose decoded value >= value.
    uint8_t lo = 0;
    uint8_t hi = 255;
 
    while (lo < hi) {
      uint8_t mid = lo + ((hi - lo) >> 1);
      uint32_t decoded = decodeTailLength(mid);
 
      if (decoded < value) {
        lo = mid + 1;
      } else {
        hi = mid;
      }
    }
 
    // lo is now the first code with decoded >= value.
    if (lo == 0) [[unlikely]] {
      return 0;
    }
 
    uint8_t above_code = lo;
    uint8_t below_code = lo - 1;
 
    uint32_t above_value = decodeTailLength(above_code);
    uint32_t below_value = decodeTailLength(below_code);
 
    uint32_t above_error = above_value - value;
    uint32_t below_error = value - below_value;
 
    // Tie-break downward.
    if (below_error <= above_error) {
      return below_code;
    } else {
      return above_code;
    }
  }
};
 
// This is an index struct to access TPC clusters inside sectors and rows. It shall not own the data, but just point to
// the data inside a buffer.
struct ClusterNativeAccess {
  const ClusterNative* clustersLinear;
  const ClusterNative* clusters[constants::MAXSECTOR][constants::MAXGLOBALPADROW];
  const o2::dataformats::ConstMCTruthContainerView<o2::MCCompLabel>* clustersMCTruth;
  unsigned int nClusters[constants::MAXSECTOR][constants::MAXGLOBALPADROW];
  unsigned int nClustersSector[constants::MAXSECTOR];
  unsigned int clusterOffset[constants::MAXSECTOR][constants::MAXGLOBALPADROW];
  unsigned int nClustersTotal; // Must be directly after clusterOffsets, --> =clusterOffset[nRows * nSectors]!
 
  void setOffsetPtrs();
 
#ifndef GPUCA_GPUCODE
  using ConstMCLabelContainer = o2::dataformats::ConstMCTruthContainer<o2::MCCompLabel>;
  using ConstMCLabelContainerView = o2::dataformats::ConstMCTruthContainerView<o2::MCCompLabel>;
  using ConstMCLabelContainerViewWithBuffer = std::pair<ConstMCLabelContainer, ConstMCLabelContainerView>;
#endif
};
 
inline void ClusterNativeAccess::setOffsetPtrs()
{
  int offset = 0;
  for (unsigned int i = 0; i < constants::MAXSECTOR; i++) {
    nClustersSector[i] = 0;
    for (unsigned int j = 0; j < constants::MAXGLOBALPADROW; j++) {
      clusterOffset[i][j] = offset;
      clusters[i][j] = &clustersLinear[offset];
      nClustersSector[i] += nClusters[i][j];
      offset += nClusters[i][j];
    }
  }
  nClustersTotal = offset;
}
} // namespace tpc
} // namespace o2
#endif